Machine Learning Prague 2024

LLMs have shown impressive performance across a wide range of tasks, including tasks that require reasoning.  Despite these results, there is considerable evidence that the reasoning capabilities of LLMs have serious limitations.  Given the significant benefits of reasoning, including improved generalization, ability to incorporate new knowledge, better explainability and reliability, it is important to better understand what LLMs can and cannot be expected to do.  In this talk I will provide an overview of reasoning and its benefits, characterize some of the reasoning limitations of LLMS, and discuss approaches to move toward AI systems that can reason more effectively.

The efficient semantic search plays a pivotal role in the Seznam.cz search engine, with the vector space method proving to be a fast and effective means of mining relevant documents for user queries. However, the effectiveness of this approach is contingent upon the alignment of the vector space with the specific requirements of the search engine. This presentation delves into the realm of contrastive learning, an efficient technique for acquiring optimal vector representations of diverse modalities, including text, images, or both. While contrastive learning offers a robust foundation, the preferential approach for downstream tasks involves fine-tuning models on labeled data. Our talk will shed light on the innovative application of contrastive learning within the fine-tuning phase of semantic models, leveraging noisy “click” user data to augment the search engine's capabilities. We will present compelling real-world use cases spanning various modalities within the Seznam.cz search domain, including images, products, web documents, and advertisements.

Quantum-inspired and Quantum Machine Learning offer the potential to solve problems intractable for classical computing. This presentation delves into the advantages of these approaches, where quantum computing can accelerate traditional machine learning methods, where traditional machine learning methods provide an advantage, and with a look at potential applications.

Recommender systems have evolved into an indispensable part of our daily experiences, delivering tailored recommendations across diverse applications that cater to individual preferences. Yet, traditional recommendation models often struggle with the complexities of comprehensively understanding user interests and adapting to the ever-expanding landscape of recommendation scenarios. Enter Large Language Models (LLMs), renowned for their unparalleled strength in language comprehension, generation, generalization and reasoning. In this talk, we will explore the transformative influence of LLMs on recommender systems. We'll provide a comprehensive overview of the strategies for seamlessly integrating LLMs into recommendation pipelines, shedding light on where and how these integrate effortlessly into the recommendation ecosystem while addressing practical challenges and opportunities that emerge. We will also delve into methodologies such as fine-tuning and prompt tuning, showcasing how these techniques unlock the full potential of LLMs. Finally, we will lay out a roadmap of exciting opportunities, offering a glimpse into a future where recommendations are not only highly personalized but also exceptionally efficient.

Have you ever wondered if the typical semantic search recipe of “embed all your documents and do the vector search” holds up in practice? Will it fit a narrow medical, law, or hardware domain? Can it handle non-English languages well? Taking a one-size-fits-all model from Huggingface does not always do the trick for your specific business needs. But the good news is that you can fine-tune such models to better fit your world, and open-source tools like sentence-transformers and nixietune are here to help you! In this talk, we will take a hands-on journey to fine-tune your own embedding model for semantic search. What kind and how much data do you need, how do you train a bi-encoder and a cross-encoder, and how far can you go quality-wise with a single GPU? If you're ready to make semantic search work for you, come join us!

While LLMs outperform humans in an ever-broadening range of tasks, they remain far behind in the ability to explain somebody else’s feelings, thoughs and behavioral drivers. This skill requires perspective-taking, the process of conceptualizing the point of view of another person. Something that humans do effortlessly and instinctively all the time - tracking the concealed mental states of others, encompassing knowledge, intentions, beliefs, and desires. This ability holds a central role in various aspects of human existence, including social interactions, self-consciousness, and ability to empathize. Foreseeing the actions or reactions of others is also for LLMs a key to choose the best action to take next. Enhancing perspective-taking capabilities of LLMs can unlock their potential to react better and safer to hints of sadness, anger, or distress, to recognize sarcasm, to engage in a more receptive argumentation, or to target an explanation to an audience. In this talk, I will review recent approaches to modelling the human context in transformers holistically -  including personal history, beliefs, and social environment – and their effect on downstream NLP tasks. I will present our recent perspectivism-motivated experiments, and discuss further opportunities for bringing the human-centered paradigm into LLMs.

