Popular posts Archives - deepsense.ai

Why do we need more data scientists and why should you become one?

If you have come across this article, you already know what data science is and how it can be used. That’s great! Now, you are probably wondering why there is so much fuss about data science. If you want to know why you should become a data scientist, the facts speak for themselves!

Demand

According to LinkedIn’s 2017 U.S. Emerging Jobs Report, the number of data scientists has grown over 650% since 2012. Yet there are still too few people exploiting the opportunities in this field. Why has it grown so fast?
Companies need to use data to run and grow their everyday business. The fundamental goal of data science is to help companies make quicker and better decisions, which can take them to the top of their market, or at least – especially in the toughest red oceans – be a matter of long-term survival. The number of companies prepared to use big data is increasing. As Dresner Advisory Services laid out in their Big Data Analytics Market Study, forty percent of non-users expect to adopt big data in the next two years.
What is more, you can apply machine learning on smaller data sets, such as ones from a local company’s social media or shopping gift card history. This provides even more opportunities and increases the demand for data scientists. Job growth in the next decade is expected to exceed growth from the past ten years, creating 11.5M jobs by 2026, according to the U.S. Bureau of Labor Statistics. Companies are building up their data science teams to embrace data analytics and will make it integral to their success. Why are these analytics so important? Is it worth working for one of these companies? You will find the answer in the next two chapters.

Influence

Data science changes how decisions are made and companies are adapting a data-driven approach on a huge scale. Data-driven decisions made with advanced data analytics benefit all manner of company, from global behemoths to medium-sized companies down to local businesses looking to get ahead. Lack of data is rarely an issue – mountains of it are collected every single second, and we are beginning to understand the potential and influence it can have. Data sets in the right hands can help predict and shape the future.
The problem is getting data sets to mingle. It is the data scientist’s role to transform organisations from reactive environments with static and aged data, to automated ones that continuously learn in real time. Forecasts are simple – data is a valuable resource and investing in it will definitely pay off.

Tractica forecasts that worldwide revenue from deployments of AI software, hardware, and services will increase from $14.9 billion in 2017 to $23.6 billion in 2018, a year-over-year increase of 58%.

Do we need more data scientists?

Now, knowing that data science is in huge demand, you are probably wondering who is going to do all the work. Do we have enough data scientists? Maybe the market is already flush with experts. Nothing could be further from the truth – data scientists are few and far between, and highly sought after. IBM predicts demand for data scientists will soar 28% by 2020. Machine learning and data science are generating more jobs than there are experts to fill them, which is why these two fields are the fastest growing tech employment areas today.

Why should you become a data scientist?

Let’s start from the bottom of Maslow’s pyramid of human needs, which you secure with money. According to Glassdoor, in 2016 data science was the single highest paid profession. If data is money, as they say, then this should come as no surprise. The combination of skills necessary to do data science the right way is not common. The good news, however, is that if you want to become a data scientist and are willing to develop yourself, you are very likely to succeed. A background in mathematics, statistics or physics is a good foundation to build upon. You don’t necessarily need to have finished a data science program. We write a lot about learning methods on our blog, which you’ll see if you read our next post. Sign up for our newsletter if you would like to be updated.

Make the world easier

Besides its financial and economic aspects, data science is simply a fascinating discipline, one which affects many areas of our everyday lives and makes the world a better place. We already use it in many fields, such as quick and easy customer service, intelligent navigation, recommendations and voice-to-text. You can even improve the resolution of an image with deep learning.
We don’t have space enough to chronicle as of the ways that data science is improving people’s lives. It is indispensable to the banking sector as it is used to detect fraud by analyzing the behavior of financial institutions in real time. Elsewhere, robots will be used to help the elderly and the disabled gain mobility and independence. Data science makes these breakthroughs accessible to individuals, solves social problems and modernizes business. Most importantly, you can take part in the revolution data science is bringing about.

