Machine Learning Prague

Self-driving is a safety-critical application. In this talk, I first present the machine learning framework used for autonomous driving, gathering contributions from computer vision, deep learning, and autonomous robotics research fields. I then discuss some of the main challenges we face at valeo.ai to improve advanced driver-assistance systems. I will give some examples such as unsupervised domain adaptation for visual segmentation, or driving behavior explanation system using natural language processing.

Deploying your ML model to production may bring you headaches for many reasons, e.g. out-of-distribution input data or low-quality user input. Recognizing those cases is a crucial step for providing actionable feedback and handling those cases properly. This talk uncovers our simple yet effective recipe for integrated confidence estimation learning alongside a practical example used in a production environment.

CFD, Computational Fluid Dynamics are numerical methods or algorithms to solve fluid flows problems. They help model fluids density, velocity, pressure, temperature, and chemical concentrations in relation to time and space. Many industries such as automotive, chemical, aerospace, biomedical, power and energy, and construction rely on fast CFD analysis turnaround time. Typical applications include weather simulations, aerodynamic characteristics modeling and optimization, and petroleum mass flow rate assessment.

ByteLAKE has been working on leveraging Artificial Intelligence (AI) and Deep Learning to significantly accelerate CFD simulations. These typically take anything between hours, days or weeks. byteLAKE's CFD Suite, a collection of AI models helps predict accurate results within minutes. During his presentation, Krzysztof Rojek will take share more details about the solution, its scalability, compatibility with CAE tools and OpenFOAM solvers and present benchmarks for commercial simulations. Also, Krzysztof will present how to get started with CFD Suite and accelerate your simulations.

Regardless of your job, you need your heart working. Early diagnostics, available through many wearable devices on the market, can capture diseases before they can prograde to the severe form. But we do not want to scare you in this talk; we will iterate from simple to more complicated ML/DL methods and their application in early diagnostics using ECG signals from telemedicine data.

Every company that takes machine learning seriously needs to ‘productionalize’ their ML pipelines. Efficiently, robustly and at scale. The emerging methodology called Machine Learning Operations (MLOps) comes to rescue.

However, there are already hundreds of convenient, stand-alone and overlapping ML tools, workflow managers, automation and orchestration frameworks, developed by both vendors and the open-source community striving to put this methodology into practice. New ones keep on appearing (and disappearing). As a result, many companies are contemplating whether to buy an MLOps platform or to build one internally. And if the latter then they hope to postpone the architectural decisions until the sheer amount of available options reduces to a widely accepted set of tooling. But will it ever?

In this talk, I will share some of Credo’s experience on how to design a modular and future-proof MLOps platform, based on open-source tooling, that hits the ground running today and survives still tomorrow in the everchanging zoo of ML tooling.

In the software engineering world, CI/CD practices have proven to be a reliable and effective approach to automating recurring tasks, like running tests, code analysis checks and even delivering final products to production. In this talk, we will present how to automate ML processes using GitHub Actions or GitLab CI/CD and Continuous Machine Learning (CML) library that will take care of:
• transferring large datasets to CI runners
• managing GPU/CPU resources for computations
and
• generating ML model report with metrics and plots right in GitHub Pull Request
so that ML specialists can focus on research.

This panel discussion will feature three panelists from world's renowned research groups where the issues of AI value alignment are taken very seriously. It's a topic inherently related to AI ethics, safety, risks, benefits and future potential. But the goal is to show in a very open discussion with the audience that it really should concern every ML practitioner.

Deep learning has achieved unprecedented performance in a wide range of domains ranging from computer vision, speech recognition, and natural language processing to game playing. However, many industrial systems still rely on human-written and maintain rule-based systems to perform classification. The reasons include better explainability of the rule-based systems and their modularity, which is crucial in dealing with non-stationary problems. We will discuss each approach's advantages and disadvantages and the possibilities of getting the best of both worlds.

AI in Pharma and Life Sciences is connected predominantly with early drug discovery in popular news. There are many more opportunities to apply both classical and modern AI in Pharma. In this session, we will present an overview of drug discovery and development phases together with examples of how AI applies to them. We will mention also a couple of selected examples from our own work: Reaction PathFinder - computation of optimal synthesis routes based on a graph of chemical reactions, Mutation Maker - a tool for designing optimized proteins, CAKE - an evaluation and recommendation engine for streamlining pharma manufacturing change requests (also presented at Amazon Re:Invent), and several Natural Language Processing use cases. The panelists are hands-on experts looking forward to an engaging discussion with you.

