We consider Graph Anisotropic Diffusion (GAD), a recently proposed [4] model of graph neural networks, that can be trained to predict desired properties of the graph by performing learnable diffusion of node features on it. In contrast with similar methods, GAD introduces anisotropy of said diffusion by incorporating filters built from the graph’s Fiedler vector. In present work we attempt to improve this approach by increasing the dimension of the space in which GAD runs – that is, adding filters built from other low-frequency eigenmodes of the graph (eigenvectors of its Laplacian). We report the performance of such “dimension-augmented” GAD in predicting the chemical properties of small organic molecules from the ZINC dataset [5].
Here we describe the OCHEM-R platform, developed to propose chemical pathways based on a consensus prediction of retrosynthetic pathways using eight different methods. The developed software allows users to visualize predicted reactions and identify similar reactions using retrieval augmented generation (RAG). The platform is publicly available at https://ochem.eu.
In this work, we present a comparative analysis of SMILES, DeepSMILES, and SELFIES string representations for chemical structures in neural machine translation tasks in cheminformatics. Using transformer-based models, we systematically evaluated their effectiveness in translating between these representations and the corresponding linguistic IUPAC nomenclature. The experimental results demonstrate comparable performance for all three string representations, with SMILES achieving a marginally higher accuracy (99.30
Recent theoretical advances have demonstrated the universality of linear Reservoir Computing (RC) models equipped with nonlinear readouts, showing their potential to approximate arbitrary inputoutput mappings. However, practical insights into the selection and performance of nonlinear readouts are limited. This paper addresses this gap by systematically benchmarking a spectrum of nonlinear readouts within linear RC frameworks. Our results reveal the practical tradeoffs in accuracy and efficiency across tasks, offering insights on how to train performant RC systems with linear recurrence. These findings provide valuable guidelines for designing efficient recurrent architectures that combine theoretical guarantees with state-of-the-art performance in sequential data processing.
Echo state networks (ESNs) are recurrent neural networks belonging to the reservoir computing framework. While ESNs are conceptually simple, their successful application can be challenging. For instance, there are no generally applicable methods for optimally setting important hyperparameters like the reservoir spectral radius. Therefore, the development of strategies for appropriately initializing ESNs is an active field of research. Plasticity-based pretraining is a bio-inspired reservoir optimization approach. We analyze if this approach is able to improve the results of a non-optimized ESN and if the pretraining effects can be explained by the influence on the spectral radius. In our experiments, we evaluate the effects of four synaptic plasticity rules (SP), namely anti-Ojas, normalized anti-Hebbian, BCM, and dual-threshold BCM, and of intrinsic plasticity (IP) on the Mackey-Glass, NARMA, and Lorenz series. IP significantly improves the ESNs performance across all three benchmarks whereas this is not the case for any of the considered SP rules. Overall, the influence of plasticity on the spectral radius is not sufficient for explaining the pretraining effects. The cases, in which plasticity significantly worsens the results, however, can be explained by the spectral radius having been moved to a disadvantageous value.
Small Boolean Liquid State Machines (B-LSMs) offer a simplified yet expressive biologically inspired model of recurrent computation, in which network attractor dynamics can be systematically analyzed. In their untrained form, B-LSMs exhibit complex, often chaotic dynamics with short-lived memory traces. This study investigates how local synaptic plasticity (STDP) and a global plasticity (GP) mechanism jointly shape the attractor landscapes of these networks. Specifically, we show that synaptic modifications can drive B-LSMs to exhibit exponentially many attractors, each corresponding to a potential memory. Such high attractor regimes are attainable through global synaptic crafting. Under noisy background conditions, STDP tends to drive the networks back to low attractor regimes; however, when receiving carefully designed inputs, STDP maintains the networks' rich attractor dynamics. Overall, our findings highlight the theoretical potential for storing an impressive number of memories in recurrent neural networks, with significant implications for theoretical neuroscience and neuromorphic computing.