Genomics is a rapidly expanding field that has recently benefited from a convergence of numerous technological and scientific advancements. In genomics, we confront a unique set of challenges, such as intricate data pipelines and metrics, alongside evaluation methods that are still not well established. In this presentation we will delve deeper into the field of metagenomics and how we use machine learning to classify microorganisms from free samples. We will show how to combine transformer models and hashing models for maximal performance.

In recent years, Transformers and self-supervised learning have achieved unprecedented success. This is exemplified by models like ChatGPT, which have garnered significant attention even among the general public. While numerous pre-trained language models are readily available for text data, either as commercial APIs or open-source models, many companies deal with large volumes of tabular, time-series, or graph data. This necessitates the creation and training of custom models for many diverse use cases. Map matching refers to the procedure of converting sparse and noisy signals into precise positions on a road network. Drawing inspiration from the methods that propelled success in the text domain, such as language modeling, machine translation, and transfer learning, Dataclair, AI centre at O2 Czech Republic, has trained its own sequence-to-sequence Transformer models, which effectively translate mobile network signals into specific road segments. By providing more accurate aggregated and anonymized data, telecommunication companies can offer valuable insights to public decision-makers, for example aiding in enhanced traffic planning.

Beyond its potential for infectious disease prophylaxis, RNA technology has emerged as one of the most promising techniques for the rapid development of treatments envisaging a wide range of pathologies ranging from cancer to cardiovascular, autoimmune, or rare diseases. While conventional vaccines and pharmaceuticals rely on the active use of cell cultures, RNA manufacturing is based on a relatively simple, scalable, and affordable cell-free production line. Furthermore, RNA-based technology is known to hold the major advantage of being disease-agnostic, hence exhibiting high versatility in terms of therapeutic scope. Nevertheless, a decisive factor in the successful design of such innovative therapeutics for personalized medicine is the use of big data and AI technologies. These account not only for the fast RNA sequencing and proteomics data analysis but also for the identification and selection of viable and relevant RNA codes with low immunogenicity, maximum potency, and superior production yields. Multi-omics data integration helps molecule design and discovery, ensuring drug connectivity analysis and drug response prediction. This presentation will showcase how to harness the potential of a disease-agnostic RNA platform for personalized medicine by making use of custom-designed big data architecture, generative AI, and deep learning but also biotech digital twins and software sensors, all within the framework of Quality by Digital Design (QbDD) with fast and interactive visualizations provided by the solution’s designated dashboard. All these have led to the development and up-scaling of the existing production processes both in terms of volume as well as timelines at unprecedented speed.

As large language models (LLMs) become more widely adopted, it is crucial to understand their effective utilization, copilot development, evaluation, operationalization, and monitoring in real-world applications. This session will provide insights into incorporating responsible AI practices and safety features into your generative AI applications. You will gain knowledge on assessing your copilots and generative AI applications, mitigating content-related risks, addressing hallucinations, jailbreak, and copywrite issues, ensuring fairness, and enhancing the overall quality and safety of your copilot.

Development of smart/autonomous cars relies on testing in the real world, with real traffic. During this type of testing, thousands of hours of camera footage is recorded and used for learning, improving and testing the autonomous mode for the cars. We have developed and studied several algorithms that can find human faces and potentially readable license plates in every frame of a video and anonymize it. Why do we anonymize? When is it mandatory? How much of it is necessary?

The Kepler and TESS missions have yielded an astounding 100,000 potential transit signals, paving the way for an intricate process of distillation to identify viable exoplanet candidates. In response to this formidable challenge, we introduced ExoMiner, a groundbreaking deep neural network meticulously designed for the classification of transit signals in the search for exoplanets. ExoMiner played a pivotal role in validating and authenticating 301 previously undiscovered exoplanets. This keynote voyage commences with a sweeping overview of the captivating realm of exoplanetary exploration. As we delve into the heart of our presentation, we embark on an exploration of the distinctive attributes that set ExoMiner apart, unraveling the intricate web of factors that contribute to its remarkable accuracy. Our narrative extends beyond the confines of ExoMiner as we navigate through the integration of additional machine learning models seamlessly layered atop the ExoMiner architecture using basic domain knowledge information. This collaborative synthesis has already led to the identification of an additional 69 exoplanets, showcasing our unwavering commitment to pushing the boundaries of exoplanetary discovery. The journey of implementing machine learning in domains of such complexity is riddled with practical challenges. Guaranteeing consistent performance and earning the trust of domain experts in the reliability of results becomes an arduous task. This keynote will intricately delve into the practical intricacies, providing profound insights into the delicate balance between machine intelligence and the discerning eye of scientific scrutiny.

