Doktoranci

Data that we need to handle in real-world is never static and keeps on changing, e.g., vertices are added to social networks, or new ties appear. Hence, typically, we need to update the solution to our problem constantly. This not only posses efficiency issus, but requires that we do not change the solution too much each time. In this research challange we want to face these problems from new perspective and create algorithms that can learn and adopt to changes.

In this research project we aim to propose tools that would provide explanations for the different basic optimization problems, e.g., assignment problem, shortest paths, minimum cuts, or basic graphical neural networks. This research is motivated by the fact that even when faced with problems that can be solved exactly, we still would like to understand why this solutions was computed.

ML tools enter as interior components into basic data structures or state-of-theart approximation algorithms resulting in solutions that have better practical properties, e.g., indices. These new hybrid constructions are called learned data-structures. As the work on these ideas has just started we miss the right framework and tools for implementing state-of-art solutions and thus the research on new tools and models is hampered. This research aims to continue research on this problem and create new algorithms and data structures together with their implementations. This could prove tools to bridge the gap between theory and practice in algorithms and show that new theoretical advances can have practical implications.

Although, different parallel computation models have be studied for years already. A new model that describes real-world systems has been proposed recently – the Massively Parallel Computation (MPC) frameworks includes systes such as MapReduce, Hadoop, Spark, or Flume. It comes with a completely new possibilities as well as requirements. MPC computations are executed in synchronous rounds, but implementing these rounds on real-world systems takes considerable time. One round takes orders of magnitude longer than on classical Cray type system. Thus we would like to solve problems, in particular graph problems, in as few rounds as possible. With this challenge in mind, this project aims to design methods to break barriers that were impossible to overcome using classical techniques and models. More specifically, we are going to work on new algorithmic tools that would improve efficiency of both parallel and non-parallel algorithms used in data science.

In recent years we observe a huge progress in development of deep NLP models. In many applications these models can effectively compete with humans, and there usage is growths. However, the main works on these models are limited to mayor languages, and recent developments are not directly available for Polish language. The aim of this project is twofold: develop cutting edge NLP models for Polish language; use the experience gained this was to extend and improve models for other languages.

In this project we aim to work on multi-purpose and multi-modal neural networks with special emphasis on physical applications and usage. The tasks we aim to cope with will different problems where we aim to integrate different kind of information and aim to deliver joint representation that would allow for example: translate text to images and vice-versa for general and medical usage; transform natural language to animations, or approach no-code programming challenges.”, a jego tytuł na “Universal and Multi-modal Neural Networks with Emphasis on Physical Applications”.

Knowledge transfer is key for obtaining a good performance for complex tasks. Intuitively, it is much more effective to pre-train a model on related (and perhaps easier/cheaper tasks) and later ‘just to fine-tune’ it to a new task. Such approaches have been widespread in practice. However, they lack a proper understanding in the case of neural networks. It is not clear what is really transfer, if these are useful features, or perhaps good weight initialization. In the project, we plan to evaluate existing hypotheses explaining transfer in the case of control tasks (e.g. robotic manipulation).

Experience replay has proven to be one of the most powerful technique mitigating forgetting in long sequences of tasks. Its main drawback is large usage of memory, which prevents in scalability for long sequences. This project aims for a systematic study of this experience replay techniques with the goal of making them more efficient. To this end, we conceptualize two major tasks:
– what are the quantitative and qualitative properties of the experience replay samples, which are needed for successful mitigation of forgetting
– what are mechanism of experience replay

For the second questions, we speculate that the experience replay loss gradients are sparse and can be distilled into much more compact form. Perhaps, also they could be factorized with respect to network weights and, therefore, expressed as a sum of simple per-weight losses.

Learning a long sequence of tasks might be facilitated by active representation learning. In this project, we aim to study two high level questions. The first one is, how much the structure of the space can facilitate efficient learning. For example, one can introduce a learning bias such that representations related to various tasks can be easily disentangled, for example expressed linearly. The second questions, is how much data augmentations can facilitate forming better representations.

When dealing with problems that require long-term planning, the search depth often needs to be reduced due to a large branching factor (for example, solving the Rubik’s cube). One promising solution to this problem is the use of subgoals, which are intermediate milestones towards the final solution. Some previous implementations of this concept have already demonstrated impressive results by allowing for deeper search and solving problems with much lower computational costs. The project aims to explore the design and testing of new methods related to subgoals for a diverse range of problems.

