Artificial Intelligence in dental health care – a case study

Artificial Intelligence in dental health care – a case study

/in , , , /by and

Dentistry is one of the oldest branches of medicine, its roots long predating agriculture. The earliest evidence of dental treatment comes from the upper paleolithic period. Like all medicine, dental healthcare today looks nothing like it did during the stone age. Machine learning, among our era’s most promising contributions, is a major factor in the difference.

Dental health is among the most important areas of modern healthcare. Oral diseases are the most common non-communicable diseases in the world. According to WHO data, an estimated half of the global population struggle with dental problems like tooth decay. The most common diseases are tooth decay and periodontal disease. Tooth and gum treatments happen also to be costly, accounting for 5% of total health expenditures and 20% of out-of-pocket expenditures.

How do AI and healthcare come together?

According to Accenture, automated image diagnosis could be worth up to a $3 billion chunk of the healthcare market. Considering the state-of-the-art image recognition solutions currently available, the most obvious applications include MRIs and other scan interpretations.


Move your mouse cursor over the image to see the difference.

According to estimates from IBM, up to 90% of all medical data is reported in image form. Thus, improving diagnostics and treatment with image recognition tools is the top-of-mind application. In healthcare, however, the road from idea to implementation can be far more tortuous than in other industries.

Challenges for AI in healthcare

Supporting modern healthcare is imperative to the quality of life humans can have. At the same time, the potential for machine learning technology to create breakthroughs that will benefit humanity is enormous. This makes it all the more important that the multiple challenges we face be overcome.

  • The challenging process of obtaining data – because medical data is sensitive, it is not easy to create a database that contains sufficient entries that are also anonymous.
  • The need for strict human supervision of artificial intelligence. There is no place for mistakes in healthcare – it is all a matter of human life and health. Also, there is no legal framework governing decisions made by AI and applied in treatment, so human supervision is key.

Keep smiling

Our client is a leading producer of dental equipment, delivering a full range of instruments – from single-use tubes and appliances to head and tooth x-ray devices.

The x-rays their machines deliver are crucial in diagnosing multiple oral afflictions, including tooth decay, periodontal disease and multiple others. Yet the raw images x-rays present, enhanced only with slight processing and basic clarifications, is often all that the physician gets. Unlike with MRIs and the process of interpreting them, x-ray machines themselves don’t support the diagnostic process.


Move your mouse cursor over the image to see the difference.

Our client was looking for a way to support dentists in their work by preprocessing x-ray images and providing more information about the potential diseases they turned up. The company also needed to automate the dull and repetitive work of spotting and writing up information in medical documentation. That information mainly concerned particular teeth depicted in the images their machines produced.

The teething problem

The initial goal was to properly mark each tooth in an image. This would prove more difficult than you might imagine, especially given the number of teeth that can be missing. In fact, according to the World Dental Federation, up to 30% of the population, aged 65-74 have no natural teeth left.

This was reflected in the data the client provided, which we used to train our model. An average healthy adult human has 32 teeth, but there were an average of 26 teeth depicted in the pictures we dealt with and included in our model. The model was responsible for spotting the right positions of particular teeth, including telling which ones were missing.

Additional challenges included:

  • Localizing implants, restorations and other things used in medical treatment
  • Uncovering signs of the most common oral diseases such as cavities and periodontitis

This task was further complicated by the fact that cavities are so indistinct in x-ray images that an untrained eye has little to no chance of finding them (in contrast to distinguishing between Aliens and Predators). And if a random person struggles to see something, an algorithm will, too.

From a machine learning point of view, the project was an object detection problem, similar to others we have handled during projects for Nielsen and the National Oceanic and Atmospheric Administration. We were given around 1000 anonymized, high-resolution x-ray pictures along with bounding boxes around teeth and other findings created by professional dentists and radiologists.


Move your mouse cursor over the image to see the difference.

We used PyTorch’s MaskRCNN neural network model to simultaneously localize teeth and other findings and determine their labels. The model was supported by our custom-made algorithm, which checked the correctness of the order of tooth labels and, more importantly, canceled some predictions. A prediction might need to be ruled out if, for example, a neural network were to spot 17 teeth (where only 16 are possible in a healthy human jaw). In any case, this bonus step was only needed for a few images with multiple chained crowns inserted to replace a row of missing teeth. For the vast majority of images, the neural network predicted the tooth labels perfectly.

We also implemented a set of heuristic postprocessing procedures addressing human knowledge not available to a model based on images only. Examples of human knowledge included:

  • You can’t have a cavity outside a natural tooth: on your collarbone, on an implant replacing a missing tooth, on a crown, etc.
  • You can have either a crown or a dental seal on your tooth.
  • Even if you have a double root canal treatment in a premolar or molar tooth, the model should highlight it with a single frame in order not to confuse the dentist.

The accuracy

Our model found it easy to spot and identify particular teeth adhering to the World Dental Federation Notation, even for ‘unusual’ patients, such as those with only a few teeth left. It localized and labeled teeth, making only a handful of mistakes while processing hundreds of images. Also, most object categories, including implants, dental work and prosthetics, were detected properly, with the mean average precision more than 0.9 on a 0-1 scale. Finally, the model’s hyper-sensitivity ensures that the early signs of decay or other oral diseases are spotted.


Move your mouse cursor over the image to see the difference.

When dealing with cavities, providing the user with more false-positives rather than overlooking the signs of a developing illness was considered a more feasible strategy. Images have to be reviewed by a dentist anyway, so delivering more warnings, even with a higher rate of false-positives, can deliver a great dose of support for dental professionals.

Summary

Using deep learning-based AI systems in healthcare is no longer confined to the realm of science fiction. In this case, a mere x-ray has armed dentists with images enhanced with ready-to-go information on the placement and state of every tooth in a patient’s mouth, as well as a cheat-sheet on the early signs of disease that could have been overlooked.

This project is a perfect example of how such systems can reduce the dull, repetitive and mistake-prone work usually done by professionals, who could be putting their expert knowledge to better use.

A business guide to Natural Language Processing (NLP)

A business guide to Natural Language Processing (NLP)

/in , , /by and

With chatbots powering up customer service on one hand and fake news farms on the other, Natural Language Processing (NLP) is getting attention as one of the most impactful branches of Artificial Intelligence (AI).

When Alan Turing proposed his famous test in 1950, he couldn’t, despite the prescience that accompanies brilliance such as his, predict how easy breaking the test would become. And how far from intelligence the machine that broke the test would be!

Modern Natural Language Processing is being used in multiple industries, in both large-scale projects delivered by tech giants and minor tweaks local companies employ to improve the user experience.

The solutions vary from supporting internal business processes in document management to improving customer service by automated responses generated for the most common questions. According to IDC data cited by Deloitte, companies leveraging the information buried in plain sight in documents and other unstructured data can achieve up to $430 billion in productivity gains by 2020.

The biggest problem with NLP is the significant difference between machines mimicking the understanding of text and actually understanding it. The difference is easily shown with ELIZA software (a famous chatbot from the 1960s), which was based on a set of scripts that paraphrased input text to produce credible-looking responses. The technology was sufficient to produce some text, but far from demonstrating real understanding or delivering business value. Things changed, however, once machine learning models came into use.

What is natural language processing?

As the name implies, natural language processing is the act of a machine processing human language, analyzing the queries in it and responding in a human manner. After several decades of NLP research strongly based on a combination of computer science and linguistic expertise, the “deep learning tsunami” (a term coined by Stanford CS and Linguistics professor Christopher Manning) has recently taken over this field of AI as well, similarly to what happened in computer vision.

Many NLP tasks today are tackled with deep neural networks, which are frequently used among various techniques that enable machines to understand a text’s meaning and its author’s intent.

Modern NLP solutions work on text by “reading” it and making a network of connections between each word. Thus, the model gets more information on the context, the sentiment and exactly what the author sought to communicate.

Tackling the context

Context and intent are critical in analyzing text. Analyzing a picture without context can be tricky  – is a fist a symbol of violence, or a bro fist-bump?

The challenge grows even further with NLP, as there are multiple social and cultural norms at work in communication. “The cafe is way too cool for me” can refer to a too-groovy atmosphere or the temperature. Depending on the age of the speaker, a “savage” punk rock concert can be either positive or negative. Before the machine learning era, the flatness of traditional, dictionary-based solutions provided information with significantly less accuracy.

The best way to deal with this challenge is to deliver a word-mapping system based on multidimensional vectors (so-called word embeddings) that provide complex information on the words they represent. Following the idea of distributional semantics (“You shall know a word by the company it keeps”), the neural network learns word representations by looking at the neighboring words. A breakthrough moment for neural NLP came in 2013, when the renowned word2vec model was introduced. However, one of the main problems that word2vec could not solve was homonymy, as the model could not distinguish between different meanings of the same word. A way to significantly improve handling the context in which a word is used in a sentence was found in 2018, when more sophisticated word embedding models like BERT and ELMo were introduced.

