Marcin Pietroń

Dominik Żurek, Kamil Faber, Marcin Pietron et al.

Continual offline reinforcement learning (CORL) aims to learn a sequence of tasks from datasets collected over time while preserving performance on previously learned tasks. This setting corresponds to domains where new tasks arise over time, but adapting the model in live environment interactions is expensive, risky, or impossible. However, CORL inherits the dual difficulty of offline reinforcement learning and adapting while preventing catastrophic forgetting. Replay-based continual learning approaches remain a strong baseline but incur memory overhead and suffer from a distribution mismatch between replayed samples and newly learned policies. At the same time, architectural continual learning methods have shown strong potential in supervised learning but remain underexplored in CORL. In this work, we propose TSN-Affinity, a novel CORL method based on TinySubNetworks and Decision Transformer. The method enables task-specific parameterization and controlled knowledge sharing through a RL-aware reuse strategy that routes tasks according to action compatibility and latent similarity. We evaluate the approach on benchmarks based on Atari games and simulations of manipulation tasks with the Franka Emika Panda robotic arm, covering both discrete and continuous control. Results show strong retention from sparse SubNetworks, with routing further improving multi-task performance. Our findings suggest that similarity-guided architectural reuse is a strong and viable alternative to replay-based strategies in a CORL setting. Our code is available at: https://github.com/anonymized-for-submission123/tsn-affinity.

Kamil Faber, Dominik Zurek, Marcin Pietron et al.

Continual learning (CL) is one of the most promising trends in recent machine learning research. Its goal is to go beyond classical assumptions in machine learning and develop models and learning strategies that present high robustness in dynamic environments. The landscape of CL research is fragmented into several learning evaluation protocols, comprising different learning tasks, datasets, and evaluation metrics. Additionally, the benchmarks adopted so far are still distant from the complexity of real-world scenarios, and are usually tailored to highlight capabilities specific to certain strategies. In such a landscape, it is hard to objectively assess strategies. In this work, we fill this gap for CL on image data by introducing two novel CL benchmarks that involve multiple heterogeneous tasks from six image datasets, with varying levels of complexity and quality. Our aim is to fairly evaluate current state-of-the-art CL strategies on a common ground that is closer to complex real-world scenarios. We additionally structure our benchmarks so that tasks are presented in increasing and decreasing order of complexity -- according to a curriculum -- in order to evaluate if current CL models are able to exploit structure across tasks. We devote particular emphasis to providing the CL community with a rigorous and reproducible evaluation protocol for measuring the ability of a model to generalize and not to forget while learning. Furthermore, we provide an extensive experimental evaluation showing that popular CL strategies, when challenged with our benchmarks, yield sub-par performance, high levels of forgetting, and present a limited ability to effectively leverage curriculum task ordering. We believe that these results highlight the need for rigorous comparisons in future CL works as well as pave the way to design new CL strategies that are able to deal with more complex scenarios.

Maxwell Miller-Golub, Collin Coil, Kamil Faber et al.

Cross-entropy loss has long been the standard choice for training deep neural networks, yet it suffers from interpretability limitations, unbounded weight growth, and inefficiencies that can contribute to costly training dynamics. The harmonic loss is a distance-based alternative grounded in Euclidean geometry that improves interpretability and mitigates phenomena such as grokking, or delayed generalization on the test set. However, the study of harmonic loss remains narrow: only Euclidean distance is explored, and no systematic evaluation of computational efficiency or sustainability was conducted. We extend harmonic loss by systematically investigating a broad spectrum of distance metrics as replacements for the Euclidean distance. We comprehensively evaluate distance-tailored harmonic losses on both vision backbones and large language models. Our analysis is framed around a three-way evaluation of model performance, interpretability, and sustainability. On vision tasks, cosine distances provide the most favorable trade-off, consistently improving accuracy while lowering carbon emissions, whereas Bray-Curtis and Mahalanobis further enhance interpretability at varying efficiency costs. On language models, cosine-based harmonic losses improve gradient and learning stability, strengthen representation structure, and reduce emissions relative to cross-entropy and Euclidean heads. Our code is available at: https://anonymous.4open.science/r/rethinking-harmonic-loss-5BAB/.

Marcin Pietroń, Dominik Żurek, Kamil Faber et al.