Neural networks has brought spectacular progress in solving many problems. In some cases, however, we expect that they have latural limitations and cannot solve each problem also. An archtypical example concerns combinatorial puzzles, like Rubik’s cube, but has much broader applicability in discrete optimization. The project will explore how to use neural network efficiently with other computational mechanisms (e.g. classical search techniqes). The core question is understanding situations, in which neural networks make errors and in which can be trusted. A proper analysis should lead to more efficient planning methods.

Transformers have been extremely successful architectures in sequential modeling, however, they have a practical limitation of a relatively short context span due to the quadratic cost of the attention mechanism. The project aims to explore practical solutions to mitigate this problem by providing access to an external memory, which can be thought of as a external knowledge system. The aim is to factorise the reasoning capabilites, which could be stored in the weights of transformer, from trivia facts, which can be stored in memory.

Large language models like GPTs have revolutionized the field of machine learning by introducing new ways of learning, such as in-context learning, chain-of-thoughts, and scratchpads. Interestingly, they also appear to possess rudimentary reasoning capabilities. This project aims to investigate how we can improve and utilize these capabilities to achieve better results.

They however struggle in tasks requiring advanced reasoning and planning, think about some complex game. It is an open question how to achieve such capabilities. One angle is to elaborated prompting techniques (e.g. scratchpad), which provide additional ‚working memory’ on which the planning might be achieved. On the other end of the spectrum there are changes in architecture, that would provide an adaptive computation budget depending on the problem difficulty. Finally, one can also hope to achieve better performance by training on structured data.

Classical reinforcement learning operates under the assumption of perfect knowledge of the environment, which is only applicable in limited, idealized scenarios. In more typical situations, an agent only has access to a subset of information about the environment, particularly in multi-agent systems like traffic control, where an agent’s understanding of other agents’ intentions may be limited. Our project aims to explore this by scaling the number of agents and observing patterns that emerge, with the goal of designing improved control mechanisms.

One of the trends observed in highly developed countries is the focus on the digitization of pathology (digital pathology) and the use of algorithms supporting the work of pathologists (computational pathology). Due to massive amounts of available clinical data, it is possible to train machine learning models to do this.  The research objective is to develop novel solutions supporting evaluation of the metabolic syndrome and non-alcoholic fatty liver disease using histological data (whole slide images). Data used in the project will be histological slides. The project is focused on image data analysis by using computer vision and artificial intelligence methods.

This research focuses on applying machine learning techniques to develop highly agile robotic systems. Agility in robotics is crucial for tasks requiring quick and precise movements. The thesis investigates various machine learning models and algorithms that can enhance the agility of robots, enabling them to perform complex maneuvers with greater speed and accuracy. The work includes developing new algorithms and adapting existing ones, specifically tailored for agility in robotic systems.

This thesis examines the application of lifelong learning in legged locomotion for robots operating in challenging environments. Lifelong learning is a machine learning approach where a system continuously learns and adapts over time. The research addresses how legged robots can improve their locomotion capabilities over their operational lifespan, adapting to various terrains and overcoming obstacles. The study includes the development of algorithms that enable these robots to learn from their experiences and refine their locomotion strategies accordingly.

This topic investigates the concept of learned dynamic manipulation in robotic systems. The research focuses on how machine learning can be employed to enhance a robot’s ability to manipulate objects dynamically. This includes the development of algorithms that allow robots to learn and adapt their manipulation strategies in real-time, improving their efficiency and effectiveness in tasks such as meal preparation, or handling of moving objects.

The effects of catastrophic winds or snowfalls sometimes result in many thousands of hectares of forest being overturned and destroyed. The damaged areas are very dangerous, so it is difficult to inventory them in order to determine the economic damage associated with the event or to plan future activities. In these areas, trees are lying on top of each other, often with varying degrees of damage, making it virtually impossible to move around the area on the ground. Another example of an area with lying trees is a situation where foresters plan to harvest raw wood for the timber industry. This involves cutting down trees that are then lying on the ground, and the forester has to measure each one, which is often labour-intensive and sometimes dangerous. The aim of this project is to use AI to detect and measure fallen trees in order to automatically inventory them on site. The development of automatic recognition methods for measuring lying trees is therefore of great practical and cognitive importance. On the one hand, research in this area is quite limited, but on the other hand, the development of such tools will improve the quality and safety of the work of many people involved in forest management and protection.

