Machine Learning Prague 2019

Most versions of IRL (inverse reinforcement learning) have assumed the human expert is rational, noisily rational, or, at best, fails to be rational is specific known ways. If the expert's rationality is completely unknown, however, then nothing can be said about their reward. More worryingly, simplicity priors don't help here - if we assume that the simplest interpretations are best, we get degenerate interpretations of human behaviour.
However, if we make a few a priori assumptions that cannot be derived from observations, we can do IRL successfully. One such assumption is that the mental models in human brains are informative about human preferences. This talk will illustrate how this can happen, and it's surprising connection with the "symbol grounding problem" - though applied to humans this time.

The taxonomy of the different business problems from a data scientific perspective is relatively finite and it all starts with an accurate system health detection. However, the traditional methods of system health detection have practical limitations when monitoring performances of applications, infrastructure or network.  These limitations arise from the imposition of frequentist statistics based frameworks on common problems such as anomaly and outlier detection. In this session, we talk about some novel patented frameworks derived from quantum mechanics along with reinforcement learning algorithms to transition from the widely used frequentists statistics approach for system health detection to bayesian statistical approach. We also elucidate how at Broadcom we capture such frameworks to address multiple problems across lifecycle of monitoring softwares like anomaly detection, root cause analysis, predictions and prescriptions. 

Many topics in unsupervised learning can be viewed as dealing with the relative geometry of data. In mathematics, topology and homotopy theory are the fields that deal with similar kinds of questions. Using ideas, techniques, and language from topology can prove fruitful for unsupervised learning. This talk will look at how topological approaches can be brought to bear upon unsupervised learning problems as diverse as dimension reduction, clustering, anomaly detection, word embedding, and metric learning. Through the lens and language of topology and category theory we can draw common threads through all these topics, pointing the way toward new approaches to these problems. I hope to convince you that topological approaches offer a rich and growing field of research for unsupervised learning.

Ruksi will discuss using our supreme overlord **data parallelism** to shard training data to achieve tenfold machine learning experimentation iteration speed. We'll go through the newest trends in distributed training, with some examples and highlighting common bottlenecks.

At Merlon Intelligence, one of the CEAi portfolio startups, we are helping banks to fight money laundering by providing a system for screening of their potential customers. Currently, we process more than 10 TB of news articles and our main goal is to index the data for efficient search and risk identification. Although the process seems to be straightforward, there are many challenging problems hidden inside application of machine learning methods, in data annotation strategies or in performance and explainability requirements. I am going to talk about several of these issues and share our solution(s) from the first prototype to the production-ready system used in a very rigid banking environment.

Assisted or Supervised models of machine learning are a great boon to automate many mundane tasks in Enterprise world. Supervised model combined with hypothesis models helps in creating “situations” and in predicting possible outcomes to a given situation, thereby making “situations” intelligent.
But the question in user’s mind is how can I believe/trust what’s proposed is correct?
It’s here explainability becomes critical in ML. Inherently Decision Trees, Random Forest and features like boosting does help in showcasing significant factors and does trace the path of decision. LIME (Local Interpretable Model-Agnostic Explanations) does play a good role here. But for a user that’s just 50%. The remaining 50% is why the other “possibly close match” was not selected. It’s here the concept of anti-model is used.
The anti-model addresses the question of why the other “possibly close match” was not selected say in a classification scenario by building anti-models to the original models and its inference models to “explain” the decision taken. Multiple anti-models are constructed and validated with original model.
We will look at this ML approach, its benefits, applicability and potential success SAP has obtained in this space.
 

Want to build a text classification pipeline and have text with high quality labels that business can act on?  Great, throw in a language model, some BiLSTMs and CNNs and viola, you have trained a high-quality classifier. 

Unfortunately, many text data available for industry projects are unlabeled and difficult to label because of their industry specific nature.  The challenge can be split into three parts:

1. Unsupervised Text Exploration – What types of texts are there?
2. Label Curation – Given the texts, which set of labels provides the most business value?
3. Active Labeling/Learning – Which texts should be labeled first/next when human labeling is expensive?

This talk shares a few technical stories for solving all three challenges.

Supervised Learning is all about the labels. From straightforward classification tasks to complex sequence to sequence models - we need labels in order to learn. Nowadays, some tasks have an abundance of labeled data available online, and for other tasks we usually just go for crowdsourcing. But what happens if both of these options are irrelevant? In this talk we will discuss exactly such cases, and present some creative ideas on how you can go along obtaining labeled data for your task.

JetBrains provides a comprehensive set of software development tools. Not so long ago, we wondered if we could develop a useful tool for our design team, too. Our designers’ technology stack is based mostly on MacOS, Adobe products, and Dropbox as the cloud storage. Designers are very talented and creative people; sometimes they might store their files with an inconsistent path or under an inconsistent name. After a while, finding some old file (among tens of thousands of others) when needed started to turn into a puzzle.