Continual Learning (CL) is a process in which there is still huge gap between human and deep learning model efficiency. Recently, many CL algorithms were designed. Most of them have many problems with learning in dynamic and complex environments. In this work new architecture based approach Ada-QPacknet is described. It incorporates the pruning for extracting the sub-network for each task. The crucial aspect in architecture based CL methods is theirs capacity. In presented method the size of the model is reduced by efficient linear and nonlinear quantisation approach. The method reduces the bit-width of the weights format. The presented results shows that low bit quantisation achieves similar accuracy as floating-point sub-network on a well-know CL scenarios. To our knowledge it is the first CL strategy which incorporates both compression techniques pruning and quantisation for generating task sub-networks. The presented algorithm was tested on well-known episode combinations and compared with most popular algorithms. Results show that proposed approach outperforms most of the CL strategies in task and class incremental scenarios.

Radosław Szostak, Marcin Pietroń, Przemysław Wachniew et al.

Unmanned aerial vehicle (UAV) photogrammetry allows for the creation of orthophotos and digital surface models (DSMs) of a terrain. However, DSMs of water bodies mapped with this technique reveal water surface distortions, preventing the use of photogrammetric data for accurate determination of water surface elevation (WSE). Firstly, we propose a new solution in which a convolutional neural network (CNN) is used as a WSE estimator from photogrammetric DSMs and orthophotos. Second, we improved the previously known "water-edge" method by filtering the outliers using a forward-backwards exponential weighted moving average. Further improvement in these two methods was achieved by performing a linear regression of the WSE values against chainage. The solutions estimate the uncertainty of the predictions. This is the first approach in which DL was used for this task. A brand new machine learning data set has been created. It was collected on a small lowland river in winter and summer conditions. It consists of 322 samples, each corresponding to a 10 by 10 meter area of the river channel and adjacent land. Each data set sample contains orthophoto and DSM arrays as input, along with a single ground-truth WSE value as output. The data set was supplemented with data collected by other researchers that compared the state-of-the-art methods for determining WSE using an UAV. The results of the DL solution were verified using k-fold cross-validation method. This provided an in-depth examination of the model's ability to perform on previously unseen data. The WSE RMSEs differ for each k-fold cross-validation subset and range from 1.7 cm up to 17.2 cm. The RMSE results of the improved "water-edge" method are at least six times lower than the RMSE results achieved by the conventional "water-edge" method. The results obtained by new methods are predominantly outperforming existing ones.

Jan Krupiński, Maciej Wielgosz, Szymon Mazurek et al.

This paper presents a computer-aided cytology diagnosis system designed for animals, focusing on image quality assessment (IQA) using Convolutional Neural Networks (CNNs). The system's building blocks are tailored to seamlessly integrate IQA, ensuring reliable performance in disease classification. We extensively investigate the CNN's ability to handle various image variations and scenarios, analyzing the impact on detecting low-quality input data. Additionally, the network's capacity to differentiate valid cellular samples from those with artifacts is evaluated. Our study employs a ResNet18 network architecture and explores the effects of input sizes and cropping strategies on model performance. The research sheds light on the significance of CNN-based IQA in computer-aided cytology diagnosis for animals, enhancing the accuracy of disease classification.

Krystian Strzałka, Szymon Mazurek, Maciej Wielgosz et al.

This paper explores the innovative use of simulation environments to enhance data acquisition and diagnostics in veterinary medicine, focusing specifically on gait analysis in dogs. The study harnesses the power of Blender and the Blenderproc library to generate synthetic datasets that reflect diverse anatomical, environmental, and behavioral conditions. The generated data, represented in graph form and standardized for optimal analysis, is utilized to train machine learning algorithms for identifying normal and abnormal gaits. Two distinct datasets with varying degrees of camera angle granularity are created to further investigate the influence of camera perspective on model accuracy. Preliminary results suggest that this simulation-based approach holds promise for advancing veterinary diagnostics by enabling more precise data acquisition and more effective machine learning models. By integrating synthetic and real-world patient data, the study lays a robust foundation for improving overall effectiveness and efficiency in veterinary medicine.

Jakub Caputa, Maciej Wielgosz, Daria Łukasik et al.