Natural Language Processing examples

Recent breakthroughs, especially GPT-2, have significantly improved NLP and delivered some very promising use cases, including the ones elaborated below.

Automated translation

One of the most widely used applications of natural language processing is automated translation between two languages, e.g. with Google Translate. The translator delivers increasingly accurate texts, good enough to serve even in court trials. Google Translate was used when a British court failed to deliver an interpreter for a Mandarin speaker.

Machine translation was one of the first successful applications of deep learning in the field of NLP. The neural approach quickly surpassed statistical machine translation, the technology that preceded it. In a translation task, the system’s input and output are sequences of words. The typical neural network architecture used for translation is therefore called seq2seq, and consists of two recurrent neural networks (encoder and decoder).

The first seq2seq paper was published in 2014, and subsequent research led Google Translate to switch from statistical to neural translation in 2016. Later that year, Google announced a single multi-lingual system that could translate between pairs of languages the system had never seen explicitly, suggesting the existence of some interlingua-like representation of sentences in vector space.

A business guide to Natural Language Processing (NLP) - inline image

Another important development related to recurrent neural networks is the attention mechanism, which allows a model to learn to focus on particular parts of sequences, greatly improving translation quality. Further improvements come from using Transformer architecture instead of Recurrent Neural Networks.

Chatbots

Automated interaction with customers causes their satisfaction with the overall user experience to rise significantly. And that’s not a thing to overcome, as up to 88% of customers are willing to pay more for better customer experience.

A great example of chatbots improving the customer experience comes from Amtrak, a US railway company that transports 31 million passengers yearly and administrates over 21,000 miles of rails across America. The company decided to employ Julie, a chatbot that supports passengers in searching for a convenient commute. She delivered 800% ROI and reduced the cost of customer service by $1 million yearly while also increasing bookings by 25%.

Speech recognition

As much as a company can use a chatbot to perform some customer service, one can have a personal assistant in the pocket. According to eMarketer data, up to 111.8 million people in the US–over a third of its population–will use a voice assistant at least once a month. The voice assistant market is growing rapidly, with companies such as Google, Apple, Amazon and Samsung developing their assistants not only for mobile devices, but also for TVs and home appliances.

Despite the privacy concerns voice assistants are raising, speech is becoming the new interface for human-machine interaction. The interface can also be used to control industrial machines, especially when employees have their hands occupied – a case common across industries from HoReCa to agriculture and construction – assuming that the noise is reduced enough for machine to register the voice properly.

Thanks to advances in NLP, speech recognition solutions are getting smarter and delivering a better experience for users. As the assistants come to understand speakers’ intentions better and better, they will provide more accurate answers to increasingly complex questions.

An unexpected example of speech recognition comes from deepsense.ai’s project renovating and digitalizating classic movies, where the machine delivers an automated transcription. When combined with a facial recognition tool, the system transcribed and annotated the actor speaking in the film.

Sentiment analysis

Social media provides numerous ways of reaching customers, gathering information on their habits and delivering excellence. It’s also a melting pot of perspectives and news, delivering unprecedented insight on public opinion. This insight can be understood using sentiment analysis tools, which check if the context where a brand is exposed in social media is positive, negative or neutral.

Sentiment analysis can be done without the assistance of AI by building up a glossary of positive and negative words and checking their frequency. If there is swearing or words like “broken” near the brand, sentiment is negative. Yet those systems cannot spot irony or more sophisticated hate. The sentence “I would be happy to see you ill” suggests an aggression and possibly hatred, yet there are no slurs or swearing. By supporting the analysis of words in the glossary by checking the relations between words in each sentence, a machine learning model can deliver a better understanding of the text and provide more information on the message’s subjectivity.

So good can that understanding be, in fact, that deepsense.ai delivered a solution that could spot terrorist propaganda and illicit content in social media in real-time. In the same way, it is possible to deliver a system that spots hate speech and other forms of online harassment. A study from the Pew Research Center shows that up to 41% of adult Americans have experienced some form of online harassment, a number that is likely to increase, mostly due to the rising prevalence of the Internet in people’s daily lives.

Natural language generation

Apart from understanding text, machines are getting better at delivering new texts. According to research published in Foreign Affairs, texts being produced by modern AI software are, for unskilled readers, comparable to those written by journalists. ML models are indeed already writing texts for world media organizations. And while that may seem a fascinating accomplishment, it was the fear of what such advanced abilities might portend that led OpenAI not to make GPT-2 public.

The most known case of automated journalism comes from the Washington Post, where Heliograf covers sport events. Its journalistic debut came in 2016, when the software was responsible for writing up coverage of the Olympic Games in Rio.

In business, natural languae generation is used to produce more polite and humane responses to FAQs. Thus, ironically, automating the conventional communication will make it more personal and humane than current, trigger-based solutions.

Text analytics

Apart from delivering real-time monitoring and sentiment analysis, NLP tools can analyze long and complicated texts, as is already being done at EY, PwC and Deloitte, all of which employ machine learning models to review contracts. The same can be applied to analyze emails or other company-owned unstructured data. According to Gartner estimates, up to 80% of all business data is unstructured and thus nonactionable for companies.

A good example of natural language processing in text analytics is a solution deepsense.ai designed for market research giant Nielsen. The company delivered reports on the ingredients in all of the FMCG products available on the market.

The process of gathering the data was time-consuming and riddled with pitfalls:  an employee had to manually read a label, check the ingredients and fill out the tables. The entire process took up to 30 minutes per product. Also, due to inconsistencies in naming, the task was riddled with inconsistencies, as the companies delivered the product ingredients in local languages, English and, especially on the beauty and skin care markets, Latin.

deepsense.ai delivered a comprehensive system that processed an image of the product label taken with a smartphone. The solution spotted the ingredients, scanned the text and sorted the ingredients into tables, effectively reducing the work time from 30 minutes to less than two minutes, including the time needed to gather and validate the data.

Another use case of text analytics is the automated question response function generated by Google, which aims not only to provide search results for particular queries, but a complete answer to the user’s needs, including a link to the referred website, and a description of the matter.

Summary

Natural language processing provides numerous opportunities for companies from multiple industries and segments. Apart from relatively intuitive ways to leverage NLP, such as processing the documents and chatbots, there are multiple other applications, including real time social media analytics and supporting journalism or research work.

NLP models can be used to further augment existing solutions–from supporting the reinforcement learning models behind autonomous cars by providing better sign recognition to augmenting demand forecasting tools with extensions to analyze headlines and deliver more event-based predictions.

Because natural language is the best way to transfer information between humans and machines, the applications NLP makes possible will only increase and will soon be augmenting business processes around the globe.

AI Monthly Digest #12 - the shadow of malicious use

AI Monthly Digest #12 – the shadow of malicious use

/in , /by and

With this edition of AI Monthly Digest, we have now for a full year been bringing readers carefully selected and curated news from the world of AI and Machine Learning (ML) that deepsense.ai’s team considers important, inspiring and entertaining.

Our aim is to deliver information that people not necessarily involved in AI and ML may find interesting. Also, the digest is curated by data scientists who ensure that the information included isn’t just hot air or marketing mumbo-jumbo, but significant news that will impact the global machine learning and reinforcement learning world.

This edition focuses on natural language processing, as the GPT-2 model is still an important element of AI-related discourse. This edition also contrasts the enthusiasm of ML-developers with concerns expressed by a renowned professor of Psychology.

1200 questions to ask

With natural language processing, a computer needs to generate natural texts in response to a human. This is at least troublesome, especially if a longer text or speech is required.

While these problems are being tackled in various ways, the gold standard is currently to run the newest solution on a benchmark. Yet delivering one is another challenge, to put it mildly.

To tackle it, researchers from the University of Maryland created a set of over 1200 questions that are easy to answer for a human and nearly impossible for a machine. To jump from “easy” to “impossible” is sometimes a matter of very subtle changes. As the researchers have said:

if the author writes “What composer’s Variations on a Theme by Haydn was inspired by Karl Ferdinand Pohl?” and the system correctly answers “Johannes Brahms,” the interface highlights the words “Ferdinand Pohl” to show that this phrase led it to the answer. Using that information, the author can edit the question to make it more difficult for the computer without altering the meaning of the question. In this example, the author replaced the name of the man who inspired Brahms, “Karl Ferdinand Pohl,” with a description of his job, “the archivist of the Vienna Musikverein,” and the computer was unable to answer correctly. However, expert human quiz game players could still easily answer the edited question correctly.

