Raul Vicente

Aqeel Labash, Florian Fletzer, Daniel Majoral et al.

Adapting to regularities of the environment is critical for biological organisms to anticipate events and plan. A prominent example is the circadian rhythm corresponding to the internalization by organisms of the $24$-hour period of the Earth's rotation. In this work, we study the emergence of circadian-like rhythms in deep reinforcement learning agents. In particular, we deployed agents in an environment with a reliable periodic variation while solving a foraging task. We systematically characterize the agent's behavior during learning and demonstrate the emergence of a rhythm that is endogenous and entrainable. Interestingly, the internal rhythm adapts to shifts in the phase of the environmental signal without any re-training. Furthermore, we show via bifurcation and phase response curve analyses how artificial neurons develop dynamics to support the internalization of the environmental rhythm. From a dynamical systems view, we demonstrate that the adaptation proceeds by the emergence of a stable periodic orbit in the neuron dynamics with a phase response that allows an optimal phase synchronisation between the agent's dynamics and the environmental rhythm.

Aqeel Labash, Ardi Tampuu, Tambet Matiisen et al.

Assisted by neural networks, reinforcement learning agents have been able to solve increasingly complex tasks over the last years. The simulation environment in which the agents interact is an essential component in any reinforcement learning problem. The environment simulates the dynamics of the agents' world and hence provides feedback to their actions in terms of state observations and external rewards. To ease the design and simulation of such environments this work introduces $\texttt{APES}$, a highly customizable and open source package in Python to create 2D grid-world environments for reinforcement learning problems. $\texttt{APES}$ equips agents with algorithms to simulate any field of vision, it allows the creation and positioning of items and rewards according to user-defined rules, and supports the interaction of multiple agents.

Marharyta Domnich, Julius Välja, Rasmus Moorits Veski et al.

As machine learning models evolve, maintaining transparency demands more human-centric explainable AI techniques. Counterfactual explanations, with roots in human reasoning, identify the minimal input changes needed to obtain a given output and, hence, are crucial for supporting decision-making. Despite their importance, the evaluation of these explanations often lacks grounding in user studies and remains fragmented, with existing metrics not fully capturing human perspectives. To address this challenge, we developed a diverse set of 30 counterfactual scenarios and collected ratings across 8 evaluation metrics from 206 respondents. Subsequently, we fine-tuned different Large Language Models (LLMs) to predict average or individual human judgment across these metrics. Our methodology allowed LLMs to achieve an accuracy of up to 63% in zero-shot evaluations and 85% (over a 3-classes prediction) with fine-tuning across all metrics. The fine-tuned models predicting human ratings offer better comparability and scalability in evaluating different counterfactual explanation frameworks.

Marharyta Domnich, Raul Vicente

A pressing issue in the adoption of AI models is the increasing demand for more human-centric explanations of their predictions. To advance towards more human-centric explanations, understanding how humans produce and select explanations has been beneficial. In this work, inspired by insights of human cognition we propose and test the incorporation of two novel biases to enhance the search for effective counterfactual explanations. Central to our methodology is the application of diffusion distance, which emphasizes data connectivity and actionability in the search for feasible counterfactual explanations. In particular, diffusion distance effectively weights more those points that are more interconnected by numerous short-length paths. This approach brings closely connected points nearer to each other, identifying a feasible path between them. We also introduce a directional coherence term that allows the expression of a preference for the alignment between the joint and marginal directional changes in feature space to reach a counterfactual. This term enables the generation of counterfactual explanations that align with a set of marginal predictions based on expectations of how the outcome of the model varies by changing one feature at a time. We evaluate our method, named Coherent Directional Counterfactual Explainer (CoDiCE), and the impact of the two novel biases against existing methods such as DiCE, FACE, Prototypes, and Growing Spheres. Through a series of ablation experiments on both synthetic and real datasets with continuous and mixed-type features, we demonstrate the effectiveness of our method.

Eduard Barbu, Marharyta Domnich, Raul Vicente et al.

