Machine Learning Prague 2016

Welcome to ML Prague 2016

Keynote speech: Deep Learning and Intelligent Applications

With recent advances in deep learning, big data, and computing infrastructure, we are able to engage and converse with intelligent machines more effectively. For example, computers will realistically reach human parity in recognizing speech in the next a few years. The computing industry’s progress in speech and vision is apparent in Microsoft products and services, including Cortana, Skype Translator, and Project Oxford cloud services. Dr. Xuedong Huang will use these examples to illustrate not only our society’s historical collective efforts, but also enumerate our remaining challenges to reaching human parity, an accomplishment that will have a profound impact on society.

Inventing General Artificial Intelligence: A Vision and Methodology

Marek Rosa, CEO/CTO of a Prague-based general AI development company, will introduce his methodological approach to general purpose artificial intelligence (AGI) research. He will highlight a key concept he is working to develop with his team: School for AI, a set of training environments where AI agents will learn in a gradual and guided way. To perform successfully in these environments, an agent needs to demonstrate a set of abilities. The analysis of these abilities leads to a specific list of functional requirements, which will then be implemented as a unified AGI architecture.

School for AI will be released as an open source tool where AI researchers and machine learning practitioners can benchmark their AI agents.

Intelligent Personal Assistants

Apple Siri, Google Now, and Microsoft Cortana are emerging new web-applications known as Intelligent Personal Assistants. They combine many latest and developing AI technologies such as speech recognition, question answering, information retrieval, knowledge management, IoT etc. The presentation will discuss the basic architecture, challenges and use cases of the future intelligent assistants.

Artificial Intelligence in Creative Writing

Have you ever thought about artificial intelligence smart enough to write novels or poetry? Meet the future. I will describe to you a machine learning model which is able to write an entire book or assist you in writing your own fictional literature.

Learning Representations for Drug Discovery

Utilizing gene expression measurements, it is possible to characterize drugs and drug candidates for their on- and off-target activities, predict treatments for new indications, and highlight potential safety concerns. Improvements in technology now permit this approach to be applied on a qualitatively larger scale, which in turn requires new analytics. We find that applying methods inspired by the deep learning community can improve the interpretability, fidelity, and predictivity of pharmaceutical genomics.

Online Hyperparameter Tuning in Non-Stationary Environments

Fitting hyperparameters online usually requires an extensive parallel search in the parameter space, which is often very demanding in time and computational resources. A comparison of parameters sequentially is often not valid, due to non-stationarity. Here, we present a method using global Bayesian optimization with Gaussian Processes to model certain kind of non-stationarities.

ML Approaches to Sentiment Analysis

Current computational linguistics witnesses a massive increase of interest in sentiment analysis, as it is a powerful means of public opinion mining. However, is it even possible to distinguish between good and evil using statistical methods? How would you treat irony, idioms, innovative vulgarisms and other inherent elements of natural language? Are emotions language-independent? I will share my thoughts and opinions towards these issues, describing the state-of-the art approaches employed in sentiment analysis.

Distributed Representations for NLP

Distributed representations are currently a hot topic both in academic research and in industrial applied machine learning. I will talk about some breakthroughs from recent years, and will try to explain what the distributed representations are about and why is there so much excitement around techniques such as word2vec.

TR Discover: A Natural Language Interface for Exploring Linked Datasets

Keywords are the dominant technology for providing non-technical users with access to linked data. This is problematic, because keywords cannot express all the necessary details of user intent. Non-technical users cannot be expected to use an expressive query language such as SQL, but still need to access the data.

They can, however, use English, Our system, called TR Discover, maps from a fragment of English into an intermediate first-order logic representation, which is in turn mapped into SPARQL or SQL. It has been tested both on a dataset relevant to drug research and on the publicly available QALD-4 dataset. The system incorporates a tailored autosuggest mechanism and a back end that delivers task appropriate analytics.

There will be a demo of a publicly visible version of the system.

Finding the Oddballs

Anomaly detection is a handy tool for a machine learning practitioner's toolbox. Not only is it useful for particular domains like fraud detection or intrusion detection. It's also nice for more general tasks, such as data cleaning or early warning of concept drift. We'll review a few approaches to anomaly detection, and then dig a little deeper into an ensemble technique called isolation forests.

Recognizing Malware

Today’s malware ecosystem produces hundreds of thousands distinct samples per day. To leverage similarities between the samples for automated classification we built a distributed database engine relying on GPUs. We use instance based classification which gives us online learning and accountability for every decision. With query times of fraction of a second even using a compound distance function, this system is able to classify the incoming samples in the real time. Samples classified as malware are directly used to generate rules to identify similar samples in machines of our customers.

MatrixNet Applications at Yandex

MatrixNet is a proprietary machine learning tool developed by Yandex and used widely throughout the company. It has different learning modes, such as ranking, regression and classification. The algorithm is based on gradient boosting over decision trees, but its implementation has a lot of heuristics that give strong defense against overfitting. We start with some details of the algorithm itself, then discuss its applications in web search, ad click prediction, churn prediction and other areas.

Closing