Capitalizing on this knowledge, researchers will be able to deliver better benchmarking for models and thus determine which part of the question confuses the computer.

Why does it matter

With each and every breakthrough, researchers get closer to delivering human-level natural language processing. At the same time, however, it is increasingly hard to determine if the neural network is understanding the processed text, or is just getting better fitted to the benchmark. Were the latter the case, the model would outperform existing solutions but register no significant improvement in real-life performance.

An example with detailed explanations are available in the video below.

A benchmark updated with those 1200 questions delivers significantly more precise information on the model’s ability to process the language and spot the drawbacks.

Large GPT-2 released

GPT-2 is probably the hottest topic among AI Trends 2019, especially considering the groundbreaking effect and controversial decision to NOT make the model public. Instead, OpenAI, the company behind the model, decided to cooperate with chosen institutions to find a way to harden the model against potential misuse.

And the threat is serious. According to research published in Foreign Affairs, readers consider GPT-2-written texts nearly as credible and convincing as those written by journalists and published in The New York Times (72% compared to 83%). Thus the articles are good enough to be especially dangerous as a weapon of mass disinformation or fake news factory – AI can produce a limitless amount of credible-looking texts with no effort.

To find the balance between supporting the development of the global science of AI and protecting models from being used for maleficent ends, OpenAI is releasing the model in iterations, starting a small one and ultimately aiming to make the model public but with the threat of misuse minimized.

Why does it matter

As research published in Foreign Affairs states, the model produces texts that an unskilled reader will find comparable to journalist-written ones. Image recognition models are already outperforming human experts in their tasks. But with all these cultural contexts, humor and irony, natural language once seemed protected by the unassailable fortress of the human mind.

The GPT-2 model has apparently cracked the gates and with business appliances it may be on the road to delivering a model that can provide human-like performance. The technology just needs to be controlled so as not to fall into the wrong hands.

What is this GPT-2 all about?

A GPT-2 model is, as stated above, one of the hottest topics of AI in 2019. But even the specialist can find it hard to understand the nitty-gritty of how the model works. To make the matter more clear, Jay Alammar has prepared a comprehensive guide to the technology.

Why does it matter

The guide is good enough to allow a person who has limited to no knowledge on the matter to understand the nuances of the model. For a moderately skilled data scientist given sufficient computing power and a dataset, the guide is sufficient to reproduce the model for example to support demand forecasting with NLP. It enables a data scientist to broaden his or her knowledge with one comprehensive article – a convenient way indeed.

Doing research is one thing, but sharing the knowledge it affords is a whole different story.

Malicious use, you say?

Jordan Peterson is a renowned professor and psychologist who studies the structure of myth and its role in shaping social behavior. If not a household name, he is certainly a public person and well-known speaker.

Using deep neural networks, AI researcher Chris Vigorito launched a notjordanpeterson.com website that allowed any user to generate any text that was later read with the neural network-generated voice of Jordan Peterson. As was the case with Joe Rogan, the output was highly convincing, mirroring the manner of speaking, breathing and natural pauses.

The networks was trained on 20 hours of transcripted Jordan Peterson speeches, an easy number to obtain where a public speaker is concerned. The amount of work was considerable, but not overwhelming.

Why does it matter

The creation of the neural network is not as interesting as Jordan Peterson’s response. He has written a blogpost entitled “I didn’t say that”, where he calls the situation “very strange and disturbing”. In the post, he notes that while it was fun to hear himself singing popular songs, the prospect of being an unwitting part of a scam is more than real. Due to the rising computing power available at affordable prices and algorithms getting better and less data-hungry, the threat of this technology being used for malicious ends is rising. If you’d like to know just how malicious he means, I’ll leave you with this to consider.

I can tell you from personal experience, for what that’s worth, that it is far from comforting to discover an entire website devoted to allowing whoever is inspired to do so produce audio clips imitating my voice delivering whatever content the user chooses—for serious, comic or malevolent purposes. I can’t imagine what the world will be like when we will truly be unable to distinguish the real from the unreal, or exercise any control whatsoever on what videos reveal about behaviors we never engaged in, or audio avatars broadcasting any opinion at all about anything at all. I see no defense, and a tremendously expanded opportunity for unscrupulous troublemakers to warp our personal and collective reality in any manner they see fit.

AI Monthly Digest #11 – From OpenAI partnering with Microsoft to battle.net Blade Runners

/in , , /by and

AI models are skilled in Chess, Go, StarCraft and, since July, six-player Texas Hold’em Poker. But the hunt for inhuman players has begun.

And when AI models get bored with beating humans in games, it’s apparently time for a bike ride. Read on to find out just why.

OpenAI partners with Microsoft

In March, OpenAI shifted to a for-profit paradigm, forging OpenAI LP. The company has now entered into a strategic partnership with Microsoft to build a new computational platform in Azure Cloud. OpenAI will port its codebase into Azure and develop new solutions with tools available there.

The main benefit for Microsoft will be the access it gains to the fruits of OpenAI’s work as a preferred partner for commercializing new AI technologies. OpenAI was formed with the goal of improving people’s lives with technology and has delivered multiple AI breakthrough models including MuseNet, which produces music in various styles, and GPT-2, a natural language processing model and gold standard in text generation.

For more on the joint venture, click on over to the official press release.

Why does it matter?

As the company recently announced, taking a non-profit approach to develop AI proved too daunting even for the likes of Elon Musk and Sam Altman, the company’s CEO. Unlike with most software, developing artificial intelligence requires not only skilled and talented people but also an astonishing amount of computing power. Training a single GPT-2 model is estimated to cost up to $50,000 – and that’s only for one of many experiments run in a given year. So, securing access to computing power is a must. By providing that, Microsoft will be getting access to the base of talented AI developers, significantly increasing its AI development potential.

AI beats skilled human players in six-player poker

Pluribus bot is the first AI-controlled agent capable of beating human pro players in six-player no-limit Hold’em poker, the most popular format of this game in the world. Unlike chess and Go, poker is a game with hidden information – the player cannot see the hands of other players. The game itself involves bluff and a great deal of psychological factors. AI bots were good at beating one opponent, but going up against more than one was a major milestone.

More details are available in the ai.facebook blogpost.

Why does it matter

Dealing with a one-on-one situation, though common in recreational games, is rare when solving real-life problems. Moreover, it is somewhere between difficult and impossible to get all the information one needs, as in chess or Go. Delivering models that work in a limited-information, multiple-agent environment is pushing models that support demand forecasting or managing cybersecurity threats.

When an autonomous car is not enough

Cars, be they traditional or autonomous, come with various disadvantages, especially in cities. They get stuck in traffic jams, and require parking, which can be hard to come by.

To provide more sustainable autonomous transport, Chinese scientists have brought out an autonomous bicycle. The machine responds to voice commands, avoids obstacles and maintains balance. It uses a new Tianjic chip, which supports processing neuroscience-inspired algorithms. Check out the video below to see how it works.

Why does it matter

The research itself sounds like a lot of fun, but it also constitutes an excellent foundation for further work on autonomous transportation. A bike can be used for delivering fast food, or modified to work as a motorized Rickshaw in our ever more crowded metropolises. Couriers can used them to deliver the mail and other documents or to transport individuals incapable of riding a bike.

Autonomous bikes need not be fully autonomous, by the way: AI can support the driving process or provide alerts to riders.

Alan Turing will be featured on 50 pound banknote

Often hailed as the godfather of modern computing, Alan Turing is widely known for his work on cracking the famed German Enigma code and as the leader of the team that enhanced the cracking methods delivered by Polish mathematicians.

Turing is also considered a pioneer of artificial intelligence. He came up with the Turing Test as a first way to determine if a machine mimicking a human in conversation is truly intelligent.

So great was Turing’s contribution to humanity that he will now be featured on Britain’s 50-pound banknote. Chosen from 227,299 nominations covering 989 eligible characters, Turing was ultimately picked by Mark Carney, Bank of England governor.

Why does it matter

The announcement is a sign that computer science is no longer considered a novelty and prominent AI researchers earn the same respect chemists, physicists or life sciences experts do, as representation on a banknote well attests.

BattleNet Blade Runners

Deepmind, in conjunction with Blizzard, has deployed AlphaStar model on Battle.net, to allow players to test their skills and mettle against artificial intelligence. Battle.net is an official platform connecting players from all around the world, enabling them to quickly find opponents for a multiplayer match.

There is just one twist: the famed, reinforcement learning-trained AlphaStar will play anonymously, thus allowing players to compete with the model as they would do in any match with a normal opponent.

AlphaStar has been developed significantly beyond the abilities it commanded in defeating human professional players MaNa and TLO. Deepmind capped the actions-per-minute and actions-per-second rate to make it more accurately appropriate human abilities limited by muscles and the need to operate a mouse and keyboard.  The model’s perception has also been narrowed to a single frame to come in line with what human players see on the screen.

