Wojciech Masarczyk

Bartosz Grabowski, Przemysław Głomb, Wojciech Masarczyk et al.

Interpretation of electrocardiography (ECG) signals is required for diagnosing cardiac arrhythmia. Recently, machine learning techniques have been applied for automated computer-aided diagnosis. Machine learning tasks can be divided into regression and classification. Regression can be used for noise and artifacts removal as well as resolve issues of missing data from low sampling frequency. Classification task concerns the prediction of output diagnostic classes according to expert-labeled input classes. In this work, we propose a deep neural network model capable of solving regression and classification tasks. Moreover, we combined the two approaches, using unlabeled and labeled data, to train the model. We tested the model on the MIT-BIH Arrhythmia database. Our method showed high effectiveness in detecting cardiac arrhythmia based on modified Lead II ECG records, as well as achieved high quality of ECG signal approximation. For the former, our method attained overall accuracy of 87:33% and balanced accuracy of 80:54%, on par with reference approaches. For the latter, application of self-supervised learning allowed for training without the need for expert labels. The regression model yielded satisfactory performance with fairly accurate prediction of QRS complexes. Transferring knowledge from regression to the classification task, our method attained higher overall accuracy of 87:78%.

Jim Zhao, Tin Sum Cheng, Wojciech Masarczyk et al.

Neural Collapse (NC) refers to the emergence of highly symmetric geometric structures in the representations of deep neural networks during the terminal phase of training. Despite its prevalence, the theoretical understanding of NC remains limited. Existing analyses largely ignore the role of the optimizer, thereby suggesting that NC is universal across optimization methods. In this work, we challenge this assumption and demonstrate that the choice of optimizer plays a critical role in the emergence of NC. The phenomenon is typically quantified through NC metrics, which, however, are difficult to track and analyze theoretically. To overcome this limitation, we introduce a novel diagnostic metric, NC0, whose convergence to zero is a necessary condition for NC. Using NC0, we provide theoretical evidence that NC cannot emerge under decoupled weight decay in adaptive optimizers, as implemented in AdamW. Concretely, we prove that SGD, SignGD with coupled weight decay (a special case of Adam), and SignGD with decoupled weight decay (a special case of AdamW) exhibit qualitatively different NC0 dynamics. Also, we show the accelerating effect of momentum on NC (beyond convergence of train loss) when trained with SGD, being the first result concerning momentum in the context of NC. Finally, we conduct extensive empirical experiments consisting of 3,900 training runs across various datasets, architectures, optimizers, and hyperparameters, confirming our theoretical results. This work provides the first theoretical explanation for optimizer-dependent emergence of NC and highlights the overlooked role of weight-decay coupling in shaping the implicit biases of optimizers.

Paweł Wawrzyński, Wojciech Masarczyk, Mateusz Ostaszewski

Effective reinforcement learning requires a proper balance of exploration and exploitation defined by the dispersion of action distribution. However, this balance depends on the task, the current stage of the learning process, and the current environment state. Existing methods that designate the action distribution dispersion require problem-dependent hyperparameters. In this paper, we propose to automatically designate the action distribution dispersion using the following principle: This distribution should have sufficient dispersion to enable the evaluation of future policies. To that end, the dispersion should be tuned to assure a sufficiently high probability (densities) of the actions in the replay buffer and the modes of the distributions that generated them, yet this dispersion should not be higher. This way, a policy can be effectively evaluated based on the actions in the buffer, but exploratory randomness in actions decreases when this policy converges. The above principle is verified here on challenging benchmarks Ant, HalfCheetah, Hopper, and Walker2D, with good results. Our method makes the action standard deviations converge to values similar to those resulting from trial-and-error optimization.

Wojciech Masarczyk, Tomasz Trzciński, Mateusz Ostaszewski

In the era of transfer learning, training neural networks from scratch is becoming obsolete. Transfer learning leverages prior knowledge for new tasks, conserving computational resources. While its advantages are well-documented, we uncover a notable drawback: networks tend to prioritize basic data patterns, forsaking valuable pre-learned features. We term this behavior "feature erosion" and analyze its impact on network performance and internal representations.