Things that matter

Among the many reasons you would want to become a data scientist is that you can make a positive contribution to society. Data science can give you some pretty super superpowers. One of them is reshaping industries like healthcare. The amount of data produced about patients and illnesses rises by the second, opening new opportunities for better structured and more informed healthcare. The challenge is to carefully analyze the data in order to be able to recognize problems quickly and accurately – like deepsense.ai did in diagnosing diabetic retinopathy with deep learning.
Did you know that deep learning can help predict dangerous seismic events and keep miners safe? Underground mining is fraught with threats including fires, methane outbreaks or seismic tremors and bumps. An automatic system for predicting and alerting against such dangers is of utmost importance – and also a great challenge for data scientists. Our deepsense.ai team created a machine learning model for the Data Mining Challenge: Predicting Dangerous Seismic Events in Active Coal Mines, which was the winning solution, and one we take great pride in.
Another superpower is saving rare species. When you think of rescuing endangered animals, you see remote jungles and scientists chasing them. This is a stereotype that has changed a lot in recent years. Complex predictive models and algorithms can create insights that help scientists analyze threats to wildlife and create a solution that can save animals – all from the relative comfort of a desk. In fact, it was at our very desktops that we created the Facebook for whales, and It works with 87% accuracy!

Fun

Psst… There’s just one more thing. Data science can simply be fun. Can deep learning play Atari games? Yes! Or perhaps you want to make art even if you aren’t an artist. Data scientists can do it. The only limitation is your imagination!
In the next post you will find inspiration on how to become a data scientist and what kinds of data science training are available on the market, along with their advantages and drawbacks. Stay tuned!

Artificial intelligence imagining and reasoning about the future

March 9, 2018/in Blog posts, Data science, Deep learning, Machine learning, Popular posts /by Anna Kowalczyk

Researchers from the deepsense.ai machine learning team, Piotr Miłoś, Błażej Osiński and Henryk Michalewski, together with Łukasz Kaiser from Google Brain’s TensorFlow team optimized infrastructure for reinforcement learning in the Tensor2Tensor project.
The team enhanced an advanced reinforcement learning package with improvements related to the state-of-the-art algorithm called Proximal Policy Optimization, which was originally developed by OpenAI. The algorithm proved to be very versatile and was used to solve games such as Dota 2, robotic tasks like Learning to Run (with our model in sixth place) and Atari games.

AI imagination and reasoning

The idea behind the improvements was to develop an artificial intelligence capable of imagining and reasoning about the future. Instead of using precise and costly simulators or even more costly real-world data, the new AI spends most of its energy on imagining possible future events. The process of imagining is much less costly than gathering real data. At the same time, a properly trained imagination is a far cry from daydreaming. In fact, it makes it possible to precisely model reality and reason about it hundreds of times faster than would be possible using simulators.
The novelty of Tensor2Tensor consists in implementation of the Proximal Policy Optimization, which is completely contained in the computation graph. This is the main technical factor behind the lightning fast imagination.

End-to-end training inside a computation graph

In the second stage of the project the researchers from deepsense.ai, the University of Warsaw and Google Brain are focusing on the end-to-end training of an reinforcement learning agent fully inside a computation graph.

One of the steps in the experiment is the implementation of the Proximal Policy Optimization algorithm entirely using TensorFlow atoms. The training will be run on Cloud Tensor Processing Units (TPUs), which are custom Google-designed chips for machine learning. Assuming that a game simulator can be represented as a neural network, we expect that the whole training process can then be kept in the memory of the Cloud TPU.
Stay tuned for the results of our project!

Logo detection and brand visibility analytics – example

October 24, 2017/in Blog posts, Data science, Deep learning, Machine learning, Neptune, Popular posts /by Michal Romaniuk

Monitoring brand visibility and measuring ROI on marketing campaigns are important business challenges, especially in ad-driven industries. Brands frequently have to function in cluttered advertising spaces, including outdoors or on websites, with limited exposition times. In this post we describe our deep learning solution for automated logo detection and visibility analytics.

How many people does your brand reach?

Brands are often promoted through sponsorship campaigns at sports and cultural events. These events attract large numbers of viewers both directly and via media reports, allowing brands to get favorable positioning. However, sponsorship contracts often come at a steep price, so brand owners are naturally more than a little interested in finding out how effectively their outlays are working for them. The problem is, it’s difficult to assess quantitatively just how great the brand exposure is.
The current approach to computing such statistics is based on manually annotating broadcast material, which is tedious and expensive. In order to address this problem, we have developed an automated tool for logo detection and visibility analysis, providing both raw detections and a rich set of statistics.