Most successful industrial ML systems of today require large amounts of labeled data for training to perform well. Yet high quality labeled data are a scarce resource. This is especially true in cybersecurity and other domains which deal with difficult to label, severely class-imbalanced and highly confidential data. We will discuss approaches to learning practical classification systems with limited amount of ground truth. We would like to focus on concrete learning approaches and also on the broader challenges related to obtaining and working with labeled data in practical classification systems.

Computational Fluid Dynamics (CFD) are numerical methods used across many industries (chemical, pharma, automotive, construction, oil&gas just to name a few) to model fluids pressure, velocity, temperature etc. Typical applications include modelling aerodynamics, chemical mixing, air flows around buildings etc. CFD simulations usually take anything between many hours to even days, depending on the amount of the information that needs to be processed i.e. geometry, boundary conditions, initial parameters like velocities, viscosity etc. byteLAKE, a company specializing in machine and deep learning, has been developing a collection of Artificial Intelligence (AI) models that are targeted to significantly reduce time to results for such simulations.

We invite you to a moderated panel discussion where byteLAKE co-founders, together with a producer of the leading CFD tool for enterprise mixing analysis (MixIT), a company named Tridiagonal Solutions will discuss how Deep Learning models accelerate complex chemical mixing simulations. Panelists will talk about how such simulations help address various industries challenges, explain how AI helps reduce the cost of trial & errors experiments and discuss the future of AI in the CFD space. We will also have representatives from Lenovo Data Center and Intel Corporation who will weigh in on scalability of the technology, and how various hardware configurations can deliver maximum value for AI+CFD adopters.

Most machine learning algorithms depend on the assumption that training and testing data are sampled independently from the same distribution. But what happens when this assumption doesn't hold? At Qminers, we are facing this problem constantly, since data from financial markets are notoriously non-stationary.
In this talk, we will discuss the different faces of distribution shift and how to fight it, both in theory and practice. Topics we will touch on include Robustness, Risk-Aversion and Invariant Risk Minimization. I will show that distribution shift is a real problem faced by ML practitioners, and that solutions exist.

Join this session to learn about our experience and challenges in designing a virtual assistant engine used in many of our IT data center monitoring products. We will discuss both the use case we have solved, its evolution as we encountered challenges and some of the machine learning models implemented to solve these problems. We will also provide in-depth review of the Kernel Density Estimation model we used to study time series in order to obtain the expected value ranges. When anomalies are detected in the expected values, we perform some additional learning on these anomalies to detect patterns and auto-correlate live events in the data center. During the session, we will share both our learnings and challenges building this enterprise grade virtual assistant.

For knowledge-based organizations, finding precise expertise to address specific projects is increasingly important. At McKinsey we’ve been improving our internal expertise search capability, by enriching colleague profiles in various ways, including ML-driven recommendations for ‘Topics to call me about’. Through a number of innovations, our Prague-based Data Science team has created a highly-effective prediction model.
 
Traditionally, expert profiling and retrieval are based on document retrieval approaches. But can the information available in profiles be used to train a supervised model? As with many retrieval applications, our challenges began with a limited amount of data available, as well as the format, which at McKinsey is mostly PowerPoint files. Several well-known approaches were combined to perform a Document Classification step in unsupervised fashion, providing data to create the expert-candidate representations. In a later step, profiling of the experts was achieved despite noisy label data (incomplete profiles) and a large amount of features (compared to the amount of samples available).
 
Despite these challenges, the model is now achieving 80% acceptance of recommendations. In turn this is materially helping the Firm find appropriate experts when needed.

AI and machine learning models are increasingly used to inform high-stakes decisions. Discrimination by AI becomes objectionable when it places certain privileged groups at systematic advantage and certain unprivileged groups at systematic disadvantage. In this talk, we will discuss the sources of unwanted bias in AI, and how it manifests along various points in the AI pipeline. We will also explore several methods of bias mitigation. Finally, we will discuss how bias can be measured and mitigated using the open source AI Fairness 360 toolkit.