The primary objective of this research was to enhance the quality of semantic segmentation in cytology images by incorporating super-resolution (SR) architectures. An additional contribution was the development of a novel dataset aimed at improving imaging quality in the presence of inaccurate focus. Our experimental results demonstrate that the integration of SR techniques into the segmentation pipeline can lead to a significant improvement of up to 25% in the mean average precision (mAP) segmentation metric. These findings suggest that leveraging SR architectures holds great promise for advancing the state of the art in cytology image analysis.

Michal Karwatowski, Marcin Pietron

Lemmatization is the process of grouping together the inflected forms of a word so they can be analysed as a single item, identified by the word's lemma, or dictionary form. In computational linguistics, lemmatisation is the algorithmic process of determining the lemma of a word based on its intended meaning. Unlike stemming, lemmatisation depends on correctly identifying the intended part of speech and meaning of a word in a sentence, as well as within the larger context surrounding that sentence. As a result, developing efficient lemmatisation algorithm is the complex task. In recent years it can be observed that deep learning models used for this task outperform other methods including machine learning algorithms. In this paper the polish lemmatizer based on Google T5 model is presented. The training was run with different context lengths. The model achieves the best results for polish language lemmatisation process.

Paweł Gajewski, Dominik Żurek, Marcin Pietroń et al.

Reinforcement learning (RL) in robotics faces significant hurdles regarding sample efficiency and generalization across varying goals. While Offline RL mitigates the need for costly online interactions, its integration with goal-conditioned policies and transformer-based architectures remains underexplored. We introduce a Goal-Conditioned Decision Transformer adapted for offline multi-goal robotics. By explicitly incorporating goal states into the sequence modeling framework, our approach efficiently solves varying tasks using only pre-collected data. We validate this method on a newly released offline dataset for the Franka Emika Panda platform. Experimental results demonstrate that our approach outperforms state-of-the-art online baselines in complex tasks and maintains robustness in sparse-reward settings, even with limited expert demonstrations.

Marcin Pietroń

Deep learning models are the most efficient models in many machine learning tasks. The main disadvantage when using them in IoT, mobile devices, independent autonomous or real-time systems is their complexity and memory size. Therefore, much research has concentrated on compression techniques of deep learning architectures. One of the most popular technique is quantization. In most of the works, the quantization is done based on the nearest neighbour quantization technique. This work focuses on improving the quantization efficiency in pretrained and quantized models. This approach has the potential to improve the final accuracy of quantized models. The main postulate of the work is that final quantization states of the network based on nearest neighbour rounding does not guarantee optimal accuracy. In the presented work, the evolution strategy is used as an optimization approach. The evolution in each iteration changes the values of the small percentage of weights. It shifts theirs values to different quantization states. The work shows that proposed evolution with an appropriate set of operators and parameters can fast improve the accuracy of the quantized models. The results are presented for popular architectures such as VGG and Resnet for image classification and detection. Additionally, simulations were carried out for the autoencoder architecture.

Kamil Faber, Marcin Pietroń, Dominik Żurek et al.

The recently proposed xLSTM is a powerful model that leverages expressive multiplicative gating and residual connections, providing the temporal capacity needed for long-horizon forecasting and representation learning. This architecture has demonstrated success in time series forecasting, lossless compression, and even large-scale language modeling tasks, where its linear memory footprint and fast inference make it a viable alternative to Transformers. Despite its growing popularity, no prior work has explored xLSTM for anomaly detection. In this work, we fill this gap by proposing xLSTMAD, the first anomaly detection method that integrates a full encoder-decoder xLSTM architecture, purpose-built for multivariate time series data. Our encoder processes input sequences to capture historical context, while the decoder is devised in two separate variants of the method. In the forecasting approach, the decoder iteratively generates forecasted future values xLSTMAD-F, while the reconstruction approach reconstructs the input time series from its encoded counterpart xLSTMAD-R. We investigate the performance of two loss functions: Mean Squared Error (MSE), and Soft Dynamic Time Warping (SoftDTW) to consider local reconstruction fidelity and global sequence alignment, respectively. We evaluate our method on the comprehensive TSB-AD-M benchmark, which spans 17 real-world datasets, using state-of-the-art challenging metrics such as VUS-PR. In our results, xLSTM showcases state-of-the-art accuracy, outperforming 23 popular anomaly detection baselines. Our paper is the first work revealing the powerful modeling capabilities of xLSTM for anomaly detection, paving the way for exciting new developments on this subject. Our code is available at: https://github.com/Nyderx/xlstmad

Marcin Pietroń, Kamil Faber, Dominik Żurek et al.