Deep leaning became one of the most active areas of research in many fields including artificial intelligence (AI), computer vision, autonomous driving, factory automation, robotics, and medicine. Deep learning is a breakthrough technology in the machine learning field, as it exceeded human performance when it was trained with a large number of cases, "big data". In many areas such as medical image diagnosis and low-volume productions in factories, however, collecting and annotating a large number of cases of patients or products is a big challenge. Deep learning models require 10,000 to 100,000 cases of patients or factory parts to adequately train, which would take months or years to collect. In the past research, less attentions were given to deep learning models that did not require big data. In this talk, "small-data" deep learning models that can be trained with a limited number of cases is introduced. Small-data deep learning is defined as deep learning models that offer the performance equivalent to the "big-data" deep-learning models but require a small number of cases (< 100) to adequately train, thus, reducing the number of necessary cases by a factor of 100-1,000. My group has been actively studying on deep learning in medical imaging in the past 25 years, including ones of the earliest deep-learning models for medical image processing, semantic segmentation of lesions and organs, lesion/organ enhancement, and detection and classification of lesions in medical imaging. We investigated the development of novel small-data deep-learning models and their applications to AI-aided diagnostic systems (“AI doctor”) and deep-learning-based imaging for medical diagnosis (“AI medical imaging”). We investigated a required number of cases for our small-data deep-learning model for AI-aided analysis of lesions in medical images. We demonstrated that our small-data deep-learning model, called a massive-training artificial neural network (MTANN), was able to achieve the state-of-the-art performance with a small number of cases (> 100). In the application to liver tumor segmentation, our small-data deep-learning model achieved the performance equivalent to the world-competition-winning deep-learning model with a very small required number of training cases of only 14. Our other verified studies include 1) AI systems for cancer detection, analysis, and diagnosis with medical images, and 2) virtual AI imaging systems for separation of bones from soft tissue in chest radiographs and those for radiation dose reduction in CT and mammography. Thus, small-data deep-learning models that can be trained with a small number of cases would be the next breakthrough technology in the AI field, and it would fill in the gap between "big-data" areas and "small-data" areas where many cases are not available such as in medicine, healthcare, and low-volume production industries.

In the rapidly evolving landscape of the travel industry, the ability to adapt and respond in real-time is not just an advantage, but a necessity. This talk offers an insightful overview of real-time online machine learning (ML), a pivotal area in modern data science. We will explore the core principles of online ML, contrasting it with traditional batch learning, and highlighting its significance in processing continuous data streams. Key focus areas include a review of various online learning algorithms, their adaptability to evolving data, and the balance between responsiveness and stability. The presentation will examine both the advantages, such as real-time data handling and adaptability, and the challenges, including computational demands and potential overfitting, associated with online ML methods. Through select examples, we will demonstrate the practical applications and limitations of these techniques across different sectors. Concluding with a look at future trends in online ML, this talk aims to provide a concise yet comprehensive understanding of the field, enabling participants to appreciate its impact and potential in diverse applications.

Churn rate as the rate of customers dropping out of a service, is an important indicator for many modern companies. Its detection thus plays a key part in customer retention and satisfaction. The detection is usually done by an AI/ML model based on neural networks or gradient boosted decision trees. Nevertheless, the sole detection of churn happening is not enough by itself. Knowing a customer is likely to drop out, it is necessary to take an action which makes them stay. Such an action might cover various options such as discounts, coupons, direct calls or email campaigns as so on. It is essential to determine the correct cause to avoid unwanted wasted resources. The second task of analysis and prediction of causes which makes a person leave is even harder than the churn detection itself. In this article we cover the process of identifying churn and with such knowledge we provide a Bayesian network based model which provides insights of reasons for such a churn happening. We demonstrate our method and results on a real world example of a taxi-hailing company which is the segment with large customer fluctuation.

