Optical Memory and Neural Networks最新文献

The task of developing and applying neuromorphic elements of an information control system for mobile robots is considered. The description of the compartmental spiking neuron model used in the work and the algorithm of its structural learning is given. The elements of the information control system used in the work are described: a neuromorphic emergency detector, a neuromorphic extrapolator, and a neuromorphic model for the formation of associative connections. Based on these elements, a scheme for the formation of a conditioned reflex reaction with negative reinforcement is proposed. In addition, a scheme is considered that allows a mobile robot to move at a given distance from the wall. The first of these schemes was tested on a real mobile robotics platform. The conclusion is made about the possibility of constructing neuromorphic information control systems from the presented elements and the prospects for the development of this approach.

The task of aircraft motion control has to be solved under conditions of numerous heterogeneous uncertainties both in the aircraft motion model and in the environment in which the aircraft is flying. These uncertainties, in particular, are caused by the fact that in the flight of the aircraft can occur various kinds of abnormal situations caused by failures of equipment and systems of the aircraft, damage to the airframe and propulsion system of the aircraft. Some of these failures and damages have a direct impact on the dynamic characteristics of the aircraft as a control object. In this regard, the problem arises of such an adjustment of aircraft control algorithms that would provide the ability to adapt to the changed dynamics of the aircraft. It is extremely difficult, and in some cases impossible, to foresee in advance all possible damages, failures and their combinations. Hence, it is necessary to implement adaptive flight control algorithms that are able to adjust to the changing situation. One of the effective tools for solving such problems is reinforcement learning in the Approximate Dynamic Programming (ADP) variant, in combination with artificial neural networks. In the last decade, a family of methods known as Adaptive Critic Design (ACD) has been actively developed within the ADP approach to control the behavior of complex dynamic systems. In our paper we consider the application of one of the variants of the ACD approach, namely SNAC (Single Network Adaptive Critic) and its development through its joint use with the method of dynamic inversion. The effectiveness of this approach is demonstrated on the example of longitudinal motion control of a supersonic transport airplane.

Large Language Models fine-tuning is an essential part of modern artificial intelligent systems that solve numerous tasks, such as natural language processing and computer vision. Among the various fine-tuning strategies, the most prominent approach for Large Language Model fine-tuning is Parameter-Efficient Fine-Tuning (PEFT), as it allows to achieve state-of-the-art performance on multiple tasks while minimizing computational resources and training time. Recently, an increasing number of PEFT methodologies have been developed, each asserting superiority based on performance metrics. However, a critical evaluation of how these methods align with the tuning dynamic of the full fine-tuning (FT) remains largely unexplored. This study focuses on bridging this gap by analyzing the learning behavior of such PEFT approaches as LoRA, LoRA+, AdaLoRA, DoRA, VeRA, PiSSA, LoKr and LoHa in comparison to FT. This work provides a comprehensive comparative analysis aimed at identifying which PEFT methods diverge significantly in weights update dynamic from the FT standard. The findings reveal insights into the underlying causes of these discrepancies, offering a deeper understanding of each method’s behavior and efficacy.

Kolmogorov-Arnold Networks (KAN), introduced in May 2024, are a novel type of artificial neural networks, whose abilities and properties are now being actively investigated by the machine learning community. In this study, we test application of KAN to solve an inverse problem for development of multimodal carbon luminescent nanosensors of ions dissolved in water, including heavy metal cations. We compare the results of solving this problem with four various machine learning methods—random forest, gradient boosting over decision trees, multi-layer perceptron neural networks, and KAN. Advantages and disadvantages of KAN are discussed, and it is demonstrated that KAN has high chance to become one of the algorithms most recommended for use in solving highly non-linear regression problems with moderate number of input features.

Recent advancements have significantly improved the skills and performance of language models, but have also increased computational demands due to the increasing number of parameters and the quadratic complexity of the attention mechanism. As context sizes expand into millions of tokens, making long-context processing more accessible and efficient becomes a critical challenge. Furthermore, modern benchmarks such as BABILong [1] underscore the inefficiency of even the most powerful LLMs in long context reasoning. In this paper, we employ finetuning and multi-task learning to train a model capable of mastering multiple BABILong long-context reasoning skills. We demonstrate that even models with fewer than 140 million parameters can outperform much larger counterparts by learning multiple essential tasks simultaneously. By conditioning Recurrent Memory Transformer [2] on task description, we achieve state-of-the-art results on multi-task BABILong QA1–QA5 set for up to 32k tokens. The proposed model also shows generalization abilities to new lengths and tasks, along with increased robustness to input perturbations.