Continual Learning (CL) is a highly relevant setting gaining traction in recent machine learning research. Among CL works, architectural and hybrid strategies are particularly effective due to their potential to adapt the model architecture as new tasks are presented. However, many existing solutions do not efficiently exploit model sparsity, and are prone to capacity saturation due to their inefficient use of available weights, which limits the number of learnable tasks. In this paper, we propose TinySubNets (TSN), a novel architectural CL strategy that addresses the issues through the unique combination of pruning with different sparsity levels, adaptive quantization, and weight sharing. Pruning identifies a subset of weights that preserve model performance, making less relevant weights available for future tasks. Adaptive quantization allows a single weight to be separated into multiple parts which can be assigned to different tasks. Weight sharing between tasks boosts the exploitation of capacity and task similarity, allowing for the identification of a better trade-off between model accuracy and capacity. These features allow TSN to efficiently leverage the available capacity, enhance knowledge transfer, and reduce computational resource consumption. Experimental results involving common benchmark CL datasets and scenarios show that our proposed strategy achieves better results in terms of accuracy than existing state-of-the-art CL strategies. Moreover, our strategy is shown to provide a significantly improved model capacity exploitation. Code released at: https://github.com/lifelonglab/tinysubnets.

Marcin Pietroń, Filip Gampel, Jakub Gomułka et al.

Argument mining (AM) is an interdisciplinary research field focused on the automatic identification and classification of argumentative components, such as claims and premises, and the relationships between them. Recent advances in large language models (LLMs) have significantly improved the performance of argument classification compared to traditional machine learning approaches. This study presents a comprehensive evaluation of several state-of-the-art LLMs, including GPT-5.2, Llama 4, and DeepSeek, on large publicly available argument classification corpora such as Args.me and UKP. The evaluation incorporates advanced prompting strategies, including Chain-of- Thought prompting, prompt rephrasing, voting, and certainty-based classification. Both quantitative performance metrics and qualitative error analysis are conducted to assess model behavior. The best-performing model in the study (GPT-5.2) achieves a classification accuracy of 78.0% (UKP) and 91.9% (Args.me). The use of prompt rephrasing, multi-prompt voting, and certainty estimation further improves classification performance and robustness. These techniques increase the accuracy and F1 metric of the models by typically a few percentage points (from 2% to 8%). However, qualitative analysis reveals systematic failure modes shared across models, including instabilities with respect to prompt formulation, difficulties in detecting implicit criticism, interpreting complex argument structures, and aligning arguments with specific claims. This work contributes the first comprehensive evaluation that combines quantitative benchmarking and qualitative error analysis on multiple argument mining datasets using advanced LLM prompting strategies.

Magdalena Stachoń, Marcin Pietroń

Many object detection models struggle with several problematic aspects of small object detection including the low number of samples, lack of diversity and low features representation. Taking into account that GANs belong to generative models class, their initial objective is to learn to mimic any data distribution. Using the proper GAN model would enable augmenting low precision data increasing their amount and diversity. This solution could potentially result in improved object detection results. Additionally, incorporating GAN-based architecture inside deep learning model can increase accuracy of small objects recognition. In this work the GAN-based method with augmentation is presented to improve small object detection on VOC Pascal dataset. The method is compared with different popular augmentation strategies like object rotations, shifts etc. The experiments are based on FasterRCNN model.

Marcin Pietron, Rafał Olszowski, Jakub Gomułka

Argument mining (AM) is defined as the task of automatically identifying and extracting argumentative components (e.g. premises, claims, etc.) and detecting the existing relations among them (i.e., support, attack, no relations). Deep learning models enable us to analyze arguments more efficiently than traditional methods and extract their semantics. This paper presents comparative studies between a few deep learning-based models in argument mining. The work concentrates on argument classification. The research was done on a wide spectrum of datasets (Args.me, UKP, US2016). The main novelty of this paper is the ensemble model which is based on BERT architecture and ChatGPT-4 as fine tuning model. The presented results show that BERT+ChatGPT-4 outperforms the rest of the models including other Transformer-based and LSTM-based models. The observed improvement is, in most cases, greater than 10The presented analysis can provide crucial insights into how the models for argument classification should be further improved. Additionally, it can help develop a prompt-based algorithm to eliminate argument classification errors.