Solution overview

We decided to break down the problem into two steps: logo detection with convolutional neural nets and an analytics step where summary statistics are computed.

The main advantage of this approach is that it is straightforward to swap the analytics module for a different one if different types of statistics are called for, or even if the neural net is to be trained for a completely different task (we had plenty of fun modifying this system to spot and count coins – stay tuned for a future blog post on that).

Logo detection with deep learning

There are two principal approaches to object detection with convolutional neural networks: region-based methods and fully convolutional methods.
Region-based methods, such as R-CNN and its descendants, first identify image regions which are likely to contain objects (region proposals), then extract these regions and process them individually with an image classifier. This can be quite slow, a problem that can be remedied to some extent with Fast R-CNN, where the image is processed by the convolutional network as a whole and then region representations are extracted from high-level feature maps. Faster R-CNN is a further improvement where region proposals are also computed from high-level CNN features, which accelerates the region proposal step.
Fully convolutional methods, such as SSD, do away with processing individual region proposals and instead aim to output class labels at the region proposal step. This approach can be much faster, since there is no need to extract and process region proposals individually. In order to make this work for objects with very different sizes, the SSD network has several detection layers attached to feature maps of different resolutions.

Since real-time video processing is one of the requirements of our system, we decided to go with the SSD method rather than Faster R-CNN. Our network also uses ResNet-50 as its convnet backbone, rather than the default VGG-16. This made it much less memory-hungry, while also helping to stabilize the training process.

Model training

In the process of refining the SSD architecture for our requirements, we ran dozens of experiments. This was an iterative process with a large delay between start and completion of an experiment (typically 1-2 days). In order to run numerous experiments in parallel, we used Neptune, our machine learning experiment manager. While the experiment is running, Neptune captures the values of the loss function and other statistics, displaying them in a friendly web UI. Additionally, it can capture images via image channels and display them, which helped greatly with troubleshooting the different variations of data augmentation that we tested.

Logo detection analytics

The model we produced generates detections just fine. However, when even a short video is analyzed, the raw description can span thousands of lines. To help humans analyze the results, we created software that translates these descriptions into a series of statistics, charts, rankings and visualizations that can be assembled into a concise report.
The statistics are calculated globally and per brand. Some of them, like brand display time, are meant to be displayed, but many are there to fuel the visual representation. Speaking of which, the charts are really expressive in this task. Some features include brand exposure size in time, a heatmap of a logo’s position on the screen and bar charts to allow you to easily compare various statistics across the brands. Last but not least, we have a module for creating highlights – visualizations of the bounding boxes detected by the model. This module serves a double purpose, because such visualizations are also a source of valuable information for data scientists tweaking the model.

Results

We processed a short video featuring a rivalry between Coca-Cola and Pepsi to see which brand received more exposure in quantitative terms. You can watch it on YouTube by following this link. Which logo has better visibility?
Below, you can compare your guesses with what our model reported:

Possible extensions

We found that our system can easily be adapted for other object detection tasks. This is a major advantage because object detection often appears in the context of computer vision projects, either as a goal in itself or as part of a longer chain of processing steps.
There are many business problems where object detection can be helpful and at deepsense.ai we have worked on a number of them. Here’s a partial list of work we’ve done:

For one of our clients we have developed a solution for extracting information about ingredients from photographs of FMCG products, using object detection networks to locate the list of ingredients on the product photograph. This helps our client make data collection more efficient and automatic.
Recently, we took part in Kaggle’s iMaterialist challenge, and rose to the task of building a machine learning model that recognizes product attributes based on images. This system can be used to automatically assign tags to products in online retail, helping customers to find more products they may be interested in (a cross-selling use case). You can read more about this solution in our blog post.
Our team also placed 4th in Kaggle’s Satellite Imagery Feature Detection challenge. The goal was to segment satellite images into 10 different classes of objects. Possible applications include agriculture, disaster relief and environmental studies. See our blog post for more details.

Solutions

Knowledge base

deepsense.ai

Support

Join our community