The first part of the session underpins the importance of Machine Learning Interpretation. Fundamentally, it is needed because machine learning by itself is incomplete as a solution. After all, the problems they solve are not deterministic, so the solution cannot cover all of it because it is an optimization. One of the most significant issues is that AI faces today is overconfidence. Given the high accuracy of AI solutions, we tend to increase our confidence level to the point we fully understand the problem. Then, we are misled into thinking our solution covers all of it! The machine learning interpretability toolkit can help us first learn from our models.  Then, leverage what was learned or our domain knowledge to place guardrails, mitigate bias, and enhance model reliability, making them safe to use even in rare and unexpected situations and free from non-discriminatory practices. One of the ways in which fairness can be ensured is through monotonic constraints. We will discuss several scenarios in which this may be needed. 

During the second part, we will dive into a law school scholarship problem. Let’s suppose a law school wants to handout merit-based scholarships to those students most likely to pass the bar exam. To that end, they want to train classifiers that score students on this probability. However, the classifier must be consistent with merit-based norms such as having the highest grades in other examinations. Employing monotonic constraints in XGBoost and Tensorflow will place the guardrails so that students with high examination scores are never unfairly penalized.  We will walk through the code that assesses, establishes, and confirms model fairness.

Federated Learning is a new approach that is picking up steam in the machine learning community as a way to improve global models by leveraging on-device training on user data. At WWDC 2019, Apple announced Private Federated Learning by combining Federated Learning with Differential Privacy. We have started to use this technology in iOS 13 for a variety of use cases, including QuickType keyboard, Found in Apps, and Smart Replies. In this talk, Vojta will provide more details about this technique.

Join my session to learn about published and unpublished studies in which we developed and applied: (i) a deep learning and NLP strategy to mine bacterial genomes in order to identify natural product with antibacterial activity (Hannigan et al., 2019, Nucleic Acids Research) (ii) a bundle of constraint-satisfaction, optimization, heuristics and backtracking algorithms to enable design of novel proteins that can be used in research, therapeutics and industrial processes (Hiraga et al., 2021, ACS Synthetic Biology) (iii) An integrated constraint-satisfaction approach to design, optimize and visualise, mRNA-based therapeutics (Vostrosablin et al., 2021, in preparation).

While most of the machine learning methods assume that samples are vectors, matrices, or sequences, in many real-world problems they have a rich structure. While this structure makes the manual design of features non-trivial, I see it as an inductive bias that should drive the design of models. In this talk, I will introduce a simple, yet powerful framework for learning on structured data. A side, yet important feature is the explainability of decisions, which is the result of ingesting data as-is instead of devising artificial features. A concrete implementation of the framework will be demoed on data from various stages of analysis of malware.

Voice biometry is a technology that overperforms humans. Petr Schwarz will present how modern voice biometry systems are built and how they are deployed. The key issues are how to collect data, what are the input features describing the human vocal tract, what machine learning techniques are used for modeling, how to train the models, and how to deliver the model to its user while keeping the best accuracy.

Machine Learning has become a major part of our lives. As more and more companies and organizations implementing ML-based solutions, we need better understand the ethical aspects of Machine Learning algorithms.

In this talk, we will speak about the key element of this field (Fairness, explainability, bias and more) and give some past examples of ethical problems in the ML field. Alongside that, We will suggest ways to solve or reduce the ethical problem in each ML project and finally, we will learn how companies like Google and Microsoft make their algorithms fairer.

What if we could predict when and where the next crimes will be committed? Crimes in Chicago is a publicly published dataset which reflects the reported incidents of crime that occurred in Chicago since 2001. Using this data, we would like not only to be able to explore specific crimes to find interesting trends, but also predict how many crimes will be taking place next week, and even next month.

One year ago, Prague-based Stories.bi were acquired by Workday. Today, the same team is building its ML-driven augmented analytics for Workday's largest customers. Filip will talk about the concept behind Stories.bi, how it is different and why it's called the next generation of BI&analytics.

During the panel, our speaker will share their view on recent advancements in speech and language processing field, and their personal experience in training large-scale natural language models. We will also discuss how accelerating pace of innovation andadoption in the field of AI leads to fast productization of research results. Panel speakers include researchers and representativesof the Speech and Language Product Groups from Microsoft, as well as researchers from organizations across Central and Eastern Europe, who will share their experience from recent projects as well as thinking on the next innovations in that area.