Finally, the model is able to control and compete in any race given, be it Terran, Protoss or Zerg, representing all the factions available in the game. This represents serious progress: during matches in January, the model could only control Protoss units fighting against other Protoss.

Why does it matter

At the moment, it doesn’t. But let the experiment run its course and tune in later for an update. We anticipate more impressive progress.

Interestingly, this most recent news thrilled the players’ community, which all too clearly remembers the wounds AlphaStar inflicted in dominating renowned pros. Given that any match with the model is counted as a normal encounter and a ranking match, on-guard players try to spot and avoid AlphaStar lurking in the muddy waters of the battle.net rankings.

Players have reported “odd” behavior of some opponents and have been uploading videos on YouTube, where they discuss if the other player is actually AlphaStar incognito. They also advise each other to check if the partner responds to messages. Being called a noob by an opponent even once can be strong evidence that the opponent on the other side of the battlefield isn’t human.

So players are on the hunt for a replicant.

AI Monthly Digest #10 - AI tackles climate change and deciphers long-forgotten languages

AI Monthly Digest #10 – AI tackles climate change and deciphers long-forgotten languages

/in , , /by and

June brought record-breaking temperatures, perfectly highlighting the global challenge of climate change. Is that AI-related news? Check and see in the latest AI Monthly Digest.

A common misconception about machine learning projects is that they are by definition big.  However, any number of AI-powered micro-tweaks and improvements are applied in everyday work. A good example of both micro and macro tweaks that can fix a major problem can be found in the paper described below.

AI tackling climate change

The world witnessed an extraordinarily hot June, with average temperatures 2 degrees celsius above normal in Europe. According to the World Meteorological Organization, the heatwave is consistent with predictions based on greenhouse gas concentrations and human-induced climate change.

Tackling this challenge will not be easy: according to World Bank Data, fossil fuel energy consumption still stacks to 79% of total. Furthermore, greenhouse gasses, particularly methane, are emitted by cattle, with livestock being responsible of 14.5% of total human-induced greenhouse emissions.

The most prominent figures in AI today, including DeepMind CEO Demis Hassabis, Turing award winner Yoshua Bengio, and Google Brain co-founder Andrew Ng, have authored a comprehensive paper on ways that AI can tackle the changing climate.

Their call for collaboration is meant to inspire practitioners, engineers and investors to deliver short- and long-term solutions for measures within our reach. Those include producing low-carbon electricity through better forecasting, scheduling, and control for variable sources of energy, mitigating the damage produced by high-carbon economies through, for example, better predictive maintenance as well as help minimize energy use in transportation, smart buildings and cities. The applications can vary from designing grid-wide control systems or optimizing scheduling with more accurate demand forecasting.

Why does it matter

Climate change is one of the greatest challenges mankind faces today, with truly cataclysmic scenarios approaching. Further temperature increases may lead to a variety of disasters, from flooding coastal regions due to melting ice caps, agricultural crises and conflicts over access to water.

Green energy promises solutions, yet these are not without their challenges, many of which could be solved with machine learning, deep learning or reinforcement learning. Responsibility is among deepsense.ai’s most important AI trends, and being responsible for the planet would be an excellent example of just why we chose to focus on that trend.

We will provide more in-depth content on climate change and AI-powered ways of tackling it. So stay tuned!

Giants racing to produce the best image recognition

If machine learning is today’s equivalent of the steam engine revolution, data and hardware are the coal and engine that power the machines. Facebook and Google are like the coal mines of yesteryear, having access to large amounts of fuel and power to build new models and experiment.

It should come as no surprise that breakthroughs are usually powered by the tech giants. Google’s state of the art in image recognition, EfficientNet, has been a recent giant step forward. The model was delivered by automated searching procedure uniformly scaling each dimension of the network in order to find the best combination.

EfficientNet stands for something.

The result is state-of-the-art in Image recognition. At least when it comes to combining efficiency and accuracy. But not when it comes to accuracy alone.

Not even a month later Facebook delivered a model that outperformed Google’s. The key lay in scaling the enormous dataset it was trained on. The social media mogul has access to Instagram’s database, which holds no less than billions of user-tagged images, a dataset ready to be chewed over by a hungry deep learning model.

The neural network was released to the public using a recently launched Pytorch Hub platform for sharing cutting edge models.

Why does it matter

Both advances show how important machine learning is for the tech giants and how much effort they invest in pushing their research forward. Every advancement in image recognition brings new breakthroughs closer. For example, models are becoming more accurate in detecting diabetic retinopathy using images of the eye. Every further development delivers new ways to solve problems that would be unsolvable without ML (Machine learning) – manufacturing for visual quality control is among the best examples.

XLNet outperforms BERT

As we noted in a past AI Monthly Digest, Google has released Bidirectional Encoder Representation from Transformations (BERT). BERT was, until recently, the state-of-the-art when it comes to Natural Language Processing benchmarks. The newly announced XLNet is an autoregressive pretraining method (as opposed to an autoencoder-like BERT) which learns a language model by predicting the next word in a sequence using the permutation of all the surrounding words. An intuitive explanation can be found (here).

The XLNet model proved more effective than BERT in beating all 20 benchmark tasks.

Why does it matter

Understanding a natural language was considered a benchmark for intelligence, with Alan Turing’s test being among the best examples. Every push forward delivers new possibilities in building new products and solving problems, be they business ones or something more uncommon, like the example below.

AI-powered archeology? Bring it on!

Deep learning-based models are getting even better at understanding natural language. But what about language that is natural, but has never been deciphered due to lack of knowledge or a frustratingly small amount of extant text?

Recent research from MIT and Google shows that a machine learning approach can deliver major improvements in deciphering ancient texts. In the basics of modern natural language processing techniques, all of the words in a given text are assumed to be related to each other. The machine itself doesn’t “understand” text it in a human way, but rather forms its own assumptions based on the relations and connotations of each word in a sentence.

Disc of Phaistos, one of the most famous mysteries of archaeology

In this approach, the translation process is not built on understanding the world, but rather finding similarly connotated words that transfer the same message. This is entirely different than humans’ approach to language.

By making the algorithm less data-hungry, the researchers deliver a model that translates texts from rare and long-lost languages. The approach is described in this paper.

Why does it matter

While there are countless examples of machine learning in business, there are also new horizons to discover in the humanities. Deciphering the secrets of the past is every bit as exciting as building defenses against the challenges of the future.

The more sophisticated approach to and possible brute-force breaking of unknown languages provides a way to uncover more language-related secrets.

A Disc of Phaistos? Or a Voynich manuscript maybe?

Outsmarting failure. Predictive maintenance powered by machine learning

Outsmarting failure. Predictive maintenance powered by machine learning

/in , , /by

Since the days of the coal-powered industrial revolution, manufacturing has become machine-dependent. As the fourth industrial revolution approaches, factories can harness the power of machine learning to reduce maintenance costs.

The internet of things (IoT) is nothing new for industry.  Worldwide, the number of cellular-enabled factory automation devices reached 270 000 in 2012 worldwide. In 2018 it will rise to a staggering 820 000. Machines are present in every stage of the production process, from assembly to shipment. Although automation makes industry more efficient, with rising complexity it also becomes more vulnerable to breakdowns, as service is both time-consuming and expensive.

Four levels of predictive maintenance

According to PricewaterhouseCoopers, there are four levels of predictive maintenance.

1.Visual inspection, where the output is entirely based on the inspector’s knowledge and intuition
2.Instrument inspection, where conclusions are a combination of the specialist’s experience and the instrument’s read-outs
3.Real-time condition monitoring that is based on constant monitoring with IoT and alerts triggered by predefined conditions
4.AI-based predictive analytics, where the analysis is performed by self-learning algorithms that continuously tweak themselves to the changing conditions

As the study indicates, a good number of the companies surveyed by PwC (36%) are now on level 2 while more than a quarter (27%) are on level 1. Only 22% had reached level 3 and 11% level 4, which is basically level 3 on machine learning steroids. The PwC report states that only 3% use no predictive maintenance at all.

Staying on track

According to the PwC data, the rail sector is the most advanced sector of those surveyed with 42% of companies at level 4, compared to 11% overall.

One of the most prominent examples is Infrabel, the state-owned Belgian company, which owns, builds, upgrades and operates a railway network which it makes available to privately-owned transportation companies. The company spends more than a billion euro annually to maintain and develop its infrastructure, which contains over 3 600 kilometers of railway and some 12 000 civil infrastructure works like crossings, bridges, and tunnels. The network is used by 4 200 trains every day, transporting both cargo and passengers.