As lead discovery increasingly targets hard-to-drug proteins, the expansion of chemical space presents unprecedented opportunities for hit identification – yet scalable, effective technologies to exploit these vast spaces remain underdeveloped. We describe recent advances to our FRASE-based hit-finding platform (FRASE-bot), including integration of an AI-powered 3D pharmacophore screening across multi-billion-compound libraries, a Hit-Triage Pretrained Transformer (Hit-TPT), and alchemical binding free energy (ABFE) simulations. We also introduce emerging strategies for leveraging phenotypic data to support both hit identification and lead optimization. The platform's utility is demonstrated across several case studies, including our winning entries in CACHE Challenges #1 and #2.
Reservoir Computing (RC) enables efficiently-trained deep Recurrent Neural Networks (RNNs) by removing the need to train the hierarchy of representations of the input sequences. In this paper, we analyze the performance and the dynamical behavior of RC models, specifically Deep Bidirectional Echo State Networks (Deep-BiESNs), applied to Natural Language Processing (NLP) tasks. We compare the performance of Deep-BiESNs against fully-trained NLP baseline models on six common NLP tasks: three sequence-to-vector tasks for sequence-level classification and three sequence-to-sequence tasks for token-level labeling. Experimental results demonstrate that Deep-BiESNs achieve comparable or superior performance to these baseline models. We then adapt the class activation mapping technique for explainability to analyze the dynamical properties of these deep RC models, highlighting how the hierarchy of representations in Deep-BiESNs layers contributes to forming the class prediction in the different NLP tasks. Investigating time scales in deep RNN layers is highly relevant for NLP because language inherently involves dependencies that occur over various temporal horizons. The findings not only underscore the potential of Deep ESNs as a competitive and efficient alternative for NLP applications, but also contribute to a deeper understanding of how to effectively model such architectures for addressing other NLP challenges.
Predicting molecular properties is an important challenge in drug discovery. Machine learning methods, particularly those based on transformer architectures, have become increasingly popular for this task by learning molecular representations directly from chemical structure [1, 2]. Motivated by progress in natural language processing, many recent approaches apply models of the BERT (Bidirectional Encoder Representations from Transformers) architecture [3] to molecular data using SMILES as the input format [4–9]. In this study, we revisit core design assumptions that originate in natural language processing but are often carried over to molecular tasks without modification. We explore how variations in masking strategies, pretraining dataset size, and model size influence downstream performance in molecular property prediction. Our findings suggest that common practices inherited from natural language processing do not always yield optimal results in this setting. In particular, we observe that increasing the masking ratio can lead to significant improvements, while scaling up the model or dataset size results in stagnating gains despite higher computational cost (Fig. 1). Building on these observations, we develop MolEncoder, a BERT-style model that achieves improved performance on standard benchmarks while remaining more efficient than existing approaches. These insights highlight meaningful differences between molecular and textual learning settings. By identifying design choices better suited to chemical data, we aim to support more effective and efficient model development for researchers working in drug discovery and related fields.
This work presents a real-time syllable-level recognition system for LIBRAS, the Brazilian Sign Language. The system extracts 2D hand landmarks using MediaPipe and a Gaussian Temporal Smoothing technique to reduce frame-wise jitter. Two deep learning models are implemented for classification: a Multilayer Perceptron (MLP) and a Convolutional Neural Network (CNN). A dataset of 27,456 samples covering all 26 LIBRAS syllables was constructed for training and evaluation. Experiments were conducted on both a desktop workstation and a Raspberry Pi 4 to assess classification accuracy and inference latency. The CNN model achieves an average accuracy of 97.4
This paper tackles the challenges of sparse and unevenly distributed of ring-like mechanical LiDAR point clouds in road environment perception for autonomous driving by proposing the LPSF-LiDARNet framework. The framework enhances fine-grained 3D semantic segmentation through inter-frame spatiotemporal feature enhancement and balanced voxel sampling. Temporal window fusion is achieved via multi-frame stacking, integrating complementary temporal features to mitigate single-frame incompleteness. Spatially, log-polar coordinate voxel sampling leverages spatial distribution patterns to improve feature consistency. Additionally, adaptive heterogeneous convolution kernels with dynamic attention mechanisms are introduced, combined with semantic-level data augmentation, to optimize feature extraction for sparse points and rare samples. The framework demonstrates superior performance over existing models in experiments using SemanticKITTI and local datasets, validating the effectiveness of its spatiotemporal fusion strategy and sampling mechanism. Ultimately, the model achieves a segmentation accuracy of 96.4
Parameter sharing is widely employed in policy networks within multi-agent reinforcement learning (MARL) methods, enhancing efficiency and performance in specific tasks. However, this approach also introduces significant challenges, particularly poor stability. In this study, we use QMIX [10] as a case study to investigate the stability issue inherent in parameter sharing. Our analysis reveals that this issue stem from the inability of parameter sharing networks to learn multiple distinct policies simultaneously. To validate this hypothesis, we introduce a specialized task designed to assess the limitations of parameter sharing in learning diverse policies. Building on these efforts, we propose a novel method to enhance the stability of parameter sharing and implement it with QMIX. Experimental results demonstrate that our approach not only improves stability but also enables QMIX to achieve superior performance across tasks. The source code and the videos of the diversity policies is available on https://github.com/Yurui-Li/ImS.
