The advance of foundation models (FM) makes it possible to avoid parametric tuning for transfer learning , taking advantage of pretrained feature spaces . In this study, we define a framework called IDEAL (Interpretable-by-design DEep learning ALgorithms) which tackles the problem of interpretable transfer learning by recasting the standard supervised classification problem into a function of similarity to a set of prototypes derived from the training data . This framework generalises previously-known prototypical approaches, such as ProtoPNet, xDNN and DNC, and decomposes the overall problem into two inherently connected stages: (A) feature extraction (FE), which maps the raw features of real-world data into a latent space, and (B) identification of representative prototypes and decision making based on similarity and association between the query and the prototypes. This addresses the issue of interpretability (stage B) while retaining the benefits of pretrained deep learning (DL) models. On a range of datasets (CIFAR-10, CIFAR-100, CalTech101, STL-10, Oxford-IIIT Pet, EuroSAT), we demonstrate, through an extensive set of experiments, how the choice of the latent space, prototype selection, and finetuning of the latent space affect accuracy and generalisation of the models on transfer learning scenarios for different backbones. Building upon this knowledge, we demonstrate that the proposed framework helps achieve an advantage over state-of-the-art baselines in class-incremental learning. The key findings can be summarised as follows: (1) the setting allows interpretability through prototypes, (2) lack of finetuning helps circumvent the issue of catastrophic forgetting, allowing efficient class-incremental transfer learning, while mitigating the issue of confounding bias, and (3) ViT architectures narrow the gap between finetuned and non-finetuned models allowing for transfer learning in a fraction of time without finetuning of the feature space on a target dataset with iterative supervised methods.

Foundation Models (FM) developed on very large generic data sets transformed the landscape of machine learning (ML). Vision transformers (ViT) closed the gap in performance between fine-tuned and unsupervised transfer learning. This opens the possibility to abandon the widely used until recently end-to-end approach. Instead, we consider a two-stage ML pipeline, where the first stage constitutes extracting features by pre-training large, multi-layer model with billions of parameters, and the second stage is a computationally lightweight learning of an entirely new, simpler model architecture based on prototypes within this feature space. In this paper we consider such two-stage approach to ML. We further analyse the use of several alternative light-weight methods in the second stage, including strategies for semi-supervised learning and a variety of strategies for linear fine-tuning. We demonstrate on the basis of nine well known benchmark data sets that the ultra-light-weight ML alternatives for the second stage (such as clustering, PCA, LDA and combinations of these) offer for the price of negligible drop in accuracy a significant (several orders of magnitude) drop of computational costs (time, energy and related CO2emissions) as well as the ability to use no labels (fully unsupervised approach) or limited amount of labels (one per cluster labels) and the ability to address interpretability.

In this paper, we apply the attention mechanism to autonomous driving for steering angle prediction. We propose the first model, applying the recently introduced sparse attention mechanism to visual domain, as well as the aggregated extension for this model. We show the improvement of the proposed method, comparing to no attention as well as to different types of attention.

Reinforcement learning methods have recently achieved impressive results on a wide range of control problems. However, especially with complex inputs, they still require an extensive amount of training data in order to converge to a meaningful solution. This limits their applicability to complex input spaces such as video signals, and makes them impractical for use in complex real world problems, including many of those for video based control. Supervised learning, on the contrary, is capable of learning on a relatively limited number of samples, but relies on arbitrary hand-labelling of data rather than task-derived reward functions, and hence do not yield independent control policies. In this article we propose a novel, model-free approach, which uses a combination of reinforcement and supervised learning for autonomous control and paves the way towards policy based control in real world environments. We use SpeedDreams/TORCS video game to demonstrate that our approach requires much less samples (hundreds of thousands against millions or tens of millions) comparing to the state-of-the-art reinforcement learning techniques on similar data, and at the same time overcomes both supervised and reinforcement learning approaches in terms of quality. Additionally, we demonstrate applicability of the method to MuJoCo control problems.

The problem of multiple group object tracking is a challenging one and has been extensively researched during the last two decades. The problem solution, proposed in this article, is an extension of the well known multi-Bernoulli filter. The model is first formulated generally, then the linear Gaussian-Wishart implementation is proposed. Contrary to many known methods, which track either individual targets or groups, the proposed filter tracks individual objects and combines them into group tracks. The experiments on simulated data have shown a sustainable ability of the method to track group objects.

