deepsense.ai, Google Brain, the University of Warsaw and the University of Illinois at Urbana-Champaign have concluded a collaborative research project, building neural networks that mimic a simulated environment and effectively enabling artificial intelligence to perform a simulation.

Designed models can be applied in text or video prediction and to solve the problems where mathematical description may be overly complicated.

Reinforcement learning (RL) enables a neural network to develop and hone its skill by interacting with its environment, which is usually a simulated one. The technique is predicted to shine in robotics and the building of autonomous vehicles.

The research team found a way to build a neural network that can mimic the signals that the RL agent usually gets from interacting with the environment. The neural network produces signals that usually would be taken from sensors, like images. “This is one of the important ideas of reinforcement learning. A recent survey done by DeepMind’s J.B. Hammrick provides a thorough account of analogies between model-based reinforcement learning and mental simulation as considered by cognitive science”, says Henryk Michalewski, the R&D Coordinator at deepsense.ai and a visiting professor in the Computer Science Department of the University of Oxford.

The challenges of education

In the learning process, the RL agent is rewarded for performing a task correctly and punished for making mistakes. Autonomous cars provide perhaps the best example–the agent is rewarded for safe driving and punished for collisions and speeding. The model seeks to maximize the rewards and minimize the punishments. Initially, all actions are random and the neural networks have yet to explore the possible ways to perform the tasks. It is thus possible that more than the first dozen rides will end with a collision with the nearest wall as the model incurs punishment and learns how to brake or avoid it.

This leads to the idea of conducting the trial and error process inside a simulator.

However, building a realistic simulator is a tedious problem for humans– says Błażej Osiński, a Senior Data Scientist at deepsense.ai. For example, Unreal Engine 4 has several million lines of code. The idea of our project is to let the neural network learn how to simulate the environment. A similar approach was suggested in a recent work from Yann LeCun’s lab, where neural networks were employed to simulate dense traffic.

The research team has built a neural network that emulates the Atari gaming environment, a popular training ground for reinforcement learning models. The network was able to create a version of the games Pong, Freeway and others that were nearly indistinguishable from Atari’s.

Ongoing research

The project was run by data scientists and researchers from deepsense.ai including Henryk Michalewski, Piotr Miłoś, Błażej Osiński and fellow researchers from Google Brain, the University of Warsaw and the University of Illinois at Urbana-Champaign. More detailed information about the research, outcomes and possible uses can be found in the Arxiv paper and a detailed blogpost about artificial imagination on deepsense.ai’s blog.

The research opens up interesting ways to apply neural networks in business settings. The first is to build an RL agent that can explore a highly complex environment and then emulate it for the needs of models that are to perform tasks within it. The best example may be teaching the model to emulate the real world in all of its complexity and unpredictability.

Another business application would be a video prediction tool that could deliver videos based on only a few frames, effectively reducing the effort animators and designers put into producing full-length videos.

We consider our research work an essential part of deepsense.ai and a key aspect of the company’s development. This time our researchers have really pushed the boundaries of knowledge with no marketing overstatements. The project brings fresh, new idea into AI research, a thing that is worth contributing into.– concludes Tomasz Kułakowski, CEO at deepsense.ai.

###

About deepsense.ai:

brand-manager-at-deepsense

AI World is an artificial intelligence and machine learning conference being held in Boston, December 3-5. deepsense.ai will once again be providing its experience and expertise as a Strategic Deep Learning Partner at the conference.

AI World Conference and Expo annually gathers artificial intelligence (AI) professionals and practitioners to explore the state of practice of machine learning and deep learning in enterprise and the possible applications of this innovative technology. deepsense.ai thas been a strategic partner since the first conference and this year will be supporting the event for the third time.

The AI World conference is organized by AI Trends, the leading industry media channel focused on the business and technology of AI. Since its merger with Cambridge Innovation Institute, AI World has been focused on machine learning-based advances in the life sciences, especially for the healthcare and pharmaceutical industries.
Eliot Weinman, AI World Conference Founder and Chair “AI World, now in our third year, has been collaborating with deepsense.ai to help us develop world-class seminars on the state of AI, machine learning and deep learning in the enterprise. deepsense.ai works hands-on with Fortune 500 class enterprises to help them deliver AI solutions and assist organizations in unlocking their data science potential. Their programs cover not only the state of the practice in machine learning and deep learning but also critical areas like cybersecurity.”

All the knowledge in one day

The conference will kick off with a range of thematic seminars, including one hosted by deepsense.ai on enterprise machine learning and deep learning.
“Business leaders will hear about the latest developments in AI so they are up to date on what they can do and how they can assess their readiness to reach for AI innovation. However, this year we will focus more on the potential of machine learning for customer analytics, cybersecurity and healthcare”, shares Robert Bogucki, Chief Technology Officer in deepsense.ai and the host of one seminar.
During the seminar, deepsense.ai experts will explain how to reduce human workloads with AI-based process automation and how to optimize revenues by applying automated customer analytics. The latter will be illustrated with a debt risk modeling prediction case done for Home Credit.
Participants will also have the opportunity to join two discussions where panelists from leading organizations such as Microsoft, Uber, Nielsen and MIT will talk about AI applications in healthcare and cybersecurity. There are already numerous machine learning and deep learning solutions being used in healthcare, such as AI-powered diabetic retinopathy diagnostics, automated nurses, surgery assistance and personalized treatment plans.
Experts will also show how AI can be also used for spotting the signs of DDoS attack before it downs a service, automated traffic analytics, advanced spam filtering technologies and fraud detection techniques.
See here for a detailed agenda.