Remote sensing image tampering detection is crucial in information security due to advanced image generation techniques. Previous research has not sufficiently explored generalization ability in satellite imagery. This paper firstly introduces domain generalization for satellite image tampering detection. Conventional spatial patterns vary across tampering methods, leading to insufficient information for feature extraction. We propose a Multi-Scale Frequency-Aware Representation Learning framework, namely MSFA-Net, which integrates spatial and frequency-domain knowledge to extract domain-invariant tampering features. To enhance local and global semantic correlation from different scale feature maps, we introduce a multi-scale contextual attention module. Furthermore, a CLIP-Guided Semantic Knowledge Distillation module is proposed to boost the generalizable models robustness through visual language knowledge. Extensive experiments conducted on three remote-sensing benchmarks demonstrate our frameworks effectiveness and superior cross-domain performance.
The exponential growth of data in IoT environments requires the development of resource-efficient feature selection methods adapted for constrained devices. This work introduces two novel efficiency modifications for the ReliefF method. The first enhancement involves fixed-point arithmetic, which lowers computational overhead while improving energy and memory efficiency. The second improvement involves a single-linkage clustering step that reduces sample size without compromising data relevance. Evaluated on nine diverse datasets, ranging from those with high sample numbers to microarrays, proposed optimizations demonstrate robust performance. A 16-bit fixed-point representation achieves feature rankings comparable to 64-bit floating-point baselines, offering significant efficiency improvements. The clustering step drastically reduces execution times for large-sample datasets while preserving classification accuracy. Key findings show that the integer part of fixed-point representation, which determines the representable range, is more critical than precision. Furthermore, 16-bit implementations provide an optimal balance for most IoT applications, and clustering is vital for scalability in datasets with numerous samples. These innovations make ReliefF an effective solution for energy-limited IoT systems.
In recent years, single-view hand mesh reconstruction has seen significant advancements. While existing methods perform well on motion-capture datasets, the lack of accurately annotated in-the-wild datasets remains a critical challenge for supervised 3D hand mesh reconstruction. Additionally, due to domain gaps, directly applying existing models to in-the-wild datasets has yielded unsatisfactory results. To address these issues, we propose UnDAHand, which aims to transfer knowledge learned from motion-capture source domains to in-the-wild target domains in an unsupervised manner. Specifically, to bridge the domain gap between different datasets, we introduce a fine-to-coarse pseudo label update strategy, which utilizes a "teacher-for-teacher" correction process for generating pseudo labels. Furthermore, recognizing the importance of consistency learning and the diversity of data in enhancing model transferability and generalization, we develop an augmentation consistency learning module to capture more robust data representations. Experimental results demonstrate the effectiveness of our method in domain adaptation across various datasets and its robustness when applied to unlabeled in-the-wild data.