Outsmarting failure. Predictive maintenance powered by machine learning

According to the PwC data, the rail sector is the most advanced sector of those surveyed with 42% of companies at level 4, compared to 11% overall.

The company faces both technical and structural challenges. Among them is its aging technical staff, which is shrinking.

At the same time, the density of railroad traffic is increasing – the number of daily passengers has increased by 50% since 2000, reaching 800 000. What’s more, the growing popularity of high-speed trains is exerting ever greater tension on the rails and other infrastructure.

To face these challenges, the company has implemented monitoring tools, such as sensors for monitoring overheating tracks, cameras which inspect the pantographs and meters to detect drifts in power consumption, which usually occur before mechanical failures in switches. All of the data is collected and analyzed by a single tool designed to apply predictive maintenance. Machine learning models are a component of that tool.

As sounding brass

Mueller Industries (Memphis, Tennessee) is a global manufacturer and distributor of copper, brass, aluminum and plastic products. The predictive maintenance solution the company uses is based on sound analysis. Every machine can be characterized by the sound it makes and any change in the tone or the sounds it makes may be a sign of impending malfunction. The analysis of the sound and the vibrations of the machine is done in real-time with the cloud-based machine learning solution that seeks patterns in the data gathered.

Both the amount and the nature of the data collected render it impossible for a human to analyze, but a machine-learning powered AI solution handles it with ease. The devices are able to gather data in ultrasonic and vibration sensors and analyze them in real time. Contrary to experience-based analytics, using the devices requires little-to-no training and can be done on the go.

Endless possibilities

With the power of machine learning enlisted, handling the tremendous amounts of data generated by the sensors in modern factories becomes a much easier task. It allows the company to detect failures before they paralyze the company, thus saving time and money. What’s more, the data that is gathered can be used to further optimize the company’s performance, including by searching for bottlenecks and managing workflows.

That’s why 98% of industrial companies expect to increase efficiency with digital technologies.

AI Monthly Digest #9 – the double-edged sword of modern technology

AI Monthly Digest #9 – the double-edged sword of modern technology

/in , , /by and

This edition is all about AI morality-related themes, with a slight tinge of Talking Heads and Modern Talking.

Earlier this year, deepsense.ai highlighted AI morality and transparency as one of 2019’s dominant AI trends. May bore out our thesis, especially as it relates to potential misuse and malicious intent.  At the same time, though, AI provides unique chances to support entertainment and education, as well as deliver new business cases.

A bigger version of GPT-2 released to the public

Open-AI has recently shown the GPT-2 model has set a new gold standard for natural language processing. Following the acclaimed success of the model, OpenAI opted not to make it public due to the risk of malicious usage, particularly to produce spam and fake news at no cost.

This sparked an uproar. The industry good practice is to release AI research work as open-source software, so other researchers can push the boundaries further without having to repeat all the work done earlier from scratch. In other words – OpenAI threw up a major hurdle to NLP-model development by keeping GPT-2 under wraps.

To support the scientific side of the equation while reducing the malicious threat, OpenAI releases some smaller-scale models to the public. The model it recently released operates on 345M parameters, while the best original model consists of 1.5B parameters. Every parameter can be seen as a virtual neuron inside a neural network, so OpenAI is basically reducing the brain it designed.

The original network was released to OpenAI partners currently working on malice-proofing the system. The first independent applications of the downscaled network are already available at talktotransformer.com and onionbot headline generator.

Why does it matter?

OpenAI is currently facing a difficult choice between supporting the global development of AI and the fear of losing control over dangerous technology. In a world facing a potential avalanche of fake news and social media being used to perpetuate propaganda, building a system that writes coherent and convincing texts is undoubtedly dangerous.

This case allows one to see all the AI-related issues in a nutshell, including the technology’s amazing potential, the real threat of misuse or malicious intent. So the case may serve as a precedent for future cases.

Talking heads unleashed

A group of scientists working for Samsung’s AI Center in Moscow and Skolkovo Institute of Science and Technology designed a model that can produce a convincing video of a talking head from a single image, such as a passport photo or even a painting.

The model renders with consistency both the background and the head’s behavior. Most impressively, the model builds a convincing video of a talking head from even a single image of the frame.

The solution is searching for a similar face that was analyzed and extracts facial features including a nose, chin, mouth and eyes. The movement of those features is then applied on the image, as shown in the video.

The results are undoubtedly impressive.

Why does it matter?

Yet another AI ethics-related issue, the talking-head technology poses the threat of deepfakes, images that show a person making statements that he or she would never make. This raises obvious questions about the malicious ways such technology could be used.

On the other hand, when deepfakes are used for special effects in popular movies, no one seems to complain and critics even weigh in with their acclaim. Some of the better-known examples come from the Star Wars franchise, particularly Rogue One, which features Leia Organa wearing the face of a young Carrie Fisher.

AI has also proved itself useful in promoting art. By leveraging this technology it is possible to deliver the talking head of Girl with a Pearl Earring or the Mona Lisa telling visitors from screens about a painting’s historical context – a great way to put more fun in art lessons for kids. Or just to have some fun seeing what a Stallone-faced Terminator would look like.

Again, AI can be used for both good and evil ends. The ethics are up to the wielder of this double-edged sword.

Modern Talking – recreating the voice of Joe Rogan

Another example of deepfake-related technology is using AI to convincingly recreate Joe Rogan’s voice. The text-to-speech technology is not a new kid on the block, yet it is easy to spot due to the robotic and inhumanely calm style of speaking. Listening to automated text-to-speech was usually boring at best while delivering the unintentional comic effects of robotic speech, all in the absence of emotion or inflection.

Dessa engineers have delivered a model that is not only transforming text to speech, but also recreating Joe Rogan’s style of speaking. Joe is a former MMA commentator who went on to become arguably the most popular podcaster in the world. Speaking with great emotion, heavily accenting and delivering power with every word, Rogan is hard to mistake.

Or is he? The team released a quiz that challenges the listener to distinguish if a given sample comes from a real podcast or was AI-generated. The details can be found on Dessa’s blog.

Why does it matter?

Hearing a convincing imitation of a public personality’s voice is nearly as unsettling as watching a talking head talk. But the technology can be used for entertainment and educational purposes. For example, delivering a new Frank Sinatra single or presenting Winston Churchill’s comprehensive and detailed speech on reasons behind World War II.

Again, the ethics are in the user’s hands, not in the tool. Despite that, and as we saw with OpenAI’s GPT-2 Natural Language Processing model, researchers have decided NOT to let the model go public.

Machine learning-powered translations increase trade by 10,9%

Researchers at Olin Business School at Washington University in St.Louis have found a direct connection between machine learning-powered translations and business efficiency. The study was conducted on e-Bay and shows that moderate improvement in the quality of language translation increased trade between countries on eBay by 10.9%.

The study examined the trade between English speakers from the United States and their trade relations with countries speaking other languages in Europe, America and Asia. More on the research can be found on the Washington University of St.Louis website.

Why does it matter?

While there is no doubt that AI provides vital support for business, the evidence, while voluminous, remains largely anecdotal (sometimes called anec-data) with little quantitative research to back up the claim. Until the Olin study, which does provide hard and reliable data. Is justified true belief knowledge? That’s an entirely different question…

A practical approach to AI in Finland

AI Monthly Digest #5 presented a bit about a Finnish way of spreading the word about AI. Long story short: contrary to many approaches of building AI strategy in a top-down model, Finns have apparently decided to build AI-awareness as a grassroots movement.

To support the strategy, the University of Helsinki has released a digital AI course on the foundations and basic principles of AI. It is available for free to everyone interested.

Why does it matter?

AI is gaining attention and the reactions are usually polarised – from fear of losing jobs and machine rebellion to arcadian visions of an automated future with no hunger or pain. The truth is no doubt far from either of those poles. Machine learning, deep learning and reinforcement learning are all built on certain technological foundations that are relatively easy to understand, including their strengths and limitations. The course provides good basic knowledge on these issues, which can do nothing but help our modern world.

A comprehensive guide to demand forecasting

A comprehensive guide to demand forecasting

/in , , , /by and

Everything you need to know about demand forecasting – from the purpose and techniques to the goals and pitfalls to avoid.

Essential since the dawn of commerce and business, demand forecasting enters a new era of big-data rocket fuel.

What is demand forecasting?

The term couldn’t be clearer: demand forecasting forecasts demand. The process of predicting the future involves processing historical data to estimate the demand for a product. An accurate forecast can bring significant improvements to supply chain management, profit margins, cash flow and risk assessment.

What is the purpose of demand forecasting?

Demand forecasting is done to optimize processes, reduce costs and avoid losses caused by freezing up cash in stock or being unable to process orders due to being out of stock. In an ideal world, the company would be able to satisfy demand without overstocking.