This study presents insights gathered from surveys and discussions with specialists in three domains, aiming to find essential elements for a universal explanation framework that could be applied to these and other similar use cases. The insights are incorporated into a software tool that utilizes GP algorithms, known for their interpretability. The applications analyzed include a medical scenario (involving predictive ML), a retail use case (involving prescriptive ML), and an energy use case (also involving predictive ML). We interviewed professionals from each sector, transcribing their conversations for further analysis. Additionally, experts and non-experts in these fields filled out questionnaires designed to probe various dimensions of explanatory methods. The findings indicate a universal preference for sacrificing a degree of accuracy in favor of greater explainability. Additionally, we highlight the significance of feature importance and counterfactual explanations as critical components of such a framework. Our questionnaires are publicly available to facilitate the dissemination of knowledge in the field of XAI.

Dmytro Shvetsov, Joonas Ariva, Marharyta Domnich et al.

Deep learning is dramatically transforming the field of medical imaging and radiology, enabling the identification of pathologies in medical images, including computed tomography (CT) and X-ray scans. However, the performance of deep learning models, particularly in segmentation tasks, is often limited by the need for extensive annotated datasets. To address this challenge, the capabilities of weakly supervised semantic segmentation are explored through the lens of Explainable AI and the generation of counterfactual explanations. The scope of this research is development of a novel counterfactual inpainting approach (COIN) that flips the predicted classification label from abnormal to normal by using a generative model. For instance, if the classifier deems an input medical image X as abnormal, indicating the presence of a pathology, the generative model aims to inpaint the abnormal region, thus reversing the classifier's original prediction label. The approach enables us to produce precise segmentations for pathologies without depending on pre-existing segmentation masks. Crucially, image-level labels are utilized, which are substantially easier to acquire than creating detailed segmentation masks. The effectiveness of the method is demonstrated by segmenting synthetic targets and actual kidney tumors from CT images acquired from Tartu University Hospital in Estonia. The findings indicate that COIN greatly surpasses established attribution methods, such as RISE, ScoreCAM, and LayerCAM, as well as an alternative counterfactual explanation method introduced by Singla et al. This evidence suggests that COIN is a promising approach for semantic segmentation of tumors in CT images, and presents a step forward in making deep learning applications more accessible and effective in healthcare, where annotated data is scarce.

Marharyta Domnich, Rasmus Moorits Veski, Julius Välja et al.

Counterfactual explanations are a widely used approach in Explainable AI, offering actionable insights into decision-making by illustrating how small changes to input data can lead to different outcomes. Despite their importance, evaluating the quality of counterfactual explanations remains an open problem. Traditional quantitative metrics, such as sparsity or proximity, fail to fully account for human preferences in explanations, while user studies are insightful but not scalable. Moreover, relying only on a single overall satisfaction rating does not lead to a nuanced understanding of why certain explanations are effective or not. To address this, we analyze a dataset of counterfactual explanations that were evaluated by 206 human participants, who rated not only overall satisfaction but also seven explanatory criteria: feasibility, coherence, complexity, understandability, completeness, fairness, and trust. Modeling overall satisfaction as a function of these criteria, we find that feasibility (the actionability of suggested changes) and trust (the belief that the changes would lead to the desired outcome) consistently stand out as the strongest predictors of user satisfaction, though completeness also emerges as a meaningful contributor. Crucially, even excluding feasibility and trust, other metrics explain 58% of the variance, highlighting the importance of additional explanatory qualities. Complexity appears independent, suggesting more detailed explanations do not necessarily reduce satisfaction. Strong metric correlations imply a latent structure in how users judge quality, and demographic background significantly shapes ranking patterns. These insights inform the design of counterfactual algorithms that adapt explanatory qualities to user expertise and domain context.

Jaan Aru, Aqeel Labash, Oriol Corcoll et al.

