International journal of neural systems最新文献

Diffusion models have achieved remarkable success in image generation, image super-resolution, and text-to-image synthesis. Despite their effectiveness, they face key challenges, notably long inference time and complex architectures that incur high computational costs. While various methods have been proposed to reduce inference steps and accelerate computation, the optimization of diffusion model architectures has received comparatively limited attention. To address this gap, we propose LDMOES (Lightweight Diffusion Models based on Multi-Objective Evolutionary Search), a framework that combines multi-objective evolutionary neural architecture search with knowledge distillation to design efficient UNet-based diffusion models. By adopting a modular search space, LDMOES effectively decouples architecture components for improved search efficiency. We validated our method on multiple datasets, including CIFAR-10, Tiny-ImageNet, CelebA-HQ [Formula: see text], and LSUN-church [Formula: see text]. Experiments show that LDMOES reduces multiply-accumulate operations (MACs) by approximately 40% in pixel space while outperforming the teacher model. When transferred to the larger-scale Tiny-ImageNet dataset, it still generates high-quality images with a competitive FID score of 4.16, demonstrating strong generalization ability. In latent space, MACs are reduced by about 50% with negligible performance loss. After transferring to the more complex LSUN-church dataset, the model surpasses baselines in generation quality while reducing computational cost by nearly 60%, validating the effectiveness and transferability of the multi-objective search strategy. Code and models will be available at https://github.com/GenerativeMind-arch/LDMOES.

In a world where social interaction presents challenges for children with Autism Spectrum Disorder (ASD), robots are stepping in as allies in emotional learning. This study examined how affective interactions with a humanoid robot elicited physiological responses in children with ASD, using electrodermal activity (EDA) and heart rate variability (HRV) as key indicators of emotional arousal. The objectives were to identify emotionally salient moments during human-robot interaction, assess whether certain individual characteristics - such as age or ASD severity - modulate autonomic responses, and evaluate the usefulness of wearable devices for real-time monitoring. Thirteen children participated in structured sessions involving a range of social, cognitive, and motor tasks alongside the robot Pepper. The results showed that the hugging phase (HS2) often generated greater autonomic reactivity in children, especially among younger children and those with higher levels of restlessness or a higher level of ASD. Children with level 2 ASD displayed higher sympathetic activation compared to level 1 participants, who showed more HRV stability. Age also played a role, as younger children demonstrated lower autonomic regulation. These findings highlight the relevance of physiological monitoring in detecting emotional dysregulation and tailoring robot-assisted therapy. Future developments will explore adaptive systems capable of adjusting interventions in real time to better support each child's unique needs.

Decoding electroencephalographic (EEG) signals is of key importance in the development of brain-computer interface (BCI) systems. However, high inter-subject variability in EEG signals requires user-specific calibration, which can be time-consuming and limit the application of deep learning approaches, due to general need of large amount of data to properly train these models. In this context, this paper proposes a multidimensional and explainable deep learning framework for fast and interpretable EEG decoding. In particular, EEG signals are projected into the spatial-spectral-temporal domain and processed using a custom three-dimensional (3D) Convolutional Neural Network, here referred to as EEGCubeNet. In this work, the method has been validated on EEGs recorded during motor BCI experiments. Namely, hand open (HO) and hand close (HC) movement planning was investigated by discriminating them from the absence of movement preparation (resting state, RE). The proposed method is based on a global- to subject-specific fine-tuning. The model is globally trained on a population of subjects and then fine-tuned on the final user, significantly reducing adaptation time. Experimental results demonstrate that EEGCubeNet achieves state-of-the-art performance (accuracy of [Formula: see text] and [Formula: see text] for HC versus RE and HO versus RE, binary classification tasks, respectively) with reduced framework complexity and with a reduced training time. In addition, to enhance transparency, a 3D occlusion sensitivity analysis-based explainability method (here named 3D xAI-OSA) that generates relevance maps revealing the most significant features to each prediction, was introduced. The data and source code are available at the following link: https://github.com/AI-Lab-UniRC/EEGCubeNet.

Deficits in auditory perception have been widely observed in Parkinson's disease (PD) patients and the literature attributes it, in part, to impaired central auditory processing. Deep brain stimulation (DBS) of the subthalamic nucleus (STN) is a well-established therapeutic option for patients with advanced PD. Analysis of auditory evoked potentials suggested a modulatory effect of DBS on central auditory processing. To better understand the latter, we investigated whether DBS modulates auditory steady state responses (ASSR) in electroencephalography (EEG) recordings of 5 PD patients. ASSRs are neural responses along central auditory pathways phase-locked to an auditory stimulus which can serve to understand the spectral aspects of central auditory processing. In our analyses, we estimate the intensity of ASSRs with a novel method based on canonical correlation analysis (CCA) and compare them in DBS ON and OFF conditions. Our results suggest that DBS effectively reduces ASSR in patients with PD. A comparison to age-matched healthy participants suggests a pathological effect of PD on ASSRs, which is disrupted by DBS. These findings support our hypothesis that DBS suppresses central auditory processing. Further research is required to assess the symptomatic effect of this modulation, as well as which cortical and subcortical generators are most affected. A better understanding of the auditory side-effects of DBS could lead to improved treatment options.