We came up with the idea to develop some deep-learning-based tool that could sort out all these files (raster and vector images along with PDFs, PSDs, and AIs!) and get some interesting stats about them.

In this talk we’ll demonstrate how SpaceKnow exploits the state-of-the-art deep learning techniques and the latest cloud technologies to generate insights and provide a global geospatial intelligence. We will focus mainly on applications related to the combination of optical satellite imagery, multispectral imagery and SAR. We will elaborate on how can recent models like GANs possibly help us.

Computer vision in aerial imagery domain can be quite challenging for some specific use cases, specifically in public safety. Based on one of the client’s needs we needed to focus not only on the neural network architecture, but also image preprocessing which is sometimes omitted. We leveraged particular series of convolutional filtering and wavelets transformation which is space depending. 

In the presentation, we will go through the process of data collection. Show benefits of preprocessing techniques to increase visual information. Discuss wavelet transformation and why wavelet transformation was used. Demonstrate the app which was used to display results. 

In recent years, artificial intelligence researchers have built impressive systems. Two of my former postdocs — Demis Hassabis and Amnon Shashua — are behind two main recent success stories of AI: AlphaGo and Mobileye, based on two key algorithms, both originally suggested by discoveries in neuroscience: deep learning and reinforcement learning. There is, however, little in terms of a theory explaining why deep networks work so well. In this talk, I will review an emerging body of theoretical results on deep learning including the conditions under which it can be exponentially better than shallow learning. The class of deep convolutional networks represent an important special case that avoids the curse of dimensionality for the class of hierarchical locally compositional functions.

Have you heard that knowing machine learning is the easiest way to get rich quickly? Let’s test this statement. Kaggle is the place to do data science projects, why not to start there? During this session we will solve simple Kaggle competition. Actually, we will submit two solutions. The first made with super-duper deep neural network (black-box approach). Then we will follow proven ML methodologies and solve the problem methodically. All that using SQL Server Machine Learning Services. Minimum slides and maximum fun guaranteed.

Automated Machine Learning was gaining momentum in 2018 with several open source projects released from top AI labs. You will learn internals of AutoML algorithms, their advantages and limitations. We will focus mainly on predictive modeling, but also demonstrate our AutoML research in the field of data clustering and recommender systems.

This talk will focus on all the engineering aspects involved in Machine Learning at scale. A common warning shared with aspiring Data Scientists & ML engineers is that 90% of the work is about gathering, cleaning and validating data plus deploying and monitoring models. Yet for a long time most of the open source ML tooling focused on the modelling part.

We will first give an overview of the different ML Engineering frameworks out there, both open and closed source. We will then focus in on Kubeflow Pipelines and TFX (Tensorflow Extended), both of which are open source and model agnostic, by giving an end-to-end example highlighting why these frameworks are incredibly powerful.

Throughout the talk we will work to implement an end-to-end example, a deep neural network for predicting customer satisfaction in taxi trips using a public dataset with over 100M rows from the city ofChigago. This example includes transforming and validating the data, training a model in a distributed way, validating and monitoring model performance and last but not least deploying the model."

Web search engine almost never searches (only) for the words that users type in. Instead, during query processing phase, the engine tries to predict additional terms to search for.  This is especially important for misspelled queries.  The talk will provide details about how machine learning can be used to "read" user's mind for a correct version of the query. It is not a neuroimaging presentation but rather a brief history and current research of the spelling error correction system in the Seznam.cz's search engine".

Robust early case assessment and search term identification has been a critical competency within forensic investigations for a long time. However, the volume of analyzed text data grows rapidly and a multi-language litigations are getting more and more common. Traditional approaches are getting costly and less practical and it is getting difficult to find a “smoking gun”. To discover patterns and trends in the data, it is no longer possible to rely on manual analysis only. This session introduces the state-of-the-art ML methods for automated unstructured content analysis within electronic discovery and crime detection as such.

The world is becoming more dynamic and real-time. We see more and more services becoming on-demand including food, transportation, delivery etc. This trend will continue to grow and become dominant. 
For example, there are more than 5 million drivers working for ride sharing companies spending 20 millions hours on the road every day. In order to allocate them more efficiently, ride sharing companies need vast amounts of accurate predictive models that can adopt to a changing and dynamic world for each one of the cities they operate in.
We will explain how AUTONOMOUS predictive technology, utilizing Automated Machine Learning for Dynamic Systems, can solve the problem at a more granular level than ever before. The way to solve it is by automating the whole pipeline of the construction of predictive models, from data preparation to feature extraction, feature set selection, hyperparameter tuning and performance evaluation of the models in Real-Time.

Over the past couple of years, there has been increasing interest in applying the latest advances in machine learning to creative projects in art, music, theatre and beyond. From Google's DeepDream and style transfer to a GAN-generated painting selling for 400,000 USD at auction, more and more creative AI projects are moving beyond the world of research and academia into the public eye. This talk will give an overview of how artists and creative technologists are using machine learning and investigating its creative potential.