Theory of Mind is an essential ability of humans to infer the mental states of others. Here we provide a coherent summary of the potential, current progress, and problems of deep learning approaches to Theory of Mind. We highlight that many current findings can be explained through shortcuts. These shortcuts arise because the tasks used to investigate Theory of Mind in deep learning systems have been too narrow. Thus, we encourage researchers to investigate Theory of Mind in complex open-ended environments. Furthermore, to inspire future deep learning systems we provide a concise overview of prior work done in humans. We further argue that when studying Theory of Mind with deep learning, the research's main focus and contribution ought to be opening up the network's representations. We recommend researchers use tools from the field of interpretability of AI to study the relationship between different network components and aspects of Theory of Mind.

Oriol Corcoll, Youssef Mohamed, Raul Vicente

Identifying controllable aspects of the environment has proven to be an extraordinary intrinsic motivator to reinforcement learning agents. Despite repeatedly achieving State-of-the-Art results, this approach has only been studied as a proxy to a reward-based task and has not yet been evaluated on its own. Current methods are based on action-prediction. Humans, on the other hand, assign blame to their actions to decide what they controlled. This work proposes Controlled Effect Network (CEN), an unsupervised method based on counterfactual measures of blame to identify effects on the environment controlled by the agent. CEN is evaluated in a wide range of environments showing that it can accurately identify controlled effects. Moreover, we demonstrate CEN's capabilities as intrinsic motivator by integrating it in the state-of-the-art exploration method, achieving substantially better performance than action-prediction models.

Florian Stelzer, André Röhm, Raul Vicente et al.

Deep neural networks are among the most widely applied machine learning tools showing outstanding performance in a broad range of tasks. We present a method for folding a deep neural network of arbitrary size into a single neuron with multiple time-delayed feedback loops. This single-neuron deep neural network comprises only a single nonlinearity and appropriately adjusted modulations of the feedback signals. The network states emerge in time as a temporal unfolding of the neuron's dynamics. By adjusting the feedback-modulation within the loops, we adapt the network's connection weights. These connection weights are determined via a back-propagation algorithm, where both the delay-induced and local network connections must be taken into account. Our approach can fully represent standard Deep Neural Networks (DNN), encompasses sparse DNNs, and extends the DNN concept toward dynamical systems implementations. The new method, which we call Folded-in-time DNN (Fit-DNN), exhibits promising performance in a set of benchmark tasks.

Oriol Corcoll, Raul Vicente

Exploration and credit assignment under sparse rewards are still challenging problems. We argue that these challenges arise in part due to the intrinsic rigidity of operating at the level of actions. Actions can precisely define how to perform an activity but are ill-suited to describe what activity to perform. Instead, causal effects are inherently composable and temporally abstract, making them ideal for descriptive tasks. By leveraging a hierarchy of causal effects, this study aims to expedite the learning of task-specific behavior and aid exploration. Borrowing counterfactual and normality measures from causal literature, we disentangle controllable effects from effects caused by other dynamics of the environment. We propose CEHRL, a hierarchical method that models the distribution of controllable effects using a Variational Autoencoder. This distribution is used by a high-level policy to 1) explore the environment via random effect exploration so that novel effects are continuously discovered and learned, and to 2) learn task-specific behavior by prioritizing the effects that maximize a given reward function. In comparison to exploring with random actions, experimental results show that random effect exploration is a more efficient mechanism and that by assigning credit to few effects rather than many actions, CEHRL learns tasks more rapidly.

Raul Vicente

Deep learning has sparked a network of mutual interactions between different disciplines and AI. Naturally, each discipline focuses and interprets the workings of deep learning in different ways. This diversity of perspectives on deep learning, from neuroscience to statistical physics, is a rich source of inspiration that fuels novel developments in the theory and applications of machine learning. In this perspective, we collect and synthesize different intuitions scattered across several communities as for how deep learning works. In particular, we will briefly discuss the different perspectives that disciplines across mathematics, physics, computation, and neuroscience take on how deep learning does its tricks. Our discussion on each perspective is necessarily shallow due to the multiple views that had to be covered. The deepness in this case should come from putting all these faces of deep learning together in the reader's mind, so that one can look at the same problem from different angles.

