The accurate classification of electrocardiogram (ECG) signals is crucial for a reliable diagnosis of cardiac arrhythmias such as ventricular tachycardia (VT) or heart block. However, ECG signals often contain variability and noise, challenging the robustness of deep neural networks (DNNs) and compromising classification accuracy. This study investigates the effectiveness of the recently proposed FlexPoints algorithm, an innovative method for detecting characteristic points within ECG signals, in improving DNN performance, including convolutional neural networks (CNNs) and long short-term memory neural networks (LSTMs) in ECG signal classification tasks. FlexPoints strategically identify characteristic points that reflect the under-lying electrophysiological dynamics, facilitating enhanced feature extraction by DNN architectures. Comparative experiments conducted on benchmark ECG datasets demonstrate that integrating FlexPoints significantly increases the classification accuracy for ECG signals. The results indicate improved robustness to noise and reduction in computational complexity, attributed to efficient representation and processing of signals. This research underscores the utility of the FlexPoints algorithm in strengthening DNN-based classification approaches, paving the way for a more accurate and clinically reliable interpretation of ECG signals.

This paper introduces a novel pruning method designed for transfer-learning models in computer vision, leveraging XGBoost to enhance model efficiency through simultaneous depth and width pruning. Contemporary computer vision tasks often rely on transfer learning due to its efficiency after short training periods. However, these pre-trained architectures, while robust, are typically overparameterized, leading to inefficiencies when fine-tuning for simpler problems. Our approach utilizes Parallel Tree Boosting as a surrogate neural classifier to evaluate and prune unnecessary convolutional filters in these architectures effectively. This method addresses the challenge of interference that is often present in transfer-learning models and enables a significant reduction in the size of the model. We demonstrate that it is feasible to achieve both depth and width pruning concurrently, allowing for substantial reductions in model complexity without compromising performance. Experimental results show that our proposed XGB-based Pruner not only reduces the model size by nearly 50% but also improves the accuracy by up to 9% in test data compared to conventional models. Moreover, it can achieve a dramatic reduction in model size by over a hundred times with minimal accuracy loss, proving its effectiveness across various architectures and datasets. This approach marks a significant advancement in optimizing transfer-learning models, promising enhanced performance in real-world applications.

This paper introduces a motivated agent scheme that enables an agent to create its own goals using prior knowledge about its environment. A motivated agent operates in a dynamically changing environment and is capable of setting and achieving its own goals, as well as those set by the designer. The agent has access to additional knowledge about the environment, which is represented in associative semantic memory. This memory is constructed based on ANAKG associative knowledge graphs, which have been shown to have several advantages over other semantic memories for processing symbolic sequential inputs. They are easy to organize and train. In this paper, we demonstrate that a motivated agent with semantic memory learns to achieve its goals more easily and utilizes environmental resources more effectively. To simplify comparisons with reinforcement learning, we represent the environment as an environmental graph that shows the principles governing it. By exploring the environment, the agent learns these principles and can use them to accomplish its tasks. Our experiments and tests confirm our claims about the higher efficiency of agents with memory. Moreover, an extensive comparison with reinforcement learning agents highlights the advantages of motivated learning over reinforcement learning.

In this work, we focus on hyperparameter optimization, which is one of the essential steps in the development of machine learning solutions. This work proposes a novel approach to hyperparameter optimization using associative self-adapting structures, which allows us to efficiently explore the hyperparameter space. The core algorithm introduced in this paper is based on the associative graph data structure (AGDS) merged with the evolutionary approach, inspired by the processes used in drug discovery and human behavior. The proposed algorithm was compared with two state-of-the-art algorithms, a Tree-Structured Parzen Estimator and differential evolution. All experiments were carried out using Penn Machine Learning Benchmarks, repeated ten times for each problem, to provide objective comparisons. We optimized Random Forest and deep neural network models for every dataset of these benchmarks. The results reveal a noticeable improvement compared to both optimization methods in both optimized models, which shows that the presented approach can be successfully used as a good alternative for current state-of-the-art solutions.

This article presents concurrent associative memories with synaptic delays useful for processing sequences of real vectors. Associative memories with synaptic delays were introduced by the authors for symbolic sequential inputs and demonstrated several advantages over other sequential memories. They were easy to organize and train. It was demonstrated that they were more robust than long short-term memories in recognition of damaged sequences. The associative memories can be applied in combination with deep neural networks to solve such symbol grounding problems, such as speech recognition, and support sequential memories triggered by sensory inputs. Several practical considerations for developed memories were discussed and illustrated. A continuous speech database was used to compare the developed method with LSTM memories. Tests demonstrated that the developed approach is more robust in recognition of speech sequences, particularly when the test sequences are damaged.

One of the most popular and widely used object detection algorithm today is the YOLOv3 due to its high performance and speed. However, YOLOv3 is not the best algorithm in terms of precision. This paper introduces a substantial change to the post-processing routine of the YOLOv3 after the prediction to increase its final accuracy. Currently, YOLOv3 uses a Non-Max Suppression algorithm to eliminate multiple detections of the same object. This algorithm is picking the most confident overlaying box on any object to present it as the final prediction. This paper presents a new algorithm called Weighted Centers of Confidence Selection that increases the precision using a confidence-weighted average bounding box as a replacement to the existing bounding boxes without making any changes to the YOLOv3 convolutional neural network. We demonstrate how this algorithm works and compare its results to the results achieved by the YOLO's Non-Max Suppression algorithm, focusing on precision and achieving almost the same frame-speed as the original YOLOv3. This new approach allowed us to improve the average accuracy on the COCO dataset in comparison to the original YOLO's Non-Max Suppression.

