Machine Learning Prague 2018

Google’s ongoing research in AI is what powers many of the products being used by hundreds of millions of people a day - from Translate to Photo Search to SmartReply for Inbox. One of the things that enables these advances is the extensive collaboration between the Google researchers in our offices across the world, all contributing their unique knowledge and disseminating ideas in state-of-the-art Machine Learning technologies and techniques in order to develop useful tools and products.

While we aim to provide everyday users with useful and innovative products, it is also important that we expand access to AI for other companies to use in their products as well.

In this talk Anna Ukhanova will speak about Google's approach to AI, collaborations with academia, our open-sourced Machine Learning platform TensorFlow and Cloud Machine Learning.

As part of the Wolfram Language, we recently developed an original machine learning framework that can be used by both beginners and experts in the field. The framework includes a symbolic and functional neural network framework, a repository of pre-trained networks, and fully automated machine learning functions. I will give an overview of this framework, demonstrate its capabilities, and talk about the scientific and technical solutions used under the hood.

Machine learning practitioners would all love a method of providing the best classifier given some data, with no algorithm selection or parameter tuning required. New techniques in the research literature, especially Bayesian parameter optimization, hold the promise of being able to do exactly this, but having a method for doing a thing is very different to actually doing it. In the first half of this talk, I'll discuss the nuts and bolts of machine learning automation, including how to know when we've actually done it right. In the second half, I'll talk about why “automation” might only be a poor substitute for our actual goal.

Shoppers are increasingly engaging in new modes of interaction such as messaging and voice. To provide our customers with a personalized experience across multiple engagement endpoints, eBay has embarked on a journey to develop AI-powered personal shopping assistants. In this talk, we will highlight the many ways we are leveraging big data and machine learning to enable conversational commerce at eBay.

Recommender systems interact repeatedly with their users. For instance, Google’s search engine processes billions of search queries daily. This gives rise to the exploration-exploitation dilemma. The recommender system may decide to exploit, recommended the best choice under its potentially-suboptimal current model of the user; or to explore, recommend suboptimal choices under the current model that allow the algorithm to improve its model, for instance from the clicks of the user. Multi-armed bandits are a principled way of dealing with the exploration-exploitation dilemma. In this talk, we introduce stochastic multi-armed bandits and show how to use them to design real-world recommender systems. These problems have multiple challenges. The space of recommended content is typically exponentially large, there are L^K lists of K web pages out of L. In addition, the responses of users are subject to strong biases. In particular, the user often does not click on the web page because other recommended web pages are better, not because this particular web page is bad.

Guaranteed display advertising has been an area of vivid interest over the recent years. Advertising agencies want to be able to make a contract with advertisers for a certain number of impressions and then to honor it. Guessing the number of impressions with a given target audience then becomes a necessity.

Deep Learning has emerged as the most successful field of artificial intelligence with overwhelming successes in industrial speech, language and vision benchmarks. I invented LSTM recurrent neural networks, which evolved into a key technology in different AI fields like speech, language, and text analysis.

We use LSTM for natural language processing in collaboration with companies like Zalando and Bayer, e.g. to analyze fashion blogs or twitter news related to health. In the AUDI Deep Learning Center and in collaboration with NVIDIA we apply Deep Learning to advance autonomous driving. With Deep Learning we won the NIH Tox21 challenge, predict biological effects of drug candidates from their chemical structure and from high content imaging.

In current research we analyze convergence properties of generative adversarial networks (GANs) using stochastic approximation (cf. TTUR). Further we investigate self-normalizing networks, which automatically converge to their optimal learning conditions (cf. SELUs). Most recently, we developed a new reinforcement learning method, which outperforms Monte Carlo Tree Search and other RL methods on delayed reward problems. This new method has a potential to initiate a paradigm shift in reinforcement learning.

Alquist is one of the three chatbots which were selected as the Alexa Prize finalist. Its task is to maintain a coherent and engaging conversation with a user about popular topics (movies, sports, news, music, etc.), and it uses a combination of generative and retrieval approaches to balance the advantages of both of the methods. It has been developed for more than a year by a group of 5 students. We will describe the Alquist architecture with all the NLP techniques we used as well as all the obstacles we encountered during the Alexa Prize competition.

Quantum mechanics offers tantalizing prospects to accelerate and improve certain machine learning tasks. Can these prospects be born out in practice? With rapid advancements in quantum processor technology, a vast international community has quickly formed to investigate the uses of current (medium sized and noisy) quantum processors. Some of the results look promising but challenges still remain. So exactly when will we see a quantum machine learning revolution?