Anti Ingel, Ilya Kuzovkin, Raul Vicente

In this work, a classification method for SSVEP-based BCI is proposed. The classification method uses features extracted by traditional SSVEP-based BCI methods and finds optimal discrimination thresholds for each feature to classify the targets. Optimising the thresholds is formalised as a maximisation task of a performance measure of BCIs called information transfer rate (ITR). However, instead of the standard method of calculating ITR, which makes certain assumptions about the data, a more general formula is derived to avoid incorrect ITR calculation when the standard assumptions are not met. This allows the optimal discrimination thresholds to be automatically calculated and thus eliminates the need for manual parameter selection or performing computationally expensive grid searches. The proposed method shows good performance in classifying targets of a BCI, outperforming previously reported results on the same dataset by a factor of 2 in terms of ITR. The highest achieved ITR on the used dataset was 62 bit/min. The proposed method also provides a way to reduce false classifications, which is important in real-world applications.

Aqeel Labash, Jaan Aru, Tambet Matiisen et al.

Perspective taking is the ability to take the point of view of another agent. This skill is not unique to humans as it is also displayed by other animals like chimpanzees. It is an essential ability for social interactions, including efficient cooperation, competition, and communication. Here we present our progress toward building artificial agents with such abilities. We implemented a perspective taking task inspired by experiments done with chimpanzees. We show that agents controlled by artificial neural networks can learn via reinforcement learning to pass simple tests that require perspective taking capabilities. We studied whether this ability is more readily learned by agents with information encoded in allocentric or egocentric form for both their visual perception and motor actions. We believe that, in the long run, building better artificial agents with perspective taking ability can help us develop artificial intelligence that is more human-like and easier to communicate with.

Tambet Matiisen, Aqeel Labash, Daniel Majoral et al.

Inferring other agents' mental states such as their knowledge, beliefs and intentions is thought to be essential for effective interactions with other agents. Recently, multiagent systems trained via deep reinforcement learning have been shown to succeed in solving different tasks, but it remains unclear how each agent modeled or represented other agents in their environment. In this work we test whether deep reinforcement learning agents explicitly represent other agents' intentions (their specific aims or goals) during a task in which the agents had to coordinate the covering of different spots in a 2D environment. In particular, we tracked over time the performance of a linear decoder trained to predict the final goal of all agents from the hidden state of each agent's neural network controller. We observed that the hidden layers of agents represented explicit information about other agents' goals, i.e. the target landmark they ended up covering. We also performed a series of experiments, in which some agents were replaced by others with fixed goals, to test the level of generalization of the trained agents. We noticed that during the training phase the agents developed a differential preference for each goal, which hindered generalization. To alleviate the above problem, we propose simple changes to the MADDPG training algorithm which leads to better generalization against unseen agents. We believe that training protocols promoting more active intention reading mechanisms, e.g. by preventing simple symmetry-breaking solutions, is a promising direction towards achieving a more robust generalization in different cooperative and competitive tasks.

Jaan Aru, Raul Vicente

Can deep learning (DL) guide our understanding of computations happening in biological brain? We will first briefly consider how DL has contributed to the research on visual object recognition. In the main part we will assess whether DL could also help us to clarify the computations underlying higher cognitive functions such as Theory of Mind. In addition, we will compare the objectives and learning signals of brains and machines, leading us to conclude that simply scaling up the current DL algorithms will most likely not lead to human level Theory of Mind.

Ardi Tampuu, Tambet Matiisen, Dorian Kodelja et al.

Multiagent systems appear in most social, economical, and political situations. In the present work we extend the Deep Q-Learning Network architecture proposed by Google DeepMind to multiagent environments and investigate how two agents controlled by independent Deep Q-Networks interact in the classic videogame Pong. By manipulating the classical rewarding scheme of Pong we demonstrate how competitive and collaborative behaviors emerge. Competitive agents learn to play and score efficiently. Agents trained under collaborative rewarding schemes find an optimal strategy to keep the ball in the game as long as possible. We also describe the progression from competitive to collaborative behavior. The present work demonstrates that Deep Q-Networks can become a practical tool for studying the decentralized learning of multiagent systems living in highly complex environments.