Wojciech Masarczyk, Mateusz Ostaszewski, Tin Sum Cheng et al. · deepmind

The softmax function is a fundamental building block of deep neural networks, commonly used to define output distributions in classification tasks or attention weights in transformer architectures. Despite its widespread use and proven effectiveness, its influence on learning dynamics and learned representations remains poorly understood, limiting our ability to optimize model behavior. In this paper, we study the pivotal role of the softmax function in shaping the model's representation. We introduce the concept of rank deficit bias - a phenomenon in which softmax-based deep networks find solutions of rank much lower than the number of classes. This bias depends on the softmax function's logits norm, which is implicitly influenced by hyperparameters or directly modified by softmax temperature. Furthermore, we demonstrate how to exploit the softmax dynamics to learn compressed representations or to enhance their performance on out-of-distribution data. We validate our findings across diverse architectures and real-world datasets, highlighting the broad applicability of temperature tuning in improving model performance. Our work provides new insights into the mechanisms of softmax, enabling better control over representation learning in deep neural networks.

Wojciech Masarczyk, Mateusz Ostaszewski, Ehsan Imani et al.

Deep neural networks are widely known for their remarkable effectiveness across various tasks, with the consensus that deeper networks implicitly learn more complex data representations. This paper shows that sufficiently deep networks trained for supervised image classification split into two distinct parts that contribute to the resulting data representations differently. The initial layers create linearly-separable representations, while the subsequent layers, which we refer to as \textit{the tunnel}, compress these representations and have a minimal impact on the overall performance. We explore the tunnel's behavior through comprehensive empirical studies, highlighting that it emerges early in the training process. Its depth depends on the relation between the network's capacity and task complexity. Furthermore, we show that the tunnel degrades out-of-distribution generalization and discuss its implications for continual learning.

Wojciech Masarczyk, Paweł Wawrzyński, Daniel Marczak et al.

We introduce a neural network architecture that logarithmically reduces the number of self-rehearsal steps in the generative rehearsal of continually learned models. In continual learning (CL), training samples come in subsequent tasks, and the trained model can access only a single task at a time. To replay previous samples, contemporary CL methods bootstrap generative models and train them recursively with a combination of current and regenerated past data. This recurrence leads to superfluous computations as the same past samples are regenerated after each task, and the reconstruction quality successively degrades. In this work, we address these limitations and propose a new generative rehearsal architecture that requires at most logarithmic number of retraining for each sample. Our approach leverages allocation of past data in a~set of generative models such that most of them do not require retraining after a~task. The experimental evaluation of our logarithmic continual learning approach shows the superiority of our method with respect to the state-of-the-art generative rehearsal methods.

Wojciech Masarczyk, Kamil Deja, Tomasz Trzciński

Catastrophic forgetting of previously learned knowledge while learning new tasks is a widely observed limitation of contemporary neural networks. Although many continual learning methods are proposed to mitigate this drawback, the main question remains unanswered: what is the root cause of catastrophic forgetting? In this work, we aim at answering this question by posing and validating a set of research hypotheses related to the specificity of representations built internally by neural models. More specifically, we design a set of empirical evaluations that compare the robustness of representations in discriminative and generative models against catastrophic forgetting. We observe that representations learned by discriminative models are more prone to catastrophic forgetting than their generative counterparts, which sheds new light on the advantages of developing generative models for continual learning. Finally, our work opens new research pathways and possibilities to adopt generative models in continual learning beyond mere replay mechanisms.

Kamil Deja, Paweł Wawrzyński, Wojciech Masarczyk et al.

We propose a new method for unsupervised generative continual learning through realignment of Variational Autoencoder's latent space. Deep generative models suffer from catastrophic forgetting in the same way as other neural structures. Recent generative continual learning works approach this problem and try to learn from new data without forgetting previous knowledge. However, those methods usually focus on artificial scenarios where examples share almost no similarity between subsequent portions of data - an assumption not realistic in the real-life applications of continual learning. In this work, we identify this limitation and posit the goal of generative continual learning as a knowledge accumulation task. We solve it by continuously aligning latent representations of new data that we call bands in additional latent space where examples are encoded independently of their source task. In addition, we introduce a method for controlled forgetting of past data that simplifies this process. On top of the standard continual learning benchmarks, we propose a novel challenging knowledge consolidation scenario and show that the proposed approach outperforms state-of-the-art by up to twofold across all experiments and the additional real-life evaluation. To our knowledge, Multiband VAE is the first method to show forward and backward knowledge transfer in generative continual learning.