Semantic-aware mapping is crucial for advancing robotic navigation and interaction within complex environments. Traditional 3D mapping techniques primarily capture geometric details, missing the semantic richness necessary for autonomous systems to understand their surroundings comprehensively. This paper presents Sea-SHINE, a novel approach that integrates semantic information within a neural implicit mapping framework for large-scale environments. Our method enhances the utility and navigational relevance of 3D maps by embedding semantic awareness into the mapping process, allowing robots to recognize, understand, and reconstruct environments effectively. The proposed system leverages dual decoders and a semantic awareness module, which utilizes Feature-wise Linear Modulation (FiLM) to condition mapping on semantic labels. Extensive experiments on datasets such as SemanticKITTI, KITTI-360, and ITLP-Campus demonstrate significant improvements in map precision and recall, particularly in recognizing crucial objects like road signs. Our implementation bridges the gap between geometric accuracy and semantic understanding, fostering a deeper interaction between robots and their operational environments. The code is publicly available at https://github.com/VitalyyBezuglyj/Sea-SHINE.

This study explores the development of classifiers for microbial images, specifically focusing on streptococci captured via microscopy of live samples. Our approach uses AutoML-based techniques and automates the creation and analysis of feature spaces to produce optimal descriptors for classifying these microscopic images. This technique leverages interpretable taxonomic features based on the external geometric attributes of various microorganisms. We have released an annotated dataset we assembled to validate our solution, featuring microbial images from unfixed microscopic scenes. Additionally, we assessed the classification performance of our method against several classifiers, including those employing deep neural networks. Our approach outperformed all others tested, achieving the highest Precision (0.980), Recall (0.979), and F1-score (0.980).

The work is devoted to the development of a classifier of motor activity patterns based on electroencephalograms (EEG) for a real-time brain-computer interface (BCI), which can be used in contactless control systems. Conducted studies of various methods for classifying motor EEG images have shown that their effectiveness significantly depends on the implementation of the stages of information processing in the BCI. The most effective classification method turned out to be the support vector machine. However, its long operating time and lack of accuracy make it difficult to use for implementing real-time BCI. Therefore, a classifier was developed using an ensemble of detectors, each of which is trained to recognize its own motor EEG image. A new EEG analysis algorithm based on event functions was applied. A study of the classifier showed that it is possible to achieve detection accuracy of 98.5% with an interface delay of 230 ms.

The paper considers the effect of transcendental factors on behavior of artificial beings, agents and robots. The boundary between the rational and transcendental relies on the type of an individual or more specifically, his sensory and intellectual abilities. For agents and simple robots all, except for tangible elements of the environment and manipulations with them, is transcendental. Such transcendental factors as environmental changes and algorithm modifications determined by the programmer, supervisor, or operator have significant effects on agents and communities of agents. Hardware malfunctions (transcendental events from an agent’s point of view) can be crucial for agents. Agents can take advantages from transcendental effects if the programmer realizes a feedback. Generation of a mental copy of an agent for making new agents allows continuity and social development. For intellectual robots the boundaries of the transcendental move away because of their ability to accommodate to new environment. However, in most cases the role of the transcendental even increases with the improvement of robots because there are consciousness and growth of communication possibilities. The consciousness changes as a result of learning of transcendental information, making robots change the behavior. Robot’s communication abilities enable transcendental (along with rational) information to be received from the data base in any amount. For people living together with intelligent robots, this sort of communication can become a tool for introducing human culture in the community of robots. This in turn would result in humanization of robots and establishment of good relations between robots and human beings.

This research develops improved Robust Adversarial Models (RAM) to enhance Intrusion Detection Systems’ (IDS) robustness against evasion attacks. Malicious packets crafted using Scapy were infused into open-source datasets NSL-KDD and CICIDS obtained from Kaggle. Experiments involved passing this traffic through baseline IDS model such as in a free open-source IDS Snort and the improved RAM. Training processes employed perturbations using Generative Adversarial Networks (GAN), Fast Gradient Sign Methods (FGSM), and Projected Gradient Descent (PGD) against reinforcement learning of features and labels from the autoencoder model. The robust adversarial model showed 34.52% higher accuracy, 59.06% higher F1-score and 85.26% higher recall than the baseline IDS Snort model across datasets. Comparative analysis demonstrated the improved RAM’s enhanced resilience, performance, and reliability in real-world scenarios, advancing IDS models' and network infrastructures' security posture.