arXiv subjects

Wenkang Du

Publications and source records attributed to Wenkang Du.

3 recordsLinked to original sources

arXiv2026

Geometric origin of adversarial vulnerability in deep learning

Balancing training accuracy and adversarial robustness has beeen a challenge since the birth of deep learning. Here, we introduce a geometry-aware deep learning framework that leverages layer-wise local training to sculpt the internal representations of deep neural networks. This framework promotes intra-class compactness and inter-class separation in feature space, leading to manifold smoothness and adversarial robustness against white or black box attacks. The performance can be explained by \blue{data-dependent statistical mechanics of integrating out the network parameters}, \blue{supplemented by a phenomenological model} with Hebbian coupling between elements of the hidden representation. Based on the current geometry-aware learning framework, the deep network can assimilate new information into existing knowledge structures while reducing representation interference.

cs.LG↗

arXiv2025

Response function as a quantitative measure of consciousness in brain dynamics

Understanding the neural correlates of consciousness remains a central challenge in neuroscience. In this study, we investigate the relationship between consciousness and neural responsiveness by analyzing intracranial ECoG recordings from non-human primates across three distinct states: wakefulness, anesthesia, and recovery. Using a nonequilibrium recurrent neural network (RNN) model, we fit state-dependent cortical dynamics to extract the neural response function as a dynamics complexity indicator. Our findings demonstrate that the amplitude of the neural response function serves as a robust dynamical indicator of conscious state, consistent with the role of a linear response function in statistical physics. Notably, this aligns with our previous theoretical results showing that the response function in RNNs peaks near the transition between ordered and chaotic regimes -- highlighting criticality as a potential principle for sustaining flexible and responsive cortical dynamics. Empirically, we find that during wakefulness, neural responsiveness is strong, widely distributed, and consistent with rich nonequilibrium fluctuations. Under anesthesia, response amplitudes are significantly suppressed, and the network dynamics become more chaotic, indicating a loss of dynamical sensitivity. During recovery, the neural response function is elevated, supporting the gradual re-establishment of flexible and responsive activity that parallels the restoration of conscious processing. Our work suggests that a robust, brain-state-dependent neural response function may be a necessary dynamical condition for consciousness, providing a principled framework for quantifying levels of consciousness in terms of nonequilibrium responsiveness in the brain.

q-bio.NC↗

arXiv2025

Synaptic plasticity alters the nature of chaos transition in neural networks

In realistic neural circuits, both neurons and synapses are coupled in dynamics with separate time scales. The circuit functions are intimately related to these coupled dynamics. However, it remains challenging to understand the intrinsic properties of the coupled dynamics. Here, we develop the neuron-synapse coupled quasi-potential method to demonstrate how learning induces the qualitative change in macroscopic behaviors of recurrent neural networks. We find that under the Hebbian learning, a large Hebbian strength will alter the nature of the chaos transition, from a continuous type to a discontinuous type, where the onset of chaos requires a smaller synaptic gain compared to the non-plastic counterpart network. In addition, our theory predicts that under feedback and homeostatic learning, the location and type of chaos transition are retained, and only the chaotic fluctuation is adjusted. Our theoretical calculations are supported by numerical simulations.

q-bio.NC↗