Demand forecasting techniques

Demand forecasting is an essential component of every form of commerce, be it retail, wholesale, online, offline or multichannel. It has been present since the very dawn of civilization when intuition and experience were used to forecast demand.

Sybilla – deepsense.ai’s demand forecasting tool

More recent techniques combine intuition with historical data. Modern merchants can dig into their data in a search for trends and patterns. At the pinnacle of these techniques, are demand forecasting machine learning models, including gradient boosting and neural networks, which are currently the most popular ones and outperform classic statistics-based methods.

The basis of more recent demand forecasting techniques is historical data from transactions. These are data that sellers collect and store for fiscal and legal reasons. Because they are also searchable, these data are the easiest to use.

Sybilla – deepsense.ai’s demand forecasting tool

How to choose the right demand forecasting method – indicators

As always, selecting the right technique depends on various factors, including:

  • The scale of operations – the larger the scale, the more challenging processing the data becomes.
  • The organization’s readiness – even the large companies can operate (efficiency aside) on fragmented and messy databases, so the technological and organizational readiness to apply more sophisticated demand forecasting techniques is another challenge.
  • The product – it is easier to forecast demand for an existing product than for a newly introduced one. When considering the latter, it is crucial to forming a set of assumptions to work on. Owning as much information about the product as possible is the first step, as it allows the company to spot the similarities between particular goods and search for correlations in the buying patterns. Spotting an accessory that is frequently bought along with the main product is one example.

How demand forecasting can help a business

Demand forecasting and following sales forecasting is crucial to shaping a company’s logistics policy and preparing it for the immediate future. Among the main advantages of demand forecasting are:

  • Loss reduction – any demand that was not fulfilled should be considered a loss. Moreover, the company freezes its cash in stock, thus reducing liquidity.
  • Supply chain optimization – behind every shop there is an elaborate logistics chain that generates costs and needs to be managed. The bigger the organization, the more sophisticated and complicated its inventory management must be. When demand is forecast precisely, managing and estimating costs is easier.
  • Increased customer satisfaction – there is no bigger disappointment for consumers than going to the store to buy something only to return empty-handed. For a business, the worst-case scenario is for said consumers to swing over to the competition to make their purchase there. Companies reduce the risk of running out of stock–and losing customers–by making more accurate predictions.
  • Smarter workforce management – hiring temporary staff to support a demand peak is a smart way for a business to ensure it is delivering a proper level of service.
  • Better marketing and sales management – depending on the upcoming demand for particular goods, sales and marketing teams can shift their efforts to support cross- and upselling of complementary products,
  • Supporting expert knowledge – models can be designed to build predictions for every single product, regardless of how many there are. In small businesses, humans handle all predictions, but when the scale of the business and the number of goods rises, this becomes impossible. Machine learning models extend are proficient at big data processing.

How to start demand forecasting – a short guide

Building a demand forecasting tool or solution requires, first and foremost, data to be gathered.

While the data will eventually need to be organized, simply procuring it is a good first step. It is easier to structure and organize data and make them actionable than to collect enough data fast. The situation is much easier when the company employs an ERP or CRM system, or some other form of automation, in their daily work. Such systems can significantly ease the data gathering process and automate the structuring.

Sybilla – deepsense.ai’s demand forecasting tool

The next step is building testing scenarios that allow the company to test various approaches and their impact on business efficiency. The first solution is usually a simple one, and is a good benchmark for solutions to come. Every next iteration should be tested to see if it is performing better than the previous one.

Historical data is usually everything one needs to launch a demand forecasting project, and obviously, there are significantly less data on the future. But sometimes it is available, for example:

  • Short-term weather forecasts – the information about upcoming shifts in weather can be crucial in many businesses, including HoReCa and retail. It is quite intuitive to cross-sell sunglasses or ice cream on sunny days.
  • The calendar – Black Friday is a day like no other. The same goes for the upcoming holiday season or other events that are tied to a given date.

Sources of data that originate from outside the company make predictions even more accurate and provide better support for making business decisions.

Common pitfalls to avoid when building a demand forecasting solution

There are numerous pitfalls to avoid when building a demand forecasting solution. The most common of them include:

  • The data should be connected with the marketing and ads history – a successful promotion results in a significant change in data, so having information about why it was a success makes predictions more accurate. If machine learning was used to make the predictions, the model could have misattributed the changes and made false predictions based on wrong assumptions.
  • New products with no history – when new products are introduced, demand must still be estimated, but without the help of historical data. The good news here is that great strides have been made in this area, and techniques such as product DNA can help a company uncover similar products its past/current portfolio. Having data on similar products can boost the accuracy of prediction for new products.
  • The inability to predict the weather – weather drives demand in numerous contexts and product areas and can sometimes be even more important than the price of a product itself! (yes, classical economists would be very upset). The good news is that even if you are unable to predict the weather, you can still use it in your model to explain historical variations in demand.
  • Lacking information about changes – In an effort to support both short- and long-term goals, companies constantly change their offering and websites. When the information about changes is not annotated in the data, the model encounters sudden dwindles and shifts in demand with apparently no reason. In the reality, it is usually a minor issue like changing the inventory or removing a section from website.
  • Inconsistent portfolio information – predictions can be done only if the data set is consistent. If any of the goods in a portfolio have undergone a name or ID change, it must be noted in order not to confuse the system or miss out on a valuable insight.
  • Overfitting the model – a vicious problem in data science. A model is so good at working on the training dataset that it becomes inflexible and produces worse predictions when new data is delivered. Avoiding overfitting is down to the data scientists.
  • Inflexible logistics chain – the more flexible the logistics process is, the better and more accurate the predictions will be. Even the best demand forecasting model is useless when the company’s logistics is a fixed process that allows no space for changes.

Sybilla – deepsense.ai’s demand forecasting tool

Summary

Demand and sales forecasting is a crucial part of any business. Traditionally it has been done by experts, based on know-how honed through experience. With the power of machine learning it is now possible to combine the astonishing scale of big data with the precision and cunning of a machine-learning model. While the business community must remain aware of the multiple pitfalls it will face when employing machine learning to predict demand, there is no doubt that it will endow demand forecasting with awesome power and flexibility.

Machine learning in drug discovery

/in , /by

Artificial intelligence is advancing various industries, including healthcare and the pharmaceutical industry. According to Accenture data, key clinical health AI applications can potentially create $150 billion in annual savings for the United States healthcare sector by 2026.

The numbers show that the healthcare industry will heavily leverage the possibilities provided by machine learning. That’s why AI companies are getting involved in various activities in the treatment process, from diagnosis to therapy and drug development.

By applying convolutional neural networks in detecting diabetic retinopathy, deepsense.ai significantly improved the diagnostic process by speeding up and automating diabetic retinopathy screenings. The next step may be building a reinforcement learning agent that can be trained to run by controlling the muscles attached to the virtual skeleton. With that doctors can predict if a patient is able to walk, jump or run properly after the treatment. Furthermore, the work done during the research might be later used to design new, AI-powered leg prostheses.

Another healthcare segment that is heavily dependent on data is drug discovery.

The potential of AI in drug discovery

Computational solutions in drug discovery help significantly reduce the cost of introducing drugs to the market. Grand View Research and its new 2018 report implies that global drug discovery informatics market size was estimated at $713.4 million in 2016 and it is anticipated to progress at a CAGR (Compound Annual Growth Rate) of 12.6% by 2025. With artificial intelligence being used in drug discovery, the market’s value is growing rapidly. In its Global Artificial Intelligence in Drug Discovery Market Size Analysis, 2018-2028, Bekryl indicates that AI has the potential to create $70 billion in savings in the drug discovery process by 2028.

The technological and paradigm shift to machine learning seen in the pharmaceutical industry enables researchers to use novel computational algorithms to support the process. As biomedical data are highly complex, using algorithms in designing new drugs has become more possible than it has ever been. Machine learning can enhance many stages of the drug discovery process:

There are different kinds of data, including genetic and imaging ones. Each of them can be analyzed with machine learning and further used to build novel solutions for drug discovery.

Challenges in machine learning for drug discovery

Ensuring drug safety is one of the main challenges in the drug discovery process. Interpreting information of the known effects of drugs and predicting their side effects are complex tasks. Scientists and engineers from research institutions and pharmaceutical companies like Roche and Pfizer have been trying to use machine learning to get meaningful information from clinical data obtained in clinical trials. Interpretation of this data in the context of drug safety is an active area of research.

Clinical trials are the most expensive stage of drug development. To reduce their costs, it is crucial to use the experience gained during previous clinical trials in the early stages of drug development. This can be achieved in two steps:

  • biomedical data from research experiments could be analyzed and interpreted using machine learning to predict a drug’s effects and side effects;
  • data from clinical trials analyzed with machine learning should support the interpretation of biological data.

