TDSNNs: Competitive Topographic Deep Spiking Neural Networks for Visual Cortex Modeling (2508.04270v1)

Abstract: The primate visual cortex exhibits topographic organization, where functionally similar neurons are spatially clustered, a structure widely believed to enhance neural processing efficiency. While prior works have demonstrated that conventional deep ANNs can develop topographic representations, these models largely neglect crucial temporal dynamics. This oversight often leads to significant performance degradation in tasks like object recognition and compromises their biological fidelity. To address this, we leverage spiking neural networks (SNNs), which inherently capture spike-based temporal dynamics and offer enhanced biological plausibility. We propose a novel Spatio-Temporal Constraints (STC) loss function for topographic deep spiking neural networks (TDSNNs), successfully replicating the hierarchical spatial functional organization observed in the primate visual cortex from low-level sensory input to high-level abstract representations. Our results show that STC effectively generates representative topographic features across simulated visual cortical areas. While introducing topography typically leads to significant performance degradation in ANNs, our spiking architecture exhibits a remarkably small performance drop (No drop in ImageNet top-1 accuracy, compared to a 3\% drop observed in TopoNet, which is the best-performing topographic ANN so far) and outperforms topographic ANNs in brain-likeness. We also reveal that topographic organization facilitates efficient and stable temporal information processing via the spike mechanism in TDSNNs, contributing to model robustness. These findings suggest that TDSNNs offer a compelling balance between computational performance and brain-like features, providing not only a framework for interpreting neural science phenomena but also novel insights for designing more efficient and robust deep learning models.

• The paper introduces TDSNNs that integrate an STC loss to optimize both spatial and temporal dynamics in spiking neural networks.
• The methodology maps SNN layers onto a virtual cortical sheet, reproducing V1 and IT-like topographies observed in the primate brain.
• Results demonstrate robust ImageNet accuracy and enhanced Brain-Score metrics, highlighting efficient temporal processing and brain-likeness.

"TDSNNs: Competitive Topographic Deep Spiking Neural Networks for Visual Cortex Modeling" (2508.04270)

Introduction and Background

The paper introduces Topographic Deep Spiking Neural Networks (TDSNNs), which combine deep Spiking Neural Networks (SNNs) with topographic constraints to model the visual cortex effectively. Spatio-temporal processing is inherent in biological neural systems, yet most conventional Artificial Neural Networks (ANNs) fail to incorporate these dynamics adequately, often resulting in performance degradation and reduced biological fidelity. The incorporation of Spatio-Temporal Constraints (STC) in TDSNNs addresses these limitations by promoting the hierarchical spatial functional organization observed in primate visual cortex, from low-level sensory inputs to high-level abstract representations.

Figure 1: TDSNNs leverage temporal information. (Left) Spike train entropy shifts reveal topography-dependent temporal dynamics across various inference timesteps. (Right) TDSNNs' spiking mechanisms solve persistent recognition degradation by leveraging these temporal patterns.

Methodology

The methodology focuses on mapping SNN layers onto a virtual cortical sheet, pre-optimizing neuronal positions, and implementing an STC loss function. Each SNN layer is represented as a set of units mapped to coordinates on a virtual cortical sheet, maintaining non-uniform embedding to reflect biological structures. Pre-optimization involves adjusting neuron positions based on firing rates to align with topographic principles.

The STC loss is a novel component, derived from the biological need to balance metabolic cost and information-theoretic efficiency. It incorporates long-timescale (firing rates) and short-timescale (spike timing synchronies) components to promote similarity in responses among spatially neighboring neurons.

Figure 2: Overview of the methodology for inducing visual cortex-like neural organization in SNN architectures.

Results: V1 and IT-like Topography

Neurons in the primate V1 exhibit organized structures with systematic maps for stimuli categories like orientation and color. TDSNNs emulate this organization, shown by preference maps and smoothness analyses indicating higher correlation in neuronal firing rates among spatially adjacent neurons.

In IT-like areas, category-specific selectivity emerges in deeper layers of TDSNNs, with neurons organized into contiguous clusters reflecting real-world stimulus categories. This organization enhances spatial-coherence of response patterns, visible in more pronounced neural clusters for similar functional representations.

Figure 3: Analysis of V1-like topography of TDSNNs highlights structured preference maps and smooth correlations akin to biological V1.

Performance and Brain-likeness

Remarkably, TDSNNs maintain, and sometimes exceed, the task performance of non-topographic models, evidenced by robust ImageNet classification accuracy without the typical accuracy drop associated with topographic architectures like TopoNet. The blend of topographic constraints and SNN architectures enhances brain-likeness, achieving higher Brain-Score metrics, particularly in V1 and IT regions.

Figure 4: Analysis of IT-like topography shows spatially coherent selectivity maps.

Improved Temporal Information Processing

Induction of topography fundamentally transforms temporal information processing within TDSNNs, enhancing network efficiency and robustness. Spiking activity patterns demonstrate synchronized neuronal responses, and changes in neural connectivity suggest a more stable and efficient representation hierarchy.

The analysis of Fisher Information reveals the progressive shift from raw signal fidelity in early layers to noise-resistant encoding in deeper layers, highlighting the role of topography in facilitating efficient temporal processing.

Figure 5: TDSNNs exhibit robust task performance and stability across varying strengths of topographic organization.

Conclusion

TDSNNs represent a significant advancement in computational models of the visual cortex, integrating biologically-feasible topographic and temporal dynamics. They offer insights into neural system evolution and practical benefits for designing efficient and robust neural networks that closely mirror brain-like processing. Future directions could explore larger-scale implementations and the integration of other biologically plausible elements, expanding the potential applications of TDSNNs in both neuroscience research and advanced AI systems.