Marcin Pietroń, Dominik Żurek, Kamil Faber et al.

Anomaly detection tools and methods enable key analytical capabilities in modern cyberphysical and sensor-based systems. Despite the fast-paced development in deep learning architectures for anomaly detection, model optimization for a given dataset is a cumbersome and time-consuming process. Neuroevolution could be an effective and efficient solution to this problem, as a fully automated search method for learning optimal neural networks, supporting both gradient and non-gradient fine tuning. However, existing frameworks incorporating neuroevolution lack of support for new layers and architectures and are typically limited to convolutional and LSTM layers. In this paper we propose AD-NEv++, a three-stage neuroevolution-based method that synergically combines subspace evolution, model evolution, and fine-tuning. Our method overcomes the limitations of existing approaches by optimizing the mutation operator in the neuroevolution process, while supporting a wide spectrum of neural layers, including attention, dense, and graph convolutional layers. Our extensive experimental evaluation was conducted with widely adopted multivariate anomaly detection benchmark datasets, and showed that the models generated by AD-NEv++ outperform well-known deep learning architectures and neuroevolution-based approaches for anomaly detection. Moreover, results show that AD-NEv++ can improve and outperform the state-of-the-art GNN (Graph Neural Networks) model architecture in all anomaly detection benchmarks.

Marcin Pietroń, Dominik Żurek, Kamil Faber et al.

Multivariate time series anomaly detection is a crucial problem in many industrial and research applications. Timely detection of anomalies allows, for instance, to prevent defects in manufacturing processes and failures in cyberphysical systems. Deep learning methods are preferred among others for their accuracy and robustness for the analysis of complex multivariate data. However, a key aspect is being able to extract predictions in a timely manner, to accommodate real-time requirements in different applications. In the case of deep learning models, model reduction is extremely important to achieve optimal results in real-time systems with limited time and memory constraints. In this paper, we address this issue by proposing a novel compression method for deep autoencoders that involves three key factors. First, pruning reduces the number of weights, while preventing catastrophic drops in accuracy by means of a fast search process that identifies high sparsity levels. Second, linear and non-linear quantization reduces model complexity by reducing the number of bits for every single weight. The combined contribution of these three aspects allow the model size to be reduced, by removing a subset of the weights (pruning), and decreasing their bit-width (quantization). As a result, the compressed model is faster and easier to adopt in highly constrained hardware environments. Experiments performed on popular multivariate anomaly detection benchmarks, show that our method is capable of achieving significant model compression ratio (between 80% and 95%) without a significant reduction in the anomaly detection performance.

Marcin Pietron, Dominik Zurek, Kamil Faber et al.

Anomaly detection tools and methods present a key capability in modern cyberphysical and failure prediction systems. Despite the fast-paced development in deep learning architectures for anomaly detection, model optimization for a given dataset is a cumbersome and time consuming process. Neuroevolution could be an effective and efficient solution to this problem, as a fully automated search method for learning optimal neural networks, supporting both gradient and non-gradient fine tuning. However, existing methods mostly focus on optimizing model architectures without taking into account feature subspaces and model weights. In this work, we propose Anomaly Detection Neuroevolution (AD-NEv) - a scalable multi-level optimized neuroevolution framework for multivariate time series anomaly detection. The method represents a novel approach to synergically: i) optimize feature subspaces for an ensemble model based on the bagging technique; ii) optimize the model architecture of single anomaly detection models; iii) perform non-gradient fine-tuning of network weights. An extensive experimental evaluation on widely adopted multivariate anomaly detection benchmark datasets shows that the models extracted by AD-NEv outperform well-known deep learning architectures for anomaly detection. Moreover, results show that AD-NEv can perform the whole process efficiently, presenting high scalability when multiple GPUs are available.

Jakub Grzeszczyk, Michał Karwatowski, Daria Łukasik et al.