With those two approaches developed simultaneously, it is possible to design better preclinical experiments to come up with the most effective therapies with the fewest side effects.

Integrating biomedical data with computational approaches

Machine learning could help optimize therapy by integrating biomedical and clinical data with computational models, and can be used to build software to test drugs and combinatorial therapies. Some computational models and approaches which support the integration of clinical data are still under development but there are also a few very good examples of successful data integration in biology and medicine.

For example, there are a number of machine learning methods for integrating genetic regulatory networks and pathway information. This can be used to predict their biological functions and efficient Python-based implementation of bioinformatic tools and approaches that are easy to interface with broadly used machine learning packages.

Genetic data analysis and personalized medicine

Many pharmaceutical companies and startups are focused on genetic data interpretation and personalized medicine. Understanding the patient’s genetic profile helps to offer appropriate drugs and therapy. Building computational approaches to analyze genetic data and propose novel therapies could be advanced with machine learning. There are only a few examples that impact current clinical practice based on machine learning solutions which bring huge potential to personalized medicine and drug discovery. They include discovering novel biomarkers of drug response and machine learning-based computational tools used in clinical practice. Such tools are used to estimate the resistance to individual drugs and to combinatorial therapies based on genotype analysis.

One of the possible approaches is based on interpreting the genetic code as a one dimensional image and then applying a standard machine learning algorithm. The data is then scoured for patterns and anomalies, just as has been done in various other deepsense.ai image recognition projects. Analyzing the genomics may be in fact done in the same way as it is applied to classical paintings, when it comes to finding a hand or any other element. For the algorithm, the nature or shape of an image to analyze is irrelevant, so the machine is equally effective at analyzing a one-dimensional DNA chain or any other type of image data

Because genomic data is usually presented as a string of letters, it is also possible to apply Natural Language Processing techniques. One advantage of doing so is that it broadens the area the algorithm is able to process. That may be important when particular changes or patterns are being sought, or the pattern to find consists of a longer sequence of genes.

A big challenge is to fully unlock the potential of machine learning for drug discovery and personalized medicine. Time series data could be useful to fully reconstruct genetic networks on the basis of expression data. To build comprehensive predictive models based on machine learning, expression genetic data and sequencing data should be acquired in time series.

Innovative startups, like Cambridge Cancer Genomics, use machine learning to analyze data gained from liquid biopsy, a diagnostic technology in which circulating tumor cells or cell-free DNA is collected from blood samples. Although it is not a fully standardized approach for cancer therapy monitoring, it is highly anticipated in personalized medicine due to its ability to acquire genetic data in time series during treatment. Applying machine learning to better understand those data and to answer the question of why cancer evolves could help scientists design less toxic therapies.

Building and getting insight from databases and datasets

Scientists use public repositories of clinical data to tackle big problems in clinics to help medical doctors in their everyday work, as medical knowledge can be extracted from public repositories. These repositories could also be used for drug discovery purposes to include clinical information in the early stage of drug development.

Attempts have been made to represent medical knowledge using deep neural networks. Data mapped with machine learning might also be easier to integrate with biomedical data analyzed with machine learning, thanks to better compatibility in the data structures generated with similar approaches.

New achievements in building databases for machine learning purposes are also promising. For example, the authors of the paper “integrative analysis and machine learning on cancer genomics data using the Cancer Systems Biology Database (CancerSysDB)” developed a database for highly flexible queries and analysis of cancer-related data across multiple data types and multiple studies. However, there are many problems in medicine and drug discovery which are very difficult to answer only on the basis of public data, of which there is a paucity if better machine learning models and approaches are to be developed.

If proper datasets are to be built to answer specific scientific questions, it is not only the way in which data is preprocessed that must be understood, but also the principles of using different bioinformatic tools and interdisciplinary knowledge in biomedicine and where computer science and medicine converge. Teams which have this knowledge and skills could help to make better use even of limited amounts of data from public repositories. Machine learning engineers usually get data from scientists, medical doctors, pharmaceutical companies and hospitals, thus the amount is limited. But models must be strong for results to be achieved.

One of the best examples of designing a model of superior strength, one that can deal with the lack of proper data, was deepsense.ai’s Right Whale Recognition engine. The model was designed to recognize an individual Right Whale in a photograph, even if there were only a few photos provided in the dataset.

To get deep insight from data, close cooperation and mutual understanding of different languages and disciplines is needed. That is difficult if there is only the occasional consultation.

Scientists with a dual biomedical and computational background are crucial members of teams building databases, datasets, machine learning models, tools and software for analyzing biomedical data and drug discovery. Leading institutions like ETH Zurich have already started educating a new generation of medical scientists with a computational and math background and have built a platform and interdisciplinary teams to analyze clinical and biomedical data. The Swiss tumor board and ETH Personalized Health Technologies Platform Nexus are actively working towards implementing individualized, biomarker-based medical decisions in clinical practice. These are crucial fundamental steps in advancing drug discovery with machine learning.

Standard machine learning approaches for genetics and genomics

Standard supervised, semi-supervised and unsupervised machine learning algorithms are applied to analyze genetic data like microarray or RNA-seq expression data. To understand how, read “machine learning in genetics and genomics”. These algorithms can reveal disease and healthy phenotypes and could be further used to uncover the mechanisms of action of drugs. In any application of machine learning methods, the researcher must decide which data to provide as input to the algorithm to answer complex biomedical questions.

There are a number of comprehensive reviews summarizing the use of large-scale analysis of genomic data and machine learning strategies to solve genomic sequencing problems, like finding specific regions in sequences and recognizing locations of transcriptomic sites. It is one of the biggest challenges in genomics with practical applications.

Machine learning has potential for this application, though the results produced with machine learning algorithms should be validated with data from laboratory experiments or clinical trials. Deep learning algorithms could be useful in genome interpretation and analysis of genetic variants,  a complex task that requires a combination of robust biological data and clinical knowledge.

Recently scientists and engineers have taken a step toward better understanding the human genome thanks to machine learning. Supervised heterogeneous ensemble methods can significantly improve our ability to address difficult biomedical prediction problems. Still, the application of machine learning algorithms to genomic problems is in a nascent stage. After all, genomic and genetic data are multidimensional and there remains a need to develop probabilistic machine learning algorithms for their analysis.

Machine learning approaches for network analysis of biomedical data

Analysis of genetic data could be helpful in elucidating genetic networks, which can reveal a drug’s mechanism of action and help understand how diseases work. This falls within the scope of an emerging new discipline called network medicine. The Barabasi group, a pioneer in network medicine, states that an unsupervised network-based approach enables the prediction of novel drug-disease associations, which offer significant opportunities for finding new applications for drugs and predicting potential side effects.

The group also found that the therapeutic effect of drugs might be localized in a small network neighborhood. This means that several genes in close network proximity of genes related to the mechanism of a disease could be targeted to effectively treat the disease.

Analyzing genetic network data with machine learning could help in finding novel targets for drugs and predict the optimal combination of drugs. There are research papers that explain how to benchmark machine learning for biological network analysis. One is “machine learning-assisted network inference approach to identify a new class of genes that coordinate the functionality of cancer networks.” This study shows usage of support vector machine (SVM) models combined with machine learning-assisted network inference (MALANI) to identify cancer-associated gene pairs. These can be used to reconstruct cancer networks to identify key cancer genes in high-dimensional data space that would otherwise go undetected by conventional approaches. These algorithms should be equally applicable to other machine learning and feature selection approaches. There is also a tutorial by Stanford lecturers which shows the basics of how to use deep learning approaches to analyze biological networks. However, for analysis of complex biological networks, non-standard machine learning algorithms are still being developed and network and machine learning approaches need better integration.

Machine learning algorithms in image analysis for drug discovery

The article Machine learning and image-based profiling in drug discovery presents how image-based screening of high-throughput experiments, in which cells are treated with drugs, could help elucidate a drug’s mechanism of action. It is mentioned that unsupervised and simple statistical inference methods seem to be in favor for analyzing image data from large-scale profiling experiments, but complex biological phenotypes and single-cell experiments could be successfully classified with supervised algorithms.

The recently explored application of supervised learning in image-based profiling, particularly deep neural networks, might be a novelty detection framework to identify unexpected phenotypes revealed in the drug discovery process. With deep learning it is possible to predict the properties of a molecule only from its structure. The technique requires using a convolutional neural network that is able to extract the shape of a molecule and then confront it with the information gathered about the properties.