Lessons from a decade of building advanced ML solutions for companies like Amazon, Hearst or Autodesk, maintaining open source software, running a Student Incubator and launching innovative AI products. This talk focuses on the journey of standing on your own two feet when it comes to applied R&D. How did we combine academia, open source and industry while following our dreams?

Ethical and moral aspects of research were always on the table in democratic countries, but when governments fail to find the line between meaningful regulations based on threat analysis and replace it with ideology, science is on the retreat. GDPR and ePrivacy regulation might be such direct threats for the substantial part of the AI research and industries in EU.

Konica Minolta launched Care Support Solution, a monitoring system for nursing care workflow innovation, in 2016. The Sensor Box, which is the core of the system, performs 24-hour monitoring using near-infrared and microwave sensors. If a patient sits up on the bed, leaves the bed, falls down, or falls out of the bed, nursing staff are notified via a smartphone with video display. The staff can also use their smartphones to input care records or vital-sign data on the spot and share them instantly. By providing one-stop total innovation that covers both the nursing care frontlines and administrative work such as record keeping, the Care Support Solution offers greater efficiency for the entire nursing care operation. In this talk, the concept of the Care Support Solution and the computer vision technologies used in the system are introduced.

Being open and transparent about scientific developments and discoveries is important for any discipline. In healthcare it is crucial: given their huge societal impact, openness can make or break the success of new therapies and drugs. Applying machine learning methods in healthcare poses new challenges to openness, for example due to the black-box nature of algorithms or due to privacy concerns. In this talk, I will give a brief overview over how machine learning methods are being applied in clinically relevant settings, what the open science challenges are and how possible solutions may look like.

One part of the anti money laundering process is a background check of the client to determine whether the client has been engaged in any risky or illegal activity. This was done mostly by searching manually news databases and reading returned articles. Vlado Boza will present how Merlon Intelligence helps automate this process by implementing a solution using recurrent neural networks and also how it has coped with limited availability of training data.

Jumpshot has a large datastream of web browser activities of an international panel. To produce accurate estimates of aggregate metrics, we need to calibrate from our data, a fraction of the actual web traffic, to its true value for any particular country, platform, domain and day. For precise calibration, correction of bias of our panelists is inevitable task that traditionally requires knowledge about both panelists and the whole population, which is not available in our case. Popular approach is to utilize sophisticated learning methods that produce phenomenological models with high accuracy but low interpretability. Contrary to that, our story shows that simpler but mathematically sound microscopical model can outperform such models and provide more clarity at the same time.

Deep Learning has the chance to be a massive force for good in the world, and I’ll discuss some of our projects in that area: Finding disaster victims using facial recognition, helping quantify the impact of illegal mining in Ghana, and tracking endangered species. We’ll take a look at the problems, their solutions, how we scale out model training on the cloud, and how we deploy solutions in the cloud for operation at scale – and I’ll point you to the open-source repositories so you can do it yourself.

DP DHL has 500K employees, making it one of the largest companies worldwide. Their business is related to massive data processing, supported by DHL ITS with its 4000+ employees. How can machine learning make their daily jobs easier? What information helps them understand ongoing issues better? What can they do to anticipate difficult periods of operation? Let us share some insights from our own experience in this domain.

In this talk we’ll explore how we combine machine learning algorithms and computer vision techniques with the latest cloud technologies for semantic segmentation, object detection and classification on multi-spectral multi-resolution satellite images. We’re developing a proprietary, scalable platform and frameworks enabling reproducible multi-GPU training and massively parallelized prediction, aiming to deliver global geospatial intelligence. Our production system is a single-point access to the world’s largest satellite providers, including DigitalGlobe, Airbus, Planet, and Urthecast, offering real-time, on-demand analytics through APIs and an web application.

Machine learning is daily bread for the Data Science team at O2 and it has been applied in many projects and data products, both for internal and external business clients. However, some of these ML efforts stand out by being extremely rewarding for us not necessarily in terms of business success, but in terms of pure joy of trying something that not many in the telco domain have tried before, and perhaps seeing some fascinating or surprising results. This talk will thus be exactly about such cases.

At CEAi, we don’t like repetitive or tedious work. On the other hand, we like to build cool stuff that simplifies our lives. That's why we decided to get rid of some pain points that are pervasive in machine learning. “I trained a model 6 months ago … which data did I use again? What parameters did I try?” Don’t tell my boss but let’s be honest: usually, I can’t remember. Hence my motivation to talk about our effort to build a system that facilitates reproducibility, automates validation, testing, auditing and continuously searches for the next best model. In addition to this, the system will use active learning strategies to shorten the total time to model deployment and reduce load on annotator pools.