Opening the black box

“Artificial intelligence is a tool for building a new, brighter future. It offers easier access to healthcare services, it’s safer and just better for everyone. Still, there is a lot of uncertainty and doubt around artificial intelligence and how it is used. Companies don’t trust the solutions and ordinary people fear losing their jobs. That’s why we consider spreading knowledge about AI as important as building the solutions of the future,” says Bogucki.
To address this challenge, AI World and deepsense.ai joined forces not only to popularize AI applications in research and business, but also to explaining how they work and how to start using them. AI Trend readers can read about modern recommendation systems, reinforcement learning development and potential use for the business and the latest advancements and challenges in NLP. They can also listen to a webinar on various approaches to using AI in business. After the conference, AI World and deepsense.ai will publish an e-book covering these topics.
AI World Conference and Expo will be held in Boston, MA on December 3-5. The full agenda, speakers list and further details can be found on the official website. Use our special codes to get Free Expo Pass (code: 186800XDS) or save $200 on VIP Conference Pass (code: 186800DSAI).

Don’t miss the e-book release and sign up to our newsletter!
[contact-form-7 404 "Not Found"]

###

About deepsense.ai:

Media contact:

deepsense.ai trademarks at boilerplate

The game “Museum Treasures” created by the DS team, including several deepsense.ai employees, took first place in the National Museum in Warsaw’s HackArt hackathon. The game was on display at the Przemiany Festival, where this year’s leitmotif was artificial intelligence.

“The National Museum in Warsaw and hackathon is not a combination one sees every day, which makes it interesting”, muses Agata Chęcińska, a data scientist at deepsense.ai and member of the DS team.

In hackathons, teams of programming, analysis, creation and research specialists produce innovative solutions in a limited time. These competitions can be organized by companies, state administrations, NGOs and, increasingly, public institutions. In May, the National Museum in Warsaw got in on the act, organizing its programming marathon with a focus on artificial intelligence.

Hunting down the museum’s treasures

The game “Museum Treasure – find hidden details in the paintings!” is designed for parents and children, who can choose their favorite paths through the museum, with names like “animals” or “body parts”, to find their favorite ones. The application brings up fragments of well-known works pre-selected by AI from a digitalized collection housed in the museum. Working by category, the algorithm looks for hands, types of clothes, cuisine or any number of other objects in the paintings. Children then go and find them in the National Museum gallery that houses the items featured. In addition to being good fun, each path awards a badge and prize to those who finish, which will encourage people to visit the museum again and keep discovering new corners.
The game is cut from the same mold as PokemonGo. In this case, instead of creatures in the city, participants seek out specific fragments of images within the museum.

“Artificial intelligence helped in selecting the appropriate fragments from among the museum’s collections”, explains Agata Chęcińska, a data scientist at deepsense.ai. “Thanks to the challenge, children will look carefully at the pictures, and thanks to the good fun they’ll remember more”.

Future versions of the application will allow visitors to play individually, or in groups that can then compare the time it took members to find all of the treasures.

“HackArt MNW was the first hackathon organized by an art museum in Poland with the help of our partner, the Polish Development Fund Foundation. 40 participants and over 20 mentors took part. The projects submitted during the hackathon exceeded our expectations, with their creativity and advanced use of technology”, says Ewa Drygalska, of the museum’s Communications and Public Programs Department, and coordinator of HackArt MNW. “The ‘Museum Treasures’ project captivated us and the jurors because it helps our visitors explore the museum’s collections in an interesting and original way”.

Transforming art

A demo version of the game was available in the AI Showroom at the Przemiany Festival. The theme of this year’s festival was artificial intelligence and its impact on people’s everyday lives.

“There is a lot of emotion surrounding AI, particularly with questions about its safety and role in society”, explains Chęcińska. “Our application shows that in many cases such fears are unjustified, because AI is only playing the role of servant and the real value is rooted elsewhere. This is an excellent example of how AI supports and develops an institution that may not be associated with innovation. I hope that thanks to the use of our application, children will see the National Museum in Warsaw as a fun place, and that visiting it will be perceived as a great way to spend time on the weekend”, Chęcińska says.

###

About deepsense.ai:

Media contact:

deepsense.ai trademarks at boilerplate

deepsense.ai researcher Henryk Michalewski is set to join Michael Benedikt’s lab as a visiting professor at the Department of Computer Science at Oxford University. Michalewski’s post will begin September 1.

deepsense.ai is committed to contributing to international-level research in deep learning and reinforcement learning. Last year our researchers created a reinforcement learning agent that learned how to run like a human, using muscles and a skeleton. It also parallelized training in Atari games so that it scales almost linearly up to 64 nodes.
deepsense.ai has recently published a five-year strategy to become a European AI powerhouse, which will see it invest heavily in R&D, particularly reinforcement learning while continuing to focus on image recognition and natural language processing.