Novel machine learning algorithms under way

Research on quantum machine learning shows that this approach should be useful for finding complex patterns in data. As biological and medical data are complex, probabilistic quantum machine learning algorithms represents a real opportunity to understand them better. Innovative pharmaceutical companies like Amgen or startups like ProteinQure have moved to apply quantum computing and quantum machine learning to drug discovery, while focusing these efforts mainly on predicting the structure of new drugs. Finally, genomics and systems biology are two important areas in which novel machine learning algorithms can be applied with a view to producing less toxic drugs based on the profound analysis of biomedical data.

The text was written in collaboration with Anna Kornakiewicz, an independent data scientist and researcher – blog, github as a consultant.

Ai-ethics-artificial-imagination

AI Monthly Digest #6 – AI ethics and artificial imagination

/in , , /by and

February came with the groundbreaking event of building a new state-of-the-art natural language processing neural network. And then the unthinkable happened.

Machine learning has already shown that it is a world-transforming technology used in various appliances, including drug discovery, saving endangered species and designing software for sophisticated prosthetic legs. Therefore it is natural that this should spark debates, both ethical and business-related. What reaction has there been to this in February?

1. Open-AI designed a new gold standard for natural language processing…

While in image processing artificial intelligence models outperform humans in most tasks, natural language processing (NLP) is still a challenge for machines to master. Although AI-based services are already usable with Google Translate being the pinnacle of today’s achievements, the texts produced by machines is easy to recognize by humans. But this could change soon with the latest state-of-the-art from OpenAI.

They designed a model called GPT-2 and prepared a dataset containing 8 million web pages for training. The network objective is to predict the next word given all the words to that point, which is the simplest way of doing unsupervised learning in NLP. The main improvements, in this case, are scaling up the model size (1.5B parameters) and training it on the gargantuan database of Internet text (40GB) at an unprecedented scale (32 x TPUv3).

The resulting network analyzes the starting word of a sentence and then it adds the next words to create a text based on “the most probable output”. The effect is surprisingly accurate. The model is able to recognize the type of the starting text. If the first sentence is a press title it will produce a legitimate-sounding short news article. It is also good at adopting various styles – something unimaginable for previously sturdy models which could produce problematic sentences.

Example:

Human written:

In a shocking finding, scientist discovered a herd of unicorns living in a remote, previously unexplored valley, in the Andes Mountains. Even more surprising to the researchers was the fact that the unicorns spoke perfect English.

Model completion (machine-written, 10 TRIES)

The scientist named the population, after their distinctive horn, Ovid’s Unicorn. These four-horned, silver-white unicorns were previously unknown to science.

Now, after almost two centuries, the mystery of what sparked this odd phenomenon is finally solved.

Dr. Jorge Pérez, an evolutionary biologist from the University of La Paz, and several companions, were exploring the Andes Mountains when they found a small valley, with no other animals or humans. Pérez noticed that the valley had what appeared to be a natural fountain, surrounded by two peaks of rock and silver snow.

Pérez and the others then ventured further into the valley. “By the time we reached the top of one peak, the water looked blue, with some crystals on top,” said Pérez.

Pérez and his friends were astonished to see the unicorn herd. These creatures could be seen from the air without having to move too much to see them – they were so close they could touch their horns.

While examining these bizarre creatures the scientists discovered that the creatures also spoke some fairly regular English. Pérez stated, “We can see, for example, that they have a common ‘language,’ something like a dialect or dialectic.”

More examples of AI-generated texts may be found at Open-AI blog.

Moreover – the network is capable of transfer learning – it is easier to perform a slight additional training to enable the model to perform more sophisticated tasks. Transfer learning is a gateway technology for modern image recognition appliances, including art recognition and visual quality control.

Why does it matter

Alan Turing, the godfather of modern information technology, was heavily convinced that the ability to understand language is a key indicator of intelligence. Natural language, the emotions hidden beneath the words and all the cultural and societal contexts behind sentences are truly unique for human beings.

Making this part of the world approachable for machines is undoubtedly the great breakthrough and passes countless possibilities, from battling hate crime to various business appliances.

2. … and they didn’t release it to the public

With great power comes great responsibility and the events from February 2019 show it clearly. Despite best practices and established custom, Open-AI decided NOT TO release the state-of-the-art neural network to the public as an open source software.

The organization did this to prevent the neural network from being used for malicious purposes, with perpetuating fake news as a top concern. The decision sparked a debate on Reddit, Twitter and other various media, where participants argued on the safety of releasing this “potentially dangerous technology” and tackling the established model of open-sourcing the effect of research. Researchers are afraid of building the doorway for “deepfakes for texts”.

Why does it matter

AI ethics and responsibility was highlighted as one of the major AI trends 2019. The situation where researchers choose safety over progress is unusual. On the other hand, fake news is considered one of the major threats of the future and was recognized one of the most dangerous online activities in 2018 by the World Economic Forum.

With the rise of autonomous cars and IoT revolution building a frame for sophisticated AI-powered solutions to work, such dilemmas may get increasingly common.

3. China’s AI strategy at a glance (or even more)

In the last AI monthly digest, we covered the AI strategy forged in Finland, where the national pursuit for the AI appliances grew from a grassroots movement. In contrast, China has a centralized strategy to make the country AI-giant.

The Center for a New American Security (CNAS) has published a report in which they provide deeper insights into understanding China’s AI strategy. The report is a long read written by Gregory C. Allen who had a chance to meet a few times with high-ranking Chinese officials on conferences focusing on Artificial Intelligence.

Topics covered in the text:

  • Chinese views on importance and security of AI
  • Strengths and weaknesses of China’s AI Ecosystem
  • China’s short-term goals in AI
  • The role of semiconductors

The author concludes that gaining expertise in and understanding of AI developments in China should help U.S. policymakers sort out their priorities. Instead of influencing China’s competitiveness, they should focus on boosting the technological and economic competitiveness of the United States.

Why does it matter

Early adopters are usually those who earn knowledge with first failures. Thus, for now, it is wise to observe and learn.

4. Introducing Paperswithcode

Staying updated on the latest developments in machine learning is not an easy task, especially considering the fact that not all the researchers decide to make their code available. Releasing the code is a good practice, but providing only the paper that “describes the matter enough” to reproduce the effect is also correct.

Using the paperswithcode makes the process of paper-and-code matching much easier, effectively saving a lot of time for people willing to broaden their knowledge.

Why does it matter

The portal itself is a handy tool to gain knowledge and, a bit unwillingly promotes the constituted approach of making the research code public. If it gains notoriety, it may be a tool of social pressure to publish code with the research.

5. Popularity of thispersondoesnotexist.com

The neural networks’ ability to create convincing human faces is one of the benchmarks for modern AI appliances. To make the images available for the public, the StyleGan-based model publishes random face on thispersondoesnotexist.com website.

As in every aspect of life, the devil is in the details. The website is a great tool to spot the weakest parts of generated faces.

  • Hair – it is common to place some hair in the air, with no attachment to other or floating in an unnatural way. Sometimes the skin around hair is “sticking” to them, creating some kind of scars that are disturbing when seen. The model sometimes messes up hairstyles, mixing dreadlocks with straight and curly hair, but it is more challenging to spot
  • Teeth – algorithms tend to mess up with teeth, placing them unnaturally, sometimes merged or blurred.
  • Glasses – it is common that the eyes inside the glasses do not fit the rest of the depicted person. Sometimes it means placing eyes of an old woman, surrounded with wrinkles, on a child’s face.

Why does it matter

Making these images publicly available with commentary, which all have been made by neural networks, is an interesting way to spread knowledge about AI-related topics. The concerns about deep fakes and other illicit content created with the support of artificial intelligence are mostly based on lack of knowledge – the AI-based techniques are seen as omnipotent. But they aren’t, as can be easily seen by the example of teeth and hair.

Moreover, the initiative is another way of encouraging people to verify news sources and distrusting information and tackling the challenge of fake news and deep fakes, mentioned above.

And a bonus one:

deepsense.ai in cooperation with Google Brain has finished research on enabling reinforcement learning (RL) agents to make predictions about their upcoming actions to reduce the number of actions needed to train skills.

RL agents typically need incomparably more actions to train skill than humans and the ability to build the skill “in a head” may be a part of the answer. Starting from childhood, people repeat their actions in their minds when they seek perfection or just recall an enjoyable or pleasant time.

Following this idea, our researchers have designed neural networks that possess the imagination that enables the model to simulate the action before executing it. A more detailed story about these projects and their outcome may be found in our blogpost about artificial imagination and Arxiv paper.

Why does it matter?

Artificial imagination is basically the idea of building the world simulated only by the mind’s power. Our researchers were able not only to reduce training time, but also to replace the simulated environment with another neural network’s imagination.

Maintaining and providing the simulated environment is one of the highest costs in building reinforcement learning agents, so tackling this challenge may be a great step to making this technique more popular.

This site uses cookies. By continuing to browse the site, you are agreeing to our use of cookies.

OKLearn more