The growing sophistication of Business Email Compromise (BEC) and spear phishing attacks poses significant challenges to organizations worldwide. The techniques featured in traditional spam and phishing detection are insufficient due to the tailored nature of modern BEC attacks as they often blend in with the regular benign traffic. Recent advances in machine learning, particularly in Natural Language Understanding (NLU), offer a promising avenue for combating such attacks but in a practical system, due to limitations such as data availability, operational costs, verdict explainability requirements or a need to robustly evolve the system, it is essential to combine multiple approaches together. We present CAPE, a comprehensive and efficient system for BEC detection that has been proven in a production environment for a period of over two years. Rather than being a single model, CAPE is a system that combines independent ML models and algorithms detecting BEC-related behaviors across various email modalities such as text, images, metadata and the email’s communication context. This decomposition makes CAPE’s verdicts naturally explainable. We describe the design principles and constraints behind its architecture, as well as the challenges of model design, evaluation and adapting the system continuously through a Bayesian approach that combines limited data with domain knowledge. Furthermore, we showcase recent developments that integrate generative LLMs into the detection pipeline.

Geometric Deep Learning (GDL) has emerged as a pivotal paradigm, enabling machines to glean profound insights from structured data such as graphs and meshes. This talk delves into the intricacies of constructing a robust and efficient Geometric Deep Learning Pipeline, a key cornerstone for unlocking the potential of geometric data.

The presentation commences with an exploration of the fundamental principles underpinning Geometric Deep Learning, elucidating its unique capacity to process and interpret non-Euclidean data structures. Attendees will gain a comprehensive understanding of how GDL extends the frontiers of traditional deep learning methodologies, offering unparalleled capabilities in tasks ranging from graph classification to 3D shape recognition.

With Retrieval-Augmented Generation (RAG), you can ask GPT models questions about your own data and get back answers with references and quotes from it. In other words, your answer generation is augmented with smart retrieval of provided data. Nvidia predicts that RAG will be one of the hottest AI topics in 2024. Applying the latest LLMs in your business can be tricky. You most likely have a lot of classified or simply niche, company-specific data that you want the LLM to take into consideration. Additionally, you can not have your LLM spout damaging hallucinations. How much background context you can send into LLMs is currently bounded by both technical and cost limitations. This is where RAG systems excel. A RAG system can automatically search your internal data for the most relevant data to answer each query and force the LLM to base its answer entirely on data provided by your defined tools. RAG is a relatively new and fast evolving technique with lots of contributors. Some common pitfalls and issues tend to crop up regardless of industry. That is why we want to share our progress with everyone so that you, too, can create actual business value with LLMs and contribute to the field. In this talk we will cover the most common problems and their solutions. For example: * asking questions about multiple sources at the same time, * adding more tools and subquerying (e.g. internal databases, web search, API calls), * handling different types of data (SQL, JSON etc.), * validating your system output (ROUGE, BERTScore, Gaia etc.) and * prohibiting topics with guardrails (like politics, talking about competitors etc.) The less technically inclined (the technically reclined, if you will), will still enjoy an analogy-heavy presentation filled with pop culture references in order to preserve client anonymity. Neural networks designed for natural language understanding are often inspired by how human brains function. They share many of the same strengths and weaknesses, which makes the whole thing easier to grasp! Entecon is a consulting firm based in Stockholm, Sweden. With a long history in NLP, we have worked with a wide variety of clients and domains. Our AI R&D consultants frequently take experimental ML theory into practice.

The necessity of combating scams resonates strongly in the face of escalating financial losses, reaching billions of dollars worldwide. Scammers use diverse communication channels such as social networks, emails, phone calls, and text messages to contact their victims. This diversity poses a significant technical challenge for designing scam-detection applications. Overcoming these technical obstacles introduces the challenge of determining whether a received message is a scam. While some scams display obvious indicators and are easily detected, categorizing other messages definitively requires additional contextual information, as these messages may not be decidable based solely on the textual content. This presentation provides insights into the methodology employed for training Large Language Models (LLMs) specifically designed for scam detection. Throughout the project, we grappled with the challenge of limited and non-comprehensive datasets, all while meeting accuracy requirements and tight deadlines. We will describe the strategic decisions and considerations taken to navigate through this triad (data scarcity, accuracy demands, and timely delivery), leading to the successful application deployment of robust scam-detection models.

Large Language Models (LLMs) have emerged as both a technological marvel and a source of fascination in the field of artificial intelligence. While they hold great promise, there is still an incredible journey from toying with LLMs to utilization in enterprise-grade solutions. Our presentation emphasizes the lessons learned and insights gained from practical LLM model productization for a specific HR data interpretation use case. We will share the journey from model selection and fine-tuning to dataset curation and model evaluation. We address issues related to hallucination, testing, model bias, ethical considerations, and data privacy, emphasizing the importance of responsible AI development and deployment.