“We see reinforcement learning as a technology of the future and that’s why we focus on it in our research. It can reshape the landscape in various ways. For instance, we are currently applying it in our work on autonomous vehicles. There’s no challenge in following the trail blazers. Real technological leadership means being the trail-blazer”, comments Tomasz Kułakowski, CEO of deepsense.ai.

Data scientists from deepsense.ai have also successfully applied image recognition techniques in whale recognition software designed for the US National Oceanic and Atmospheric Administration and designed an award-winning crime forecasting model for Portland, Oregon. Michalewski’s activity at Oxford is another step in deepsense’s long-term research, which are intended to combine deep learning and reinforcement learning with practical applications.
Michalewski says of the opportunity, “The University of Oxford is known for excellent work on robotics and reinforcement learning with its Whiteson Research Lab and Autonomous Machines and Systems Lab. One of the first milestones in modern deep learning was achieved by the Visual Geometry Group, with its impressive work on image recognition and classification resulting in their taking 1st place in the 2014 ImageNet competition”.

Oxford practitioners

Both students and researchers at Oxford have successfully launched startups providing innovative and interesting AI-based services.
Among them is the well-known DiffBlue, which uses AI to support programmers in the most repetitive tasks, including code debugging and software testing.
Student-founded Onfido performs identity background checks by scrapping public databases and the data available online to make HR and hiring tasks easier. Other interesting startups showing out-of-the-box thinking include IntelligentX, the AI powered brewery and OpenMined, which builds open-source and transparent AI.
“With his inquiring nature and expertise in AI, Henryk will be able both to support Oxford University’s research on AI and share his knowledge with students”, said deepsense Chief Technology Officer Robert Bogucki. “Oxford is a great destination for a researcher and I am proud that deepsense can contribute to it”.

###

About deepsense.ai:

Media contact:

deepsense.ai trademarks at boilerplate

“Since the Vancouver police use predictive modelling to fight crime, it is a great city to gather crime prediction experts and confront theory with practice” says Patryk Miziuła, data scientist and lead instructor at deepsense.ai. Miziula is slated to be a panelist at the Joint Statistical Meetings 2018 conference held in Vancouver on July 28 – August 2.

Crime forecasting is usually associated with the film “Minority Report”, based on Philip K. Dick’s short story. It depicts a dystopian vision of people convicted of a crime before actually committing it.  In the story, information about potential killers or burglars comes from “precogs”, people with precognition talents who saw the future crimes.
“Fortunately, statistical data collected by the police, surveillance systems and municipal services is enough to build a system that can predict the probability of crimes in the area,” says Miziuła. “deepsense.ai prepared such a system for Portland, Oregon and we see a similar solution working efficiently every day in Vancouver”.

Reaching for PreCrime

The Vancouver crime prediction system was introduced in August 2017, following six months of testing in 2016. The pilot ran from April 1 to September 30, when people tend to go on vacation and leave their houses vulnerable to burglaries. During the testing time police recorded a 27% drop in break-ins due to the better distribution of police patrols.
According to Vancouver Police Department data, the total crime rate in the city decreased slightly (-1,5%) in 2016-2017, with the burglary rate falling the most  (18.4%).
The crimes are predicted by a machine learning model, which analyzes the data collected by the city police department and looks for patterns in crimes.
“Machine learning is currently the best tool to handle the massive amounts of data generated on crime. The experiment we did in Portland showed that bad neighborhoods tend to remain bad and crimes usually occur within close proximity. Given that, statistics and machine learning are great tools not only for predicting crime, but also for spotting the problem areas and inspire other public services to act. If vandalism is common in the district, maybe it is not only a matter of crime, but also suggests the need to implement social policy”, Miziuła explained. Further details of the experiment, the results and the technology used in the Portland predictive model are described in a blog post on deepsense.ai’s blog. The experiment will also be described during the upcoming Joint Statistical Meetings 2018 conference held in Vancouver on July 28 – August 2.

Datagazers

The model was built for a US National Institute of Justice competition. The crimes were split into a few categories – burglary, car theft and street crimes including assaults, robberies and shots fired. To predict the place and time a crime would probably occur, the Institute provided historical data with all the crimes registered in Portland between March 2012 and February 2017. deepsense.ai researchers fed the data to its machine learning model, which analyzed the dataset.
“During my speech I will discuss the details of our solution and show how statistics is the best fuel for machine learning techniques,” Miziuła says. “We are essentially datagazers, who can see the future by looking at data, rather than the stars”.
“Vancouver provides a great environment for a conference with crime forecasting on the agenda.  Not only is it a really safe city, but also gave us the opportunity to prove just how efficient these technologies are” concludes Miziuła.

###

About deepsense.ai:

Media contact:

deepsense.ai trademarks at boilerplate