As an essential cross-modal task, radiology report generation (RRG) has drawn much attention recently. The existing approaches normally rely on large-scale annotated data (coupled image-report pairs) to generate satisfying radiology reports. As we know, it is time-consuming and labor-intensive to build these large-scale annotated datasets and the data demand of the growing model increases simultaneously. Thanks to the global understanding and proper unlabeled exploitation of masked auto-encoder (MAE), we attempt to extend MAE to learn both global and specific features for RRG, probably with unlabeled image information. In this way, we can alleviate the annotation burden and leverage less labeled data to achieve competitive performance. Therefore, we propose a novel approach, namely joint MAE (JMAE), to simultaneously mine the global and specific representation for RRG. Moreover, we also intend to inject potential unlabeled Chest X-ray images to augment our joint model. Extensive experiments on two widely used datasets demonstrate that our approach can outperform many representative baselines, and detailed analysis further shows the feasibility and effectiveness of our approach.
Federated learning is a decentralized machine learning approach where models are trained collaboratively across multiple devices or nodes holding local data without sharing that data directly. It enables privacy-preserving, scalable, and collaborative machine learning. One of the key challenges in federated learning is its inefficiency in handling scenarios where data is highly imbalanced and non-independent and identically distributed (non-IID) across local nodes, leading to biased global models and slow convergence. This paper introduces a peer-to-peer refinement mechanism combined with FedAvg aggregation to enhance model accuracy in highly imbalanced and non-IID federated learning scenarios. Experiments were conducted on the MNIST, Fashion-MNIST and CIFAR-10 datasets using a Dirichlet distribution with α = 0.1 to simulate highly imbalanced and non-IID data scenarios. The results demonstrated that the proposed approach achieved higher accuracy, 98.17
Time series forecasting remains a fundamental challenge in data science. We introduce SWIFT, a novel neural architecture that synergistically combines selective state space models (Mamba) with multi-scale dilated convolutions for enhanced time series forecasting. Our approach incorporates: (1) a Selective Temporal State Space module extending Mamba with time series-specific gating; (2) a Multi-Scale Dilated Convolutional Network with adaptive receptive fields; and (3) a Feature Interaction Bridge facilitating cross-pathway information exchange. Experiments on six benchmark datasets demonstrate SWIFT outperforms other common methods, achieving 6.5
In medical image segmentation, single-source domain generalization aims to improve model robustness and performance on unseen target domains using only single-source training data. Single-source domain generalization can mitigate the challenge of limited domain diversity in training, which arises from the impracticality of freely collecting patient data from frontline healthcare institutions due to ethical, legal, and privacy constraints. In this paper, we propose Constrained Learnable Channel-wise Normalization (CLCN), a novel method for single-source domain generalization in medical image segmentation. CLCN comprises two key components: Learnable Channel-wise Normalization (LCN) and Feature Consistency Constraint (FCC). LCN dynamically adjusts feature distributions by learning channel-wise normalization parameters, reducing redundancy and improving the models adaptability to diverse and unknown target domains. To address the risk of over-transformation, FCC is introduced to regulate the extent of feature transformations by enforcing consistency between features with and without channel-wise normalization. For evaluation, we conduct experiments on the Prostate dataset across six single-source domain generalization tasks, demonstrating that CLCN outperforms existing methods. Ablation studies further validate the effectiveness of LCN and FCC, showcasing their complementary roles in balancing generalization ability and feature representation. These results highlight the potential of CLCN in advancing domain-adaptive medical image segmentation.
Point clouds often suffer from noise during capture, which can negatively impact downstream tasks such as analysis and surface reconstruction. Denoising is thus a fundamental task in 3D vision. Large point clouds frequently exceed hardware capabilities, requiring division into smaller patches for efficient processing. Existing denoising methods typically use kNN or ballquery to divide point clouds into patches to learn shape features. However, these methods struggle to recover local features accurately and can be misled by outliers. To address these challenges, we propose the RegionExpansion method, which iteratively selects patches with a relatively small radius on the CUDA platform, enabling localized region expansion. By refining patch extraction, our method balances efficiency and accuracy, successfully removing most outliers during preprocessing and preserving sharp geometric features. Experimental results demonstrate that our method achieves better performance compared to common patch-fetching methods. The source code is available at: https://github.com/only-tao/RegionExpansion.
