Path integration is essential for spatial navigation. Experimental studies have identified neural correlates for path integration, but exactly how the neural system accomplishes this computation remains unresolved. Here, we adopt recurrent neural networks (RNNs) trained to perform a path integration task to explore this issue. After training, we borrow neuroscience prior knowledge and methods to unfold the black box of the trained model, including: clarifying neuron types based on their receptive fields, dissecting information flows between neuron groups by pruning their connections, and analyzing internal dynamics of neuron groups using the attractor framework. Intriguingly, we uncover a hierarchical information processing pathway embedded in the RNN model, along which velocity information of an agent is first forwarded to band cells, band and grid cells then coordinate to carry out path integration, and finally grid cells output the agent location. Inspired by the RNN-based study, we construct a neural circuit model, in which band cells form one-dimensional (1D) continuous attractor neural networks (CANNs) and serve as upstream neurons to support downstream grid cells to carry out path integration in the 2D space. Our study challenges the conventional view of considering grid cells as the principal velocity integrator, and supports a neural circuit model with the hierarchy of band and grid cells.

Abstract--Dendrite-inspired neurons have been widely used in tasks such as image classification due to low computational complexity and fast inference speed. T emporal data prediction, as a key machine learning task, plays a key role in real-time scenarios such as sensor data analysis, financial forecasting, and urban traffic management. However, existing dendrite-inspired neurons are mainly designed for static data. Studies on capturing dynamic features and modeling long-term dependencies in temporal sequences remain limited. Efficient architectures specifically designed for temporal sequence prediction are still lacking. In this paper, we propose Memory-DD, a low-complexity dendrite-inspired neuron model. Memory-DD consists of two dendrite-inspired neuron groups that contain no nonlinear activation functions but can still realize nonlinear mappings. Compared with traditional neurons without dendritic functions, Memory-DD requires only two neuron groups to extract logical relationships between features in input sequences. This design effectively captures temporal dependencies and is suitable for both classification and regression tasks on sequence data. Experimental results show that Memory-DD achieves an average accuracy of 89.41% on 18 temporal classification benchmark datasets, outperforming LSTM by 4.25%. On 9 temporal regression datasets, it reaches comparable performance to LSTM, while using only 50% of the parameters and reducing computational complexity (FLOPs) by 27.7%. These results demonstrate that Memory-DD successfully extends the low-complexity advantages of dendrite-inspired neurons to temporal prediction, providing a low-complexity and efficient solution for time-series data processing. ITH the rapid development of information technology, massive temporal sequence data have become the foundation of modern society, ranging from industrial IoT to financial markets.

Regret in Large Language Models refers to their explicit regret expression when presented with evidence contradicting their previously generated misinformation. Studying the regret mechanism is crucial for enhancing model reliability and helps in revealing how cognition is coded in neural networks. To understand this mechanism, we need to first identify regret expressions in model outputs, then analyze their internal representation. This analysis requires examining the model's hidden states, where information processing occurs at the neuron level. However, this faces three key challenges: (1) the absence of specialized datasets capturing regret expressions, (2) the lack of metrics to find the optimal regret representation layer, and (3) the lack of metrics for identifying and analyzing regret neurons. Addressing these limitations, we propose: (1) a workflow for constructing a comprehensive regret dataset through strategically designed prompting scenarios, (2) the Supervised Compression-Decoupling Index (S-CDI) metric to identify optimal regret representation layers, and (3) the Regret Dominance Score (RDS) metric to identify regret neurons and the Group Impact Coefficient (GIC) to analyze activation patterns. Our experimental results successfully identified the optimal regret representation layer using the S-CDI metric, which significantly enhanced performance in probe classification experiments. Additionally, we discovered an M-shaped decoupling pattern across model layers, revealing how information processing alternates between coupling and decoupling phases. Through the RDS metric, we categorized neurons into three distinct functional groups: regret neurons, non-regret neurons, and dual neurons.

Large language models struggle with representing and generating rare tokens despite their importance in specialized domains. In this study, we identify neuron structures with exceptionally strong influence on language model's prediction of rare tokens, termed as rare token neurons, and investigate the mechanism for their emergence and behavior. These neurons exhibit a characteristic three-phase organization (plateau, power-law, and rapid decay) that emerges dynamically during training, evolving from a homogeneous initial state to a functionally differentiated architecture. In the activation space, rare token neurons form a coordinated subnetwork that selectively co-activates while avoiding co-activation with other neurons. This functional specialization potentially correlates with the development of heavy-tailed weight distributions, suggesting a statistical mechanical basis for emergent specialization.