This paper shows the machine learning system which performs instance segmentation of cytological images in veterinary medicine. Eleven cell types were used directly and indirectly in the experiments, including damaged and unrecognized categories. The deep learning models employed in the system achieve a high score of average precision and recall metrics, i.e. 0.94 and 0.8 respectively, for the selected three types of tumors. This variety of label types allowed us to draw a meaningful conclusion that there are relatively few mistakes for tumor cell types. Additionally, the model learned tumor cell features well enough to avoid misclassification mistakes of one tumor type into another. The experiments also revealed that the quality of the results improves with the dataset size (excluding the damaged cells). It is worth noting that all the experiments were done using a custom dedicated dataset provided by the cooperating vet doctors.

Marcin Pietroń, Dominik Żurek

This work is focused on the pruning of some convolutional neural networks (CNNs) and improving theirs efficiency on graphic processing units (GPU) by using a direct sparse algorithm. The Nvidia deep neural network (cuDnn) library is the most effective implementations of deep learning (DL) algorithms for GPUs. GPUs are the most commonly used accelerators for deep learning computations. One of the most common techniques for improving the efficiency of CNN models is weight pruning and quantization. There are two main types of pruning: structural and non-structural. The first enables much easier acceleration on many type of accelerators, but with this type it is difficult to achieve a sparsity level and accuracy as high as that obtained with the second type. Non-structural pruning with retraining can generate a weight tensors up to 90% or more of sparsity in some deep CNN models. In this article the pruning algorithm is presented which makes it possible to achieve high sparsity levels without accuracy drop. In the next stage the linear and non-linear quantization is adapted for further time and footprint reduction. This paper is an extended of previously published paper concerning effective pruning techniques and present real models pruned with high sparsities and reduced precision which can achieve better performance than the CuDnn library.

Radosław Szostak, Marcin Pietroń, Mirosław Zimnoch et al.

Development of the new methods of surface water observation is crucial in the perspective of increasingly frequent extreme hydrological events related to global warming and increasing demand for water. Orthophotos and digital surface models (DSMs) obtained using UAV photogrammetry can be used to determine the Water Surface Elevation (WSE) of a river. However, this task is difficult due to disturbances of the water surface on DSMs caused by limitations of photogrammetric algorithms. In this study, machine learning was used to extract a WSE value from disturbed photogrammetric data. A brand new dataset has been prepared specifically for this purpose by hydrology and photogrammetry experts. The new method is an important step toward automating water surface level measurements with high spatial and temporal resolution. Such data can be used to validate and calibrate of hydrological, hydraulic and hydrodynamic models making hydrological forecasts more accurate, in particular predicting extreme and dangerous events such as floods or droughts. For our knowledge this is the first approach in which dataset was created for this purpose and deep learning models were used for this task. Additionally, neuroevolution algorithm was set to explore different architectures to find local optimal models and non-gradient search was performed to fine-tune the model parameters. The achieved results have better accuracy compared to manual methods of determining WSE from photogrammetric DSMs.

M. Pietroń, D. Żurek, K. Faber et al.

Neuroevolution is one of the methodologies that can be used for learning optimal architecture during training. It uses evolutionary algorithms to generate the topology of artificial neural networks and its parameters. The main benefits are that it is scalable and can be fully or partially non gradient method. In this work, a modified neuroevolution technique is presented which incorporates multi-level optimisation. The presented approach adapts evolution strategies for evolving an ensemble model based on the bagging technique, using genetic operators for optimising single anomaly detection models, reducing the training dataset to speedup the search process and perform non-gradient fine tuning. Multivariate anomaly detection as an unsupervised learning task is the case study upon which the presented approach is tested. Single model optimisation is based on mutation and crossover operators and is focused on finding optimal window sizes, the number of layers, layer depths, hyperparameters etc. to boost the anomaly detection scores of new and already known models. The proposed framework and its protocol shows that it is possible to find architecture within a reasonable time frame which can boost all well known multivariate anomaly detection deep learning architectures. The work concentrates on improvements to the multi-level neuroevolution approach for anomaly detection. The main modifications are in the methods of mixing groups and single model evolution, non-gradient fine tuning and a voting mechanism. The presented framework can be used as an efficient learning network architecture method for any different unsupervised task where autoencoder architectures can be used. The tests were run on SWAT, WADI, MSL and SMAP datasets and the presented approach evolved the architectures that achieved the best scores among other deep learning models.

