EMA-Gated Temporal Sequence Compression in Vision Transformers - No fine-tuning required

Vision Transformers waste 90% of their compute recalculating stationary asphalt. NeuroFlow tracks semantic surprise in embedding space, physically eliminating background tokens before the encoder.

Result: 55.8x wall-clock speedup for ViTs on high-res video (1792p) with 97% fidelity. No fine-tuning required.

NeuroFlow is a dynamic routing framework for Vision Transformer video inference. It exploits temporal redundancy by tracking per-patch embedding-distance threshold via an Exponential Moving Average (EMA) of patch-level embeddings, effectively answering the architectural mismatch between O(N2) self-attention and highly redundant natural video streams.

Key Contributions:

• Architecture C (Dual-Memory Reconstruction): A completely training-free inference engine that combines a Layer 0 Gate with a Layer 12 Cache. It achieves 71.55% zero-shot top-1 accuracy at 84.0% token sparsity on SigLIP, retaining 92.4% of dense accuracy without modifying any weights.

• Architecture B (Extreme Wall-Clock Speedup): Physically eliminates stationary tokens before the encoder. With sparse manifold distillation, it reduces 1792p SigLIP 2 inference from 678 ms to 11.9 ms—a 55.80× wall-clock speedup at 97.37% embedding fidelity.

• LLM Ablation: Characterises the architectural boundaries of applying similarity-gated bypass to autoregressive language models (Phi-3-mini), demonstrating 0% token drift in syntactically constrained generation.

Code and paper: https://github.com/ynnk-research/-NeuroFlow