In the context of neural system structure modeling and complex visual tasks, the effective integration of multi-scale features and contextual information is critical for enhancing model performance. This paper proposes a biologically inspired hybrid neural network architecture - CompEyeNet - which combines the global modeling capacity of transformers with the efficiency of lightweight convolutional structures. The backbone network, multi-attention transformer backbone network (MATBN), integrates multiple attention mechanisms to collaboratively model local details and long-range dependencies. The neck network, compound eye neck network (CENN), introduces high-resolution feature layers and efficient attention fusion modules to significantly enhance multi-scale information representation and reconstruction capability. CompEyeNet is evaluated on three authoritative medical image segmentation datasets: MICCAI-CVC-ClinicDB, ISIC2018, and MICCAI-tooth-segmentation, demonstrating its superior performance. Experimental results show that compared to models such as Deeplab, Unet, and the YOLO series, CompEyeNet achieves better performance with fewer parameters. Specifically, compared to the baseline model YOLOv11, CompEyeNet reduces the number of parameters by an average of 38.31%. On key performance metrics, the average Dice coefficient improves by 0.87%, the Jaccard index by 1.53%, Precision by 0.58%, and Recall by 1.11%. These findings verify the advantages of the proposed architecture in terms of parameter efficiency and accuracy, highlighting the broad application potential of bio-inspired attention-fusion hybrid neural networks in neural system modeling and image analysis.

The real-time synthesis of 3D views, facilitated by convolutional neural networks like NeX, is increasingly pivotal in various computer vision applications. These networks are trained using photographs taken from different perspectives during the training phase. However, these images may be susceptible to contamination from noise originating from the vision sensor or the surrounding environment. This research meticulously examines the impact of noise on the resulting image quality of 3D views synthesized by the NeX network. Various noise levels and scenes have been incorporated to substantiate the claim that the presence of noise significantly degrades image quality. Additionally, a new strategy is introduced to improve image quality by calculating consensus among NeX networks trained on images pre-processed with a denoising algorithm. Experimental results confirm the effectiveness of this technique, demonstrating improvements of up to 1.300 dB and 0.032 for Peak Signal Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM), respectively, under certain scenes and noise levels. Notably, the performance gains are especially significant when using synthesized images generated by NeX from noisy inputs in the consensus process.

In this paper, we present a hybrid learning framework that integrates two model-driven AI paradigms: Deep unfolding and Variable Projections (VPs). The core idea is to unfold the iterations of VP solvers for separable nonlinear least squares (SNLLS) problems into trainable neural network layers. As a consequence, the network is capable of learning optimal nonlinear VP parameters during inference, which is a form of model-based meta-learning. Furthermore, the architecture incorporates prior knowledge of the underlying SNLLS problem, such as basis function expansions and signal structure, which enhance interpretability, reduce model size, and lower data requirements. As a case study, we adapt the proposed deep unfolded VPNet to learn ECG representations for the classification of five arrhythmias. Experimental results on the MIT-BIH Arrhythmia Database show that VPNet achieves performance comparable to state-of-the-art ECG classifiers, attaining  95% accuracy while maintaining a compact architecture. Its low computational complexity enables efficient training and inference, making it highly suitable for real-time, power-efficient edge computing applications. This is further validated through embedded implementation on STM32 microcontrollers.

Cellular traffic forecasting is a critical task that enables network operators to efficiently allocate resources and address anomalies in rapidly evolving environments. The exponential growth of data collected from base stations poses significant challenges to processing and analysis. While machine learning (ML) algorithms have emerged as powerful tools for handling these large datasets and providing accurate predictions, their environmental impact, particularly in terms of energy consumption, is often overlooked in favor of their predictive capabilities. This study investigates the potential of two bio-inspired models: Spiking Neural Networks (SNNs) and Reservoir Computing through Echo State Networks (ESNs) for cellular traffic forecasting. The evaluation focuses on both their predictive performance and energy efficiency. These models are implemented in both centralized and federated settings to analyze their effectiveness and energy consumption in decentralized systems. Additionally, we compare bio-inspired models with traditional architectures, such as Convolutional Neural Networks (CNNs) and Multi-Layer Perceptrons (MLPs), to provide a comprehensive evaluation. Using data collected from three diverse locations in Barcelona, Spain, we examine the trade-offs between predictive accuracy and energy demands across these approaches. The results indicate that bio-inspired models, such as SNNs and ESNs, can achieve significant energy savings while maintaining predictive accuracy comparable to traditional architectures. Furthermore, federated implementations were tested to evaluate their energy efficiency in decentralized settings compared to centralized systems, particularly in combination with bio-inspired models. These findings offer valuable insights into the potential of bio-inspired models for sustainable and privacy-preserving cellular traffic forecasting.

This research looks at the use of long-short-term memory (LSTM) networks to predict psychosis, in patients within the schizophrenia spectrum, based on Heart Rate Variability (HRV) data acquired from wearable devices. Our primary objective is to test whether the personalized relapse prediction remains accurate when eliminating the artifact-removal preprocessing. We first analyzed 7 patients sleep HRV recordings (7-113 days each), and then validated the methodology on a separate 30-patient psychosis cohort from another clinical setting. In this framework, HRV characteristics are computed directly from the unprocessed time series for each patient, without artifact correction, at any stage prior to feature extraction. HRV features are then, organized into sequential inputs, where the model uses the first n-1 steps to predict the nth step. This structure allows the model to learn from temporal relationships and individual physiological trends in HRV. The sequence length used by the LSTM is optimized for each patient, allowing the model to account for individual physiological patterns. Through this, on the 7 patient cohort, the LSTM model reaches a mean F1 score of 0.9817, marking its strength across diverse patient profiles. The method provides predictions for each individual by learning from their own HRV history. Results using both traditional and state-of-the-art noise-removal techniques, like wavelet and GAN-based denoising, showed that omitting these data cleaning steps did not reduce, and in some cases even improved, prediction accuracy. These findings indicate that, for psychosis prediction based on wearable HRV data, additional data cleaning may not be necessary.