Mateusz Ostaszewski, Lea M. Trenkwalder, Wojciech Masarczyk et al.

The study of Variational Quantum Eigensolvers (VQEs) has been in the spotlight in recent times as they may lead to real-world applications of near-term quantum devices. However, their performance depends on the structure of the used variational ansatz, which requires balancing the depth and expressivity of the corresponding circuit. In recent years, various methods for VQE structure optimization have been introduced but the capacities of machine learning to aid with this problem has not yet been fully investigated. In this work, we propose a reinforcement learning algorithm that autonomously explores the space of possible ans{ä}tze, identifying economic circuits which still yield accurate ground energy estimates. The algorithm is intrinsically motivated, and it incrementally improves the accuracy of the result while minimizing the circuit depth. We showcase the performance of our algorithm on the problem of estimating the ground-state energy of lithium hydride (LiH). In this well-known benchmark problem, we achieve chemical accuracy, as well as state-of-the-art results in terms of circuit depth.

Kamil Deja, Paweł Wawrzyński, Daniel Marczak et al.

We introduce a binary latent space autoencoder architecture to rehearse training samples for the continual learning of neural networks. The ability to extend the knowledge of a model with new data without forgetting previously learned samples is a fundamental requirement in continual learning. Existing solutions address it by either replaying past data from memory, which is unsustainable with growing training data, or by reconstructing past samples with generative models that are trained to generalize beyond training data and, hence, miss important details of individual samples. In this paper, we take the best of both worlds and introduce a novel generative rehearsal approach called BinPlay. Its main objective is to find a quality-preserving encoding of past samples into precomputed binary codes living in the autoencoder's binary latent space. Since we parametrize the formula for precomputing the codes only on the chronological indices of the training samples, the autoencoder is able to compute the binary embeddings of rehearsed samples on the fly without the need to keep them in memory. Evaluation on three benchmark datasets shows up to a twofold accuracy improvement of BinPlay versus competing generative replay methods.

Wojciech Masarczyk, Ivona Tautkute

Catastrophic forgetting is a problem caused by neural networks' inability to learn data in sequence. After learning two tasks in sequence, performance on the first one drops significantly. This is a serious disadvantage that prevents many deep learning applications to real-life problems where not all object classes are known beforehand; or change in data requires adjustments to the model. To reduce this problem we investigate the use of synthetic data, namely we answer a question: Is it possible to generate such data synthetically which learned in sequence does not result in catastrophic forgetting? We propose a method to generate such data in two-step optimisation process via meta-gradients. Our experimental results on Split-MNIST dataset show that training a model on such synthetic data in sequence does not result in catastrophic forgetting. We also show that our method of generating data is robust to different learning scenarios.

Wojciech Masarczyk, Przemysław Głomb, Bartosz Grabowski et al.

Hyperspectral imaging is a rich source of data, allowing for multitude of effective applications. However, such imaging remains challenging because of large data dimension and, typically, small pool of available training examples. While deep learning approaches have been shown to be successful in providing effective classification solutions, especially for high dimensional problems, unfortunately they work best with a lot of labelled examples available. To alleviate the second requirement for a particular dataset the transfer learning approach can be used: first the network is pre-trained on some dataset with large amount of training labels available, then the actual dataset is used to fine-tune the network. This strategy is not straightforward to apply with hyperspectral images, as it is often the case that only one particular image of some type or characteristic is available. In this paper, we propose and investigate a simple and effective strategy of transfer learning that uses unsupervised pre-training step without label information. This approach can be applied to many of the hyperspectral classification problems. Performed experiments show that it is very effective at improving the classification accuracy without being restricted to a particular image type or neural network architecture. The experiments were carried out on several deep neural network architectures and various sizes of labeled training sets. The greatest improvement in overall accuracy on the Indian Pines and Pavia University datasets is over 21 and 13 percentage points, respectively. An additional advantage of the proposed approach is the unsupervised nature of the pre-training step, which can be done immediately after image acquisition, without the need of the potentially costly expert's time.