ModelOps is a holistic approach for rapidly and iteratively moving models through the analytics life cycle so they are deployed faster and deliver expected business value. ModelOps is based on the application development community's DevOps approach. But where DevOps focuses on application development, ModelOps focuses on getting models from the lab through validation, testing and deployment phases as quickly as possible, while ensuring quality results. It also focuses on ongoing monitoring and retraining of models to ensure peak performance.

To help you cross the "last mile" of deployment much faster, and ensure that your analytic models deliver expected value, ModelOps defines people (or culture), process and technology changes that facilitate smooth, efficient and continuous development and deployment of high-impact analytic models.

Join us to this Mastermind session, where experts from different areas will discuss current challenges that often prevent organization to get the full potential of their analytics. In the second half of the session, we will also invite person from audience to come on a stage and share their experience or ask questions.

Critical success factor is having data in a good shape, let’s discusss what could be good practices in terms of organizational point of view or suitable technical tools in order to support data quality. What data quality means in particular? We would like also focus on effectivity within small start-up product teams without dedicated „quality department“ and what could be done in good enough principle. Let’s discuss techniques how to properly monitor incoming data or model itself, setup smart alerts based on time periods comparison, production data issues and how to avoid them, validation automation.

Machine Learning technology has become an important part of the ongoing Fourth Industrial Revolution. This revolution transforms manufacturing industry into a new era through digitization and automation of various processes. Especially Computer Vision-based systems have already brought significant benefits to the companies. These benefits include e.g. increased efficiencies and decreased waste, thus having an impact not only on the company itself but also on the whole society. Let’s discuss together both the applications and implications of ML technology in the manufacturing domain.

In Socialbakers, we analyze around 100M texts every day. We process them using our natural language processing (sentiment, NER, ...) or computer vision (image & video classification) systems. In the session, a large part of our research team will be available to share the knowledge and answer any questions you might have.

Machine learning has been tremendously successful in the last decade. The core concepts for training the models is to use supervision, error backpropagation and stochastic gradient descent. However, many scientists believe that to make steps towards more autonomous AI systems, we need to discover learning approaches that are fundamentally less supervised than the current ones. In this talk, I will describe a project where we attempted to define a system which can evolve for indefinitely long, possibly reach arbitrary complexity, and use no supervision. It is based on an old idea of cellular automaton which can be seen as a special type of a recurrent-convolutional network. I will show examples of interesting behavior that we did observe in automatically constructed models. We were able to discover these interesting automata using a novel metric which measures structured complexity growth in time. This work could be a basis of a new generation of machine learning models which can continue learning in interesting ways in situations where no supervision of even rewards are available.

The Keras ecosystem now has two new members, Keras Tuner and AutoKeras. They are built with AutoML techniques to dramatically reduce the manual work for designing and training deep learning models. They work seamlessly with Keras and TensorFlow for model export, saving, and deployment. The talk not only covers how to use them but their underlying mechanism as well.

Deep neural networks (DNNs) have revolutionized AI due to their remarkable performance in pattern recognition, comprising of both memorizing complex training sets and demonstrating intelligence by generalizing to previously unseen data (test sets). The high generalization performance in DNNs has been explained by several mathematical tools, including optimization, information theory, and resilience analysis. In humans, it is the ability to abstract concepts from examples that facilitates generalization; this presentation describes DNN generalization from that perspective. A recent computational neuroscience study revealed a correlation between abstraction and particular neural firing patterns. We express these brain patterns in a closed-form mathematical expression, termed the “Cognitive Neural Activation metric” (CNA) and apply it to DNNs. Our findings reveal parallels in the mechanism underlying abstraction in DNNs and those in the human brain. Beyond simply measuring similarity to human abstraction, the CNA is able to predict and rate how well a DNN will perform on test sets, and determines the better  network architectures for a given task in a manner not possible with extant tools. These results were validated on a broad range of datasets (including ImageNet and random labeled datasets) and neural architectures.

Machine Learning is one of the key component that enables systems that operate under uncertainty. For example, AI systems and robots might employ sensors together with a machine learned system to identify obstacles. However, such data driven system are far from perfect and can result in failure cases that can jeopardize safety. In this talk we will explore a framework that aims to preserve safety invariants despite the uncertainties in the environment arising due to incomplete information. We will describe various methods to reason about safe plans and control strategies despite perceiving the world through noisy sensors and machine learning systems. We will also consider extensions of these ideas, using high-fidelity simulation, to a sequential decision making framework that considers the trade-off in risk and reward in a near-optimal manner.