Marcin Pietroń, Dominik Żurek

This paper is focused on the improvement the efficiency of the sparse convolutional neural networks (CNNs) layers on graphic processing units (GPU). The Nvidia deep neural network (cuDnn) library provides the most effective implementation of deep learning (DL) algorithms for GPUs. GPUs are one of the most efficient and commonly used accelerators for deep learning computations. The modern CNN models need megabytes of coefficients and needed millions MAC operations to perform convolution. One of the most common techniques for compressing CNN models is weight pruning. There are two main types of pruning: structural (based on removing whole weight channels) and non-structural (removing individual weights). The first enables much easier acceleration, but with this type it is difficult to achieve a sparsity level and accuracy as high as that obtained with the second type. Non-structural pruning with retraining can generate a matrix-weight up to $\sim90\%$ or more of sparsity in some deep CNN models. This work shows when is worth using a direct sparse operation to speed-up the calculation of the convolution layers. The VGG-16, CNN-non-static and 1x1 layers from ResNet models were used as a benchmarks. In addition, we present the impact of using reduced precision on time efficiency.

Dominik Żurek, Marcin Pietroń

This paper presents the impact of using quantization on the efficiency of multi-class text classification in the training process of a support vector machine (SVM). This work is focused on comparing the efficiency of SVM model trained using reduced precision with its original form. The main advantage of using quantization is decrease in computation time and in memory footprint on the dedicated hardware platform which supports low precision computation like GPU (16-bit) or FPGA (any bit-width). The paper presents the impact of a precision reduction of the SVM training process on text classification accuracy. The implementation of the CPU was performed using the OpenMP library. Additionally, the results of the implementation of the GPU using double, single and half precision are presented.

Marcin Pietron, Maciej Wielgosz

Convolutional neural networks (CNN) play a major role in image processing tasks like image classification, object detection, semantic segmentation. Very often CNN networks have from several to hundred stacked layers with several megabytes of weights. One of the possible methods to reduce complexity and memory footprint is pruning. Pruning is a process of removing weights which connect neurons from two adjacent layers in the network. The process of finding near optimal solution with specified drop in accuracy can be more sophisticated when DL model has higher number of convolutional layers. In the paper few approaches based on retraining and no retraining are described and compared together.

Anna Aniol, Marcin Pietron

Machine comprehension, answering a question depending on a given context paragraph is a typical task of Natural Language Understanding. It requires to model complex dependencies existing between the question and the context paragraph. There are many neural network models attempting to solve the problem of question answering. The best models have been selected, studied and compared with each other. All the selected models are based on the neural attention mechanism concept. Additionally, studies on a SQUAD dataset were performed. The subsets of queries were extracted and then each model was analyzed how it deals with specific group of queries. Based on these three model ensemble model was created and tested on SQUAD dataset. It outperforms the best Mnemonic Reader model.

Mo'taz Al-Hami, Marcin Pietron, Rishi Kumar et al.

This invention addresses fixed-point representations of convolutional neural networks (CNN) in integrated circuits. When quantizing a CNN for a practical implementation there is a trade-off between the precision used for operations between coefficients and data and the accuracy of the system. A homogenous representation may not be sufficient to achieve the best level of performance at a reasonable cost in implementation complexity or power consumption. Parsimonious ways of representing data and coefficients are needed to improve power efficiency and throughput while maintaining accuracy of a CNN.

Krzysztof Wróbel, Marcin Pietroń, Maciej Wielgosz et al.

Convolutional neural networks are modern models that are very efficient in many classification tasks. They were originally created for image processing purposes. Then some trials were performed to use them in different domains like natural language processing. The artificial intelligence systems (like humanoid robots) are very often based on embedded systems with constraints on memory, power consumption etc. Therefore convolutional neural network because of its memory capacity should be reduced to be mapped to given hardware. In this paper, results are presented of compressing the efficient convolutional neural networks for sentiment analysis. The main steps are quantization and pruning processes. The method responsible for mapping compressed network to FPGA and results of this implementation are presented. The described simulations showed that 5-bit width is enough to have no drop in accuracy from floating point version of the network. Additionally, significant memory footprint reduction was achieved (from 85% up to 93%).