In this paper, we propose an approach to data analysis, which is based entirely on the empirical observations of discrete data samples and the relative proximity of these points in the data space. At the core of the proposed new approach is the typicality—an empirically derived quantity that resembles probability. This nonparametric measure is a normalized form of the square centrality (centrality is a measure of closeness used in graph theory). It is also closely linked to the cumulative proximity and eccentricity (a measure of the tail of the distributions that is very useful for anomaly detection and analysis of extreme values). In this paper, we introduce and study two types of typicality, namely its local and global versions. The local typicality resembles the well-known probability density function (pdf), probability mass function, and fuzzy set membership but differs from all of them. The global typicality, on the other hand, resembles well-known histograms but also differs from them. A distinctive feature of the proposed new approach, empirical data analysis (EDA), is that it is not limited by restrictive impractical prior assumptions about the data generation model as the traditional probability theory and statistical learning approaches are. Moreover, it does not require an explicit and binary assumption of either randomness or determinism of the empirically observed data, their independence, or even their number (it can be as low as a couple of data samples). The typicality is considered as a fundamental quantity in the pattern analysis, which is derived directly from data and is stated in a discrete form in contrast to the traditional approach where a continuous pdf is assumed a priori and estimated from data afterward. The typicality introduced in this paper is free from the paradoxes of the pdf. Typicality is objectivist while the fuzzy sets and the belief-based branch of the probability theory are subjectivist. The local typicality is expressed in a closed analytical form and can be calculated recursively, thus, computationally very efficiently. The other nonparametric ensemble properties of the data introduced and studied in this paper, namely, the square centrality, cumulative proximity, and eccentricity, can also be updated recursively for various types of distance metrics. Finally, a new type of classifier called naïve typicality-based EDA class is introduced, which is based on the newly introduced global typicality. This is only one of the wide range of possible applications of EDA including but not limited for anomaly detection, clustering, classification, control, prediction, control, rare events analysis, etc., which will be the subject of further research.

In this paper, a novel empirical data analysis approach (abbreviated as EDA) is introduced which is entirely data-driven and free from restricting assumptions and pre-defined problem- or user-specific parameters and thresholds. It is well known that the traditional probability theory is restricted by strong prior assumptions which are often impractical and do not hold in real problems. Machine learning methods, on the other hand, are closer to the real problems but they usually rely on problem- or user-specific parameters or thresholds making it rather art than science. In this paper we introduce a theoretically sound yet practically unrestricted and widely applicable approach that is based on the density in the data space. Since the data may have exactly the same value multiple times we distinguish between the data points and unique locations in the data space. The number of data points, k is larger or equal to the number of unique locations, l and at least one data point occupies each unique location. The number of different data points that have exactly the same location in the data space (equal value), f can be seen as frequency. Through the combination of the spatial density and the frequency of occurrence of discrete data points, a new concept called multimodal typicality, τMMis proposed in this paper. It offers a closed analytical form that represents ensemble properties derived entirely from the empirical observations of data. Moreover, it is very close (yet different) from the histograms, from the probability density function (pdf) as well as from fuzzy set membership functions. Remarkably, there is no need to perform complicated pre-processing like clustering to get the multimodal representation. Moreover, the closed form for the case of Euclidean, Mahalanobis type of distance as well as some other forms (e.g. cosine-based dissimilarity) can be expressed recursively making it applicable to data streams and online algorithms. Inference/estimation of the typicality of data points that were not present in the data so far can be made. This new concept allows to rethink the very foundations of statistical and machine learning as well as to develop a series of anomaly detection, clustering, classification, prediction, control and other algorithms.

In this article the novel clustering and regression methods TEDACluster and TEDAPredict methods are described additionally to recently proposed evolving classifier TEDAClass. The algorithms for classification, clustering and regression are based on the recently proposed AnYa type fuzzy rule based system. The novel methods use the recently proposed TEDA framework capable of recursive processing of large amounts of data. The framework is capable of computationally cheap exact update of data per sample, and can be used for training `from scratch'. All three algorithms are evolving that is they are capable of changing its own structure during the update stage, which allows to follow the changes within the model pattern.