How can individuals and communities protect their privacy against social network analysis techniques, algorithms and other tools? How do criminals or terrorists organizations evade detection by such tools? Under which conditions can these tools be made strategy proof? These fundamental questions have attracted little attention in the literature to date, as most tools are built around the assumption that individuals or groups in a network do not act strategically to evade social network analysis. To address this issue, a recently novel paradigm is social network analysis explicitely models strategic behaviour of network actors using the aparatus of game theory. Addressing this research challenge has various implications. For instance, it may allow two individuals to keep their relationship secret or private. It may also allow members of an activist group to conceal their membership, or even conceal the existence of their group from authoritarian regimes. Furthermore, it may assist security agencies and counter terrorism units in understanding the strategies that covert organizations use to escape detection, and give rise to new strategy-proof countermeasures.

Social Networks have become a primary media for cybercrimes. For instance, attackers may compromise accounts to diffuse misinformation (e.g., fake news, rumors, hate speeches, etc.) through a social network. Fraudsters may also trick innocent customers into conducting fraudulent transactions over online trading platforms. Meanwhile, on the defense side, defenders (e.g., network administrators) are increasingly employing machine-learning-based tools to detect malicious behaviors. Graph Neural Networks (GNNs) have become the \textit{de facto} choice of social detection tools due to their superior performance over a wide spectrum of tasks. In this project, the overall goal is to develop robust and effective GNN-based social detection tools in an adversarial environment. This goal is decomposed into three coherent objectives. First, design more effective GNN-based tools to detect crimes in social networks that could achieve a better detection accuracy as well as a lower false positive rate. Second, from the standpoint of an attacker, investigate effective evasion techniques to bypass the detection of the GNN-based tools. Third, as a defender, enhance the robustness of the GNN-based detection tools to mitigate evasion attacks. Overall, the expected outcomes significantly advance our knowledge in developing trustworthy AI systems in a real-world adversarial environment.

Machine learning, especially deep learning, has transformed the way how data is processed. Recent studies have revealed that deep learning systems lack transparency and are also vulnerable to adversarial attacks. The fundamental reason is that deep learning systems rely on a large amount of data possibly collected from the wild, which gives the opportunity for attacks to inject \textit{adversarial noise} to mislead the systems. Meanwhile, an active line of research, termed Explainable AI (XAI), aims to interpret the decisions made by AI systems, which essentially identify a subset of data that is important for the decision. In this project, we will investigate how to use XAI to build robust deep learning systems against attacks. This goal is decomposed into two major objectives. First, enhance existing or develop new XAI techniques to effectively identify adversarial noises from data. That is, we employ more advanced XAI to sanitize the data for deep learning systems. Second, provided with the sanitation results, develop new algorithms to train robust deep learning systems from the \textit{noisy} data. Overall, the expected outcomes will make significant contributions toward developing more transparent and robust deep learning systems.

Cryptocurrency based on blockchain technology has significantly reduced our dependence on the central authority. Meanwhile, due to its anonymity nature, cryptocurrency trading platforms have also become the perfect media for financial crimes. For example, many known studies have revealed that criminals are increasingly using Bitcoin transaction networks for money laundering. Thus, a very significant while underexplored problem is how to effectively detect fraudsters in bitcoin transaction networks utilizing machine learning techniques. The major objectives of this project are as follows. First, design unsupervised machine learning algorithms (e.g., clustering, contrastive learning, etc) to effectively identify fraudulent transactions and malicious accounts in a transaction network. Essentially, this objective calls for new approaches to detect anomalies at the node level, edge level, and subgraph level within a graph. Second, investigate the vulnerabilities of prior detection methods by designing more practical evasion techniques. Especially, besides considering the evasion objective, the design of evasion attacks should simultaneously consider the need for stealthiness and preserving malicious utilities. Third, faced with strategical evaders, further improve the robustness of the detection methods. Successfully achieving these objectives will contribute to applying unsupervised machine learning in anomaly detection from a technical perspective, and enhancing the security of the trading environment of cryptocurrencies.

Federated learning is a computation paradigm for training machine learning models from distributed data while preserving data privacy. Most of the existing research has been devoted to investigating federated learning algorithms over well-structured data such as tabular data. Since graphs are widely used to represent various kinds of relational data (e.g., social networks, recommendation systems, communication networks, etc.), there is an urgent need to investigate and design new federated learning algorithms for graphs. Especially, graphs have some unique features which make previous algorithms not suitable. For example, the features of nodes in graphs are highly heterogeneous, which makes federated training algorithms hard to converge. Also, graphs distributed into different subgraphs will inevitably miss those interconnected edges, which represent a kind of information loss for learning. Thus, in this project, the primary goal is to design new federated learning algorithms for graph learning models (e.g., graph neural networks) over distributed graphs. In expectation, these algorithms will mitigate a series of issues of learning over graph data, including heterogeneity, information loss, and so on.

