Research map / Architectures and efficiency / CNN design

Don’t Think It Twice: Exploit Shift Invariance for Efficient Online Streaming Inference of CNNs

AAAI 2025 · Architectures and efficiency › CNN design · approach: Theory of convolutional inductive bias · 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 StreamiNNC, a way to run pretrained CNNs in online streaming mode by reusing overlapping-window activations, using signal padding (with optional retraining) and analyzing pooling shift-error bounds.

Problem

Overlapping-window CNN inference for time-series recomputes much of the same activations; zero-padding and pooling break translation invariance, so prior streaming methods do not generalize to standard CNN architectures.

Approach

Exploit convolution shift-invariance to compute only new sub-window activations and buffer previous ones, replacing zero-padding with signal padding (buffered previous-layer activations, retrained with an extended input) and deriving upper error bounds for approximately shift-invariant pooling, with alignment of window step and pooling size.

Key results

• Signal-padding retraining of f_PPG reduces streaming NRMSE to 2.03% (MAE 3.38 BPM vs 3.36 full), versus 5.83% with pretrained zero-padding weights.
• Without retraining, streaming NRMSE is 3.32-3.55% for f_EEG and f_ACC, and approximate streaming gives 2.12% for f_ACC.
• Streaming speedup scales linearly with L/S: coefficient 0.13 for approximate and 0.15 for exact streaming inference.

Measured effects reported in the paper

Surprising findings

• Approximate streaming with zero-padding causes a large error for f_PPG (NRMSE 19.9%), while the same approach stays low for f_ACC (2.12%).
• Pooling misalignment between window step and pooling size increases f_PPG MAE to 6.73 BPM, showing alignment is critical.

Limitations

• Only the feature extractor is streamed; the fully connected classifier is not shift-invariant and is excluded.
• Signal padding requires retraining and extra activation buffers, increasing memory and latency.
• Pooling error bounds saturate and can be large in the worst case, so shift approximation is only guaranteed when aligned.

Experimental setup

Datasets: PPGDalia CHB-MIT

Benchmarks: PPGDalia CHB-MIT

Metrics: NRMSE MAE (BPM) speedup

Baselines: full inference zero-padding streaming approximate streaming

Builds on: Khandelwal real-time CNN inference Stream Buffer causal convolution

Closest papers by meaning

• Auto-reconfiguration for Latency Minimization in CPU-based DNN Serving (ICML 2025 · similarity 0.84)
• Convolutions and More as Einsum: A Tensor Network Perspective with Advances for Second-Order Methods (NeurIPS 2024 · similarity 0.83)
• Algorithm Development in Neural Networks: Insights from the Streaming Parity Task (ICML 2025 · similarity 0.83)
• SuBiTO: Synopsis-based Training Optimization for Continuous Real-Time Neural Learning over Big Streaming Data (AAAI 2025 · similarity 0.83)
• Don't Pause! Every prediction matters in a streaming video (NeurIPS 2026 · similarity 0.83)
• LMUFormer: Low Complexity Yet Powerful Spiking Model With Legendre Memory Units (ICLR 2024 · similarity 0.83)