In this paper, we introduce a new concept of associative memories in which synaptic connections of the self-organizing neural network learn time delays between input sequence elements. Synaptic connections represent both the synaptic weights and expected delays between the network inputs. This property of synaptic connections facilitates recognition of time sequences and provides context-based associations between sequence elements. Characteristics of time delays are learned and are updated each time an input sequence is presented. There are no separate learning and testing modes typically used in other neural networks, as the network starts to predict the next input element as soon as there is no expected input signal. The network generates output signals useful for associative recall and prediction. These output signals depend on the presented input context and the knowledge stored in the graph. Such a mode of operation is preferred for the organization of episodic memories used to store the observed episodes and to recall them if a sufficient context is provided. The associative sequential recall is useful for the operation of working memory in a cognitive agent. Test results demonstrate that the network correctly recognizes the input sequences with variable delays and that it is more efficient than other recently developed sequential memory networks based on associative neurons.

A generalization of active neural associative knowledge graphs (ANAKGs) to their minicolumn form is presented in this paper. Each minicolumn represents a single symbol, and the activation of an individual neuron in a minicolumn depends on the context of the activation of the presynaptic neuron. The implemented memory model combines the ANAKG associative spiking neuron idea with the idea of the hierarchical temporal memory. This new associative memory organization preserves all properties of ANAKG memories, such as storage of knowledge based on the association of spatiotemporal input sequences, self-organization, quick learning, and recall of the sequential memories, while increasing the recall quality and the memory capacity. The recall quality advantage of the new approach over ANAKG increases with the length of the recalled episodes and the number of neurons used in each minicolumn. We introduced a new distance measure to compare the recalled sequences and defined a recall quality to determine the memory capacity. Performed tests confirmed our claims. Additional tests were performed to illustrate the computational complexity and the efficiency of the developed approach.

This paper introduces a new efficient associative graph data structure with efficient access to all stored data inspired by biological neural networks and associative brain-like approaches. It also introduces new self-balancing, self-organizing, and self-sorting AVB+trees which not only store data but it also allows for replacing many search operations with this structure. This structure aggregates representations of all duplicated of values and objects. It also allows for faster computation of many useful functions as medians, average, minima, maxima, neighbor values in the defined order etc. It always stores all data in order for all attributes simultaneously. Automatic aggregations of duplicates allow it for access to all stored data in less than logarithmic time. These associative structures can also be successfully used as a kind of artificial neural network for classification, clustering, various inferences, estimation of similarities, differences, or correlations of the stored objects.

This paper introduces the use of a new model of associative pulsing neurons (APN) for multi-class and multi-label classification tasks which are usually performed by separate artificial neural networks. The presented associative pulsing neurons have similar capabilities as various spiking models of neurons, but they additionally have built-in conditional plastic mechanisms which allow creating a neural structure with any given training dataset. Associative pulsing neurons can be connected and adapted very quickly. They have been implemented in the described research to define static patterns and their relations. They have been successfully used to automatically construct associative pulsing neural networks (APNN) to provide a classification of some well-known benchmark training data. These networks use special receptors which transform external stimuli into their internal representation of pulses. Receptors charge connected neurons in different periods of time according to the similarity of the presented input value to the values that they represent. This paper also presents the answer to one of the most challenging tasks in neuroscience, i.e. whether neurons communicate by a rate of pulses or temporal differences between pulses, and how the frequency of pulses influences the neural network activations.

This paper describes the integration of semantic and episodic memory (EM) models and the benefits of such integration. Semantic memory (SM) is used as a foundation of knowledge and concept learning, and is needed for the operation of any cognitive system. EM retains personal experiences stored based on their significance-it is supported by the SM, and in return, it supports SM operations. Integrated declarative memories are critical for cognitive system development, yet very little research has been done to develop their computational models. We considered structural self-organization of both semantic and episodic memories with a symbolic representation of input events. Sequences of events are stored in EM and are used to build associations in SM. We demonstrated that integration of semantic and episodic memories improves the native operation of both types of memories. Experimental results are presented to illustrate how the two memories complement each other by improving recognition, prediction, and context-based generalization of individual memories.

Self-optimizing neural networks (SONNs) are very effective in solving different classification tasks. They have been successfully used to many different problems. The classical SONN adaptation process has been defined as supervised. This paper introduces a new very interesting SONN feature - the unsupervised clustering ability. The unsupervised SONNs (US-SONNs) are able to find out most differentiating features for some training data and recursively divide them into subgroups. US-SONNs can also characterize the importance of features differentiating these groups. The division of the data is recursively performed till the data in subgroups differ imperceptibly. The SONN clustering proceeds very fast in comparison to other unsupervised clustering methods.

Artificial neural networks (ANNs) were inspired by natural neural networks (NNNs) and natural processes of training. The NNNs receive data in time still tuning the inner model of the surrounding world. These valuable features of our brains let us to dynamically accommodate themselves to the changes surround. These features make us possible to forget some irrelevant information, correct our knowledge and meet truth. ANNs usually work on the training data (TD) acquired in the past and totally known at the beginning of the adaptation process. Because of this the adaptation methods of the ANNs can be sometimes more effective than the natural training process observed in the NNNs. This paper discusses the ability of ANNs to adapt more effectively than NNNs do if only the TD is completely given at the beginning of the adaptation process. In this case the adaptation process of ANNs can be divided into two steps: analyze or examining the set of TD and construction of neural network topology and weights computation. Two different applications areas of such approach are presented in the paper.