The objective is to enable a paradigm shift from correlation-driven to scaled-up causality-driven machine learning. The project deals with the unresolved learning challenge in logical engineering, the scaling challenge of probabilistic causal models, and the correlationreliance of deep learning using explainable AI, advanced time series analysis, and multimodal deep learning.

Another critical weakness in the vision of decentralizing internet services is that interacting with blockchains is more complicated than in the case of centralized solutions. Moreover, the decentralization makes it impossible to revert the transactions that were posted by mistake or as a result of an attack. Due to this, the users often rely on the socalled hardware wallets, which are dedicated devices protected against cyber-attacks. This student will work on analyzing the security of the existing hardware wallets. In particular, we will be interested in their side-channel security, i.e., security against attacks based on information such as power consumption or electromagnetic radiation.

Several machine learning applications involve issues where privacy plays a special role. This includes cases in which secrecy applies to the training data (e.g., when it contains medical information) and those in which the algorithm itself is subject to protection because, for example, it reveals specific information about the training data. The student will work on addressing these problems using methods such as multiparty computation protocols, differential privacy, and trusted execution environments.

One of the main problems in the blockchain space is that decentralized solutions are typically more complex and error-prone than centralized ones. In particular, errors in smart contracts can lead to considerable financial losses. Furthermore, some blockchain algorithms in the past had serious mistakes that could be exploited to steal large amounts of money. This student will work on addressing these problems using tools from formal methods from machine learning, in combination with proof assistants and formal theorem provers such as Coq, Easycrypt, Why3, and others.

Large language models (LLMs) have proven to be very strong generalpurpose architectures. Recent successes of systems like chatGPT highlighted several important areas that the research community needs to address. These include truthfulness, alignment, or uncertainty awareness. The lack of truthfulness results in model making stuff up, or hallucinating. We want strong models to be aligned with human values and act in accordance with human intentions. Lastly, models that are not aware of their uncertainty may either unnecessarily withhold information, or hallucinate; whereas if the contrary happens the model can e.g., delegate queries to external APIs. The goal of this research stream is to formulate the aforementioned problems as a solution to a multi-player coopetitive game and via agents autonomous interaction.

Natural language is a very exciting modality, which has opened up to other parts of Machine Learning mostly due to the power of large language models (LLMs). In particular it allows constructing reinforcement learning (RL) agents that can interact with the world via instructions or communicate their internal state in natural language. Furthermore, we can use LLMs as AIgenerated environments to be used in RL, which opens up new possibilities such as performing interventions or asking counterfactual questions in a natural way. The goal is to study the capacity of RL agents to learn in this regime and simultaneously ask questions about LLms consistency, truthfulness, or their knowledge graph.

Classically, RL algorithms use some approximation of future rewards as a learning signal to improve a policy. Recent research has shown that RL can be viewed through the lens of representation learning. Here the premise is that the policy guided by the similarity between representation of the current state and the goal state can be a valid alternative. The goal of this research stream is to investigate old and recent ideas from statistics and selfsupervised learning to propose new RL algorithms. In particular, study the impact of the methods in subgoal search, for subgoal generation, learning the latent, or guiding the search.

Human ethical biases (conscious and unconscious) can significantly skew data and influence the behavior of AI systems; the way data is collected, labelled, and processed affects training datasets that mould AI systems. The aim of this topic is to identify and overcome such biases regarding mental health. When these biases seep into AI algorithms, they perpetuate harmful stereotypes and misinformation, potentially leading to misdiagnosis and improper treatment. One of the most pervasive demographic biases in AI research is the overrepresentation of WEIRD (White, Educated, Industrialized, Rich, Democratic) populations. If data labels categorize certain behaviors as indicative of mental illness, the AI systems will likely perpetuate these misconceptions. When an algorithm is optimized for a certain goal that reflects a human bias (such as prioritizing efficiency over fairness), an AI system such as a therapeutic bot can advise solutions that are ethically skewed. An AI system might e.g. maximize overall utility leading to treatments or resources being disproportionately directed towards groups that are easier or less costly to treat, potentially neglecting those with more complex or costly needs, thus disregarding the bioethical principle of autonomy or individualized care. The broader aim of this topic is to stimulate conversation around the ethical use of AI in mental health, inspire the development of more robust, bias-aware data collection methods, and promote transparency in AI algorithm design.