Understanding the inner workings of neural networks is essential for enhancing model performance and interpretability. Current research predominantly focuses on examining the connection between individual neurons and the model's final predictions. Which suffers from challenges in interpreting the internal workings of the model, particularly when neurons encode multiple unrelated features. In this paper, we propose a novel framework that transitions the focus from analyzing individual neurons to investigating groups of neurons, shifting the emphasis from neuron-output relationships to functional interaction between neurons. Our automated framework, NeurFlow, first identifies core neurons and clusters them into groups based on shared functional relationships, enabling a more coherent and interpretable view of the network's internal processes. This approach facilitates the construction of a hierarchical circuit representing neuron interactions across layers, thus improving interpretability while reducing computational costs. Our extensive empirical studies validate the fidelity of our proposed NeurFlow. Additionally, we showcase its utility in practical applications such as image debugging and automatic concept labeling, thereby highlighting its potential to advance the field of neural network explainability.

Safety fine-tuning algorithms are widely used to reduce harmful outputs in language models, but how they achieve this remain unclear. Studying the Direct Preference Optimization (DPO) algorithm for toxicity reduction, current explanations claim that DPO achieves this by dampening the activations of toxic MLP neurons. However, through activation patching, we show that this explanation is incomplete. Projections onto a toxicity probe's direction show that only 4.9% of toxicity reduction comes from dampened toxic neurons. Instead, DPO reduces toxicity through distributed activation shifts across a majority of neurons, progressively shifting MLP layer outputs away from toxicity. These shifts accumulate across four neuron groups: two reducing toxicity and two promoting anti-toxicity. Activation patching validates the cumulative roles of these groups, where patching all identified groups effectively replicates DPO's effects. These findings illustrate DPO's mechanism: it reduces toxicity by accumulating small activation shifts across many neurons throughout the layers. Our findings provide new mechanistic insights into how safety fine-tuning reduces harmful outputs in language models.

Uncovering the fundamental neural correlates of biological intelligence, developing mathematical models, and conducting computational simulations are critical for advancing new paradigms in artificial intelligence (AI). In this study, we implemented a comprehensive visual decision-making model that spans from visual input to behavioral output, using a neural dynamics modeling approach. Drawing inspiration from the key components of the dorsal visual pathway in primates, our model not only aligns closely with human behavior but also reflects neural activities in primates, and achieving accuracy comparable to convolutional neural networks (CNNs). Moreover, magnetic resonance imaging (MRI) identified key neuroimaging features such as structural connections and functional connectivity that are associated with performance in perceptual decision-making tasks. A neuroimaging-informed fine-tuning approach was introduced and applied to the model, leading to performance improvements that paralleled the behavioral variations observed among subjects. Compared to classical deep learning models, our model more accurately replicates the behavioral performance of biological intelligence, relying on the structural characteristics of biological neural networks rather than extensive training data, and demonstrating enhanced resilience to perturbation.

Spiking Neural Networks (SNNs) and neuromorphic models are more efficient and have more biological realism than the activation functions typically used in deep neural networks, transformer models and generative AI. SNNs have local learning rules, are able to learn on small data sets, and can adapt through neuromodulation. Although research has shown their advantages, there are still few compelling practical applications, especially at the edge where sensors and actuators need to be processed in a timely fashion. One reason for this might be that SNNs are much more challenging to understand, build, and operate due to their intrinsic properties. For instance, the mathematical foundation involves differential equations rather than basic activation functions. To address these challenges, we have developed CARLsim++. It is an integrated toolbox that enables fast and easy creation of neuromorphic applications. It encapsulates the mathematical intrinsics and low-level C++ programming by providing a graphical user interface for users who do not have a background in software engineering but still want to create neuromorphic models. Developers can easily configure inputs and outputs to devices and robots. These can be accurately simulated before deploying on physical devices. CARLsim++ can lead to rapid development of neuromorphic applications for simulation or edge processing.

A widely used interpretation by example We design, implement, and evaluate DeepEverest, a system for the query is, "find the top-inputs that produce the highest activation efficient execution of interpretation by example queries over the values for an individual neuron or a group of neurons" [12, 14, 21, activation values of a deep neural network. DeepEverest consists 33, 50, 57, 58, 61]. Another common query is, "for any input, find of an efficient indexing technique and a query execution algorithm the k-nearest neighbors in the dataset using the activation values of a with various optimizations. We prove that the proposed query group of neurons based on the proximity in the latent space defined execution algorithm is instance optimal.

This can lead to various inefficiencies in terms of potential parallelization. Local learning is an approach to model-parallelism that removes the standard end-to-end learning setup and utilizes local objective functions to permit parallel learning amongst model components in a deep network. Recent works have demonstrated that variants of local learning can lead to efficient training of modern deep networks. However, in terms of how much computation can be distributed, these approaches are typically limited by the number of layers in a network. In this work we propose to study how local learning can be applied at the level of splitting layers or modules into sub-components, adding a notion of width-wise modularity to the existing depth-wise modularity associated with local learning. We investigate local-learning penalties that permit such models to be trained efficiently. Our experiments on the CIFAR-10, CIFAR-100, and Imagenet32 datasets demonstrate that introducing width-level modularity can lead to computational advantages over existing methods based on local learning and opens new opportunities for improved model-parallel distributed training.