Krzysztof Wróbel, Maciej Wielgosz, Marcin Pietroń et al.

This paper presents the analysis of the impact of a floating-point number precision reduction on the quality of text classification. The precision reduction of the vectors representing the data (e.g. TF-IDF representation in our case) allows for a decrease of computing time and memory footprint on dedicated hardware platforms. The impact of precision reduction on the classification quality was performed on 5 corpora, using 4 different classifiers. Also, dimensionality reduction was taken into account. Results indicate that the precision reduction improves classification accuracy for most cases (up to 25% of error reduction). In general, the reduction from 64 to 4 bits gives the best scores and ensures that the results will not be worse than with the full floating-point representation.

Maciej Wielgosz, Marcin Pietroń

This paper examines the performance of a Spatial Pooler (SP) of a Hierarchical Temporal Memory (HTM) in the task of noisy object recognition. To address this challenge, a dedicated custom-designed system based on the SP, histogram calculation module and SVM classifier was implemented. In addition to implementing their own version of HTM, the authors also designed a profiler which is capable of tracing all of the key parameters of the system. This was necessary, since an analysis and monitoring of the system performance turned out to be extremely difficult using conventional testing and debugging tools. The system was initially trained on artificially prepared videos without noise and then tested with a set of noisy video streams. This approach was intended to mimic a real life scenario where an agent or a system trained to deal with ideal objects faces a task of classifying distorted and noisy ones in its regular working conditions. The authors conducted a series of experiments for various macro parameters of HTM SP, as well as for different levels of video reduction ratios. The experiments allowed them to evaluate the performance of two different system setups (i.e. 'Multiple HTMs' and 'Single HTM') under various noise conditions with 32--frame video files. Results of all the tests were compared to SVM baseline setup. It was determined that the system featuring SP is capable of achieving approximately 12 times the noise reduction for a video signal with with distorted bits accounting for 13\% of the total. Furthermore, the system featuring SP performed better also in the experiments without a noise component and achieved a max F1 score of 0.96. The experiments also revealed that a rise of column and synapse number of SP has a substantial impact on the performance of the system. Consequently, the highest F1 score values were obtained for 256 and 4096 synapses and columns respectively.

Maciej Wielgosz, Marcin Pietroń

We present a method to classify objects in video streams using a brain-inspired Hierarchical Temporal Memory (HTM) algorithm. Object classification is a challenging task where humans still significantly outperform machine learning algorithms due to their unique capabilities. We have implemented a system which achieves very promising performance in terms of recognition accuracy. Unfortunately, conducting more advanced experiments is very computationally demanding; some of the trials run on a standard CPU may take as long as several days for 960x540 video streams frames. Therefore we have decided to accelerate selected parts of the system using OpenCL. In particular, we seek to determine to what extent porting selected and computationally demanding parts of a core may speed up calculations. The classification accuracy of the system was examined through a series of experiments and the performance was given in terms of F1 score as a function of the number of columns, synapses, $min\_overlap$ and $winners\_set\_size$. The system achieves the highest F1 score of 0.95 and 0.91 for $min\_overlap=4$ and 256 synapses, respectively. We have also conduced a series of experiments with different hardware setups and measured CPU/GPU acceleration. The best kernel speed-up of 632x and 207x was reached for 256 synapses and 1024 columns. However, overall acceleration including transfer time was significantly lower and amounted to 6.5x and 3.2x for the same setup.

M. Pietron, M. Wielgosz, K. Wiatr

This paper introduces mathematical formalism for Spatial (SP) of Hierarchical Temporal Memory (HTM) with a spacial consideration for its hardware implementation. Performance of HTM network and its ability to learn and adjust to a problem at hand is governed by a large set of parameters. Most of parameters are codependent which makes creating efficient HTM-based solutions challenging. It requires profound knowledge of the settings and their impact on the performance of system. Consequently, this paper introduced a set of formulas which are to facilitate the design process by enhancing tedious trial-and-error method with a tool for choosing initial parameters which enable quick learning convergence. This is especially important in hardware implementations which are constrained by the limited resources of a platform. The authors focused especially on a formalism of Spatial Pooler and derive at the formulas for quality and convergence of the model. This may be considered as recipes for designing efficient HTM models for given input patterns.