The main goal of this topic is the design of a digital decision-making tool for mental health professionals working with people on the autism spectrum. In contrast to prevailing perceptions, extensive research underscores the vulnerability of expert decision-making to myriad of systematic cognitive biases, such as representativeness, anchoring or blind spot bias. The project will analyze (and implement in the form of appropriate modules) factors influencing expert decisions, such as: prevalent social stereotypes regarding autism, cognitive humility, cognitive fatigue, etc. to equip experts with support enhancing the precision and reliability of their judgment as well as mitigating the adverse repercussions stemming from cognitive errors that may arise during the evaluation or diagnosis of individuals with autism. The larger context for this topis is prioritizing patient-centered approaches to autism.

The aim of this topis is the algorithmization of some key analytic tools used in applied (qualitative) phenomenology, such as reduction and imaginative variation. Synthetic phenomenology in this context refers to the creation of algorithms specifically designed to analyze the “essential” facets of phenomenal states via linguistic/conceptual representations. The goal is to merge the realms of phenomenology and machine learning by leveraging a neural network architecture trained on carefully preselected corpora of lived experiences. The core challenge is the meticulous selection of first-person training data encapsulating both regular and pathological lived experiences. Another critical aspect involves the application of the phenomenological principle of epoché signifying the suspension of judgment via the implementation of debiasing strategies reflecting various levels of phenomenological reduction. Blending phenomenological methodologies with advanced machine learning tools offers an unprecedented level of insight into subjectivity, unattainable through conventional phenomenological tools reliant solely on the human mind.

This research opportunity focuses on advancing geometry generation for computer games and movie productions through the innovative use of generative modeling techniques. This interdisciplinary project aims to revolutionize content creation by developing novel approaches and architectures to enhance the generation of geometric assets and virtual worlds. In this project, we will build on top of state-of-the-art generative modeling techniques and research novel methods for geometry generation using implicit neural representations (INRs). The application fields of the developed algorithms are generations of large scenes or open worlds for computer games and movie productions. By leveraging the power of generative models, we aim to enable more efficient, realistic, and artistically controllable content creation processes.

This project is an exciting opportunity to delve into the promising and rapidly evolving field of differentiable neural rendering. This is a novel area at the crossroads of machine learning and computer graphics, where traditional rendering techniques meet advanced neural networks. The primary objective of this research is to develop innovative methods for differentiable neural rendering with applications in CGI for visual effects for film, TV, and advertisement productions. The project involves developing a framework where changes in the output image can be traced back to changes in input parameters, such as object shapes, light positions, or material properties. Leveraging this, the project aims to optimize these parameters to improve the quality of the rendered image.

Geometric modeling involves the creation, representation, and manipulation of 2D and 3D shapes, serving as a fundamental tool in various fields such as computer-aided design (CAD), architecture, animation, virtual reality, and industrial manufacturing. This interdisciplinary project aims to leverage the power of neural networks and implicit representations to improve the way we model and manipulate complex 3D shapes. In this project, we will research implicit neural representations, such as neural signed distance functions (SDFs), to advance geometric modeling. Implicit neural representations have shown great potential in capturing intricate shape details. However, efficient shape manipulation of neural representations is challenging and not well-researched. This project seeks to push the boundaries of geometric modeling by harnessing the capabilities of neural networks.

We consider an entity such as a power prosumer which continually trades a certain commodity in the market. This entity requires a technology to automatically designate buy and sell bids for this market. The goal of the project is to develop such a technology within the area of reinforcement learning. Specifically, we address the following issues: How relevant market information can be transformed into a set of bids? How can this transformation be trained in a simulation? How can this transformation be trained based on offline data only?

We consider networks such as the power grid and look to provide the best control of devices in the network. We consider reinforcement learning as a general approach to control optimization. The goal of the project is to design efficient learning algorithms that take into account the setting where many agents learn simultaneously and encounter constraints defined by the network.

The goal of the project is to design neural architectures that generate graphs that meet the given functional requirements. The anticipated use ofthese architectures is the design of structures represented by graphs. For example, given a molecule (a functional requirement) the architecture generates another molecule (output graph) that is in the required relation with the given one, e.g. the second is an inhibitor of the first.