Research map / Architectures and efficiency / Spiking networks

TIDE: Asymmetric Neural Circuits for Stabilized Temporal Inhibitory-Excitatory Dynamics

NeurIPS 2026 · Architectures and efficiency › Spiking networks · approach: Spiking transformers · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TIDE, a Continuous Thought Machine variant whose recurrent weights follow Dale-constrained Wilson-Cowan E-I dynamics with a spectral stability regularizer and a game-theoretic loss, reporting faster training and better robustness on ImageNet corruptions.

Problem

Continuous Thought Machine (CTM) uses unconstrained MLP recurrence without Dale's principle or stability guarantees, so its internal neural dynamics lack biological E-I structure and stability.

Approach

Replace CTM's unconstrained recurrent weights with non-negative magnitudes under a fixed Dale sign mask, integrate Wilson-Cowan E-I dynamics with population-specific time constants (tau_E=20 ms, tau_I=5 ms), enforce an 80:20 E-I ratio via a homeostatic loss, regularize the Perron eigenvalue with a differentiable power-iteration surrogate, and add a surprise-gated memory with an E-I retention gate.

Key results

• On ImageNet-1K, TIDE (100K steps) reaches 68.74% best top-1 versus CTM's 51.00% at 100K steps and 71.78% at 500K steps.
• TIDE improves CTM's top-1 accuracy by an average of +1.65% across ImageNet-C perturbations, while CTM is more robust under fog and snow.
• On MNIST, TIDE reaches 99.67% best accuracy vs CTM's 99.59% (200K steps), and on CIFAR-10 90.60% vs 86.16%.

Surprising findings

• Population collapse (n_I=0, i.e., CTM-like without inhibitory neurons) still achieves high top-1 accuracy, suggesting CTM is a special case of TIDE.
• CTM trained 500K steps on ImageNet-1K attains 71.78%, above TIDE's 68.74% best at 100K steps; the ImageNet gain holds only at fewer steps.

Limitations

• Many ablations and CTM comparisons use a single seed; multi-seed results are only for the main table.
• Training instability and population collapse at some hyperparameters (e.g., T=75/100, n_E/d=0.8 on Fashion-MNIST).
• Mismatched backbones: TIDE's HRF differs from CTM's pretrained ResNet-152 backbone, so comparisons are not like-for-like.

Experimental setup

Datasets: MNIST Fashion-MNIST CIFAR-10 CIFAR-100 ImageNet-1K

Benchmarks: MNIST Fashion-MNIST CIFAR-10 CIFAR-100 ImageNet-1K ImageNet-C ImageNet-R Tiny-ImageNet

Metrics: top-1 accuracy

Baselines: CTM

Builds on: Continuous Thought Machine Wilson-Cowan model Titans Dale's principle

Closest papers by meaning

• Unconditional stability of a recurrent neural circuit implementing divisive normalization (NeurIPS 2024 · similarity 0.87)
• The Expressive Leaky Memory Neuron: an Efficient and Expressive Phenomenological Neuron Model Can Solve Long-Horizon Tasks. (ICLR 2024 · similarity 0.86)
• Scaling Laws and Tradeoffs in Recurrent Networks of Expressive Neurons (NeurIPS 2026 · similarity 0.85)
• SPICED: A Synaptic Homeostasis-Inspired Framework for Unsupervised Continual EEG Decoding (NeurIPS 2025 · similarity 0.85)
• BioOSS: A Bio-Inspired Oscillatory State System with Spatio-Temporal Dynamics (NeurIPS 2025 · similarity 0.85)
• The Bayesian sampling in a canonical recurrent circuit with a diversity of inhibitory interneurons (NeurIPS 2024 · similarity 0.85)