Research map / Architectures and efficiency / RNNs and normalization

MEET: Towards Memory-Efficient Temporal Sparse Deep Neural Networks

CVPR 2025 · Architectures and efficiency › RNNs and normalization · approach: Efficient network compression and training · 2 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 MEET, a co-designed NAS and mixed spatial-temporal suppression framework that trades activations for weights to shrink neuron-state memory of temporal delta-sigma DNNs so they fit in on-chip SRAM while preserving accuracy and compute savings.

Problem

Temporal delta-sigma DNNs save computation on event-driven platforms but need large neuron-state memory that exceeds on-chip SRAM; directly compressing states breaks delta-sigma linearity and accumulates errors in long-term processing.

Approach

Trade fewer activations for more weights (A→W tradeoff) via compute- and state-memory-aware NAS (CSM-NAS) with fused inverted-bottleneck blocks, then compress the extra weights with pruning; co-design the network with mixed spatial (early stages) and temporal (deep stages) suppression.

Key results

• MEET reduces total memory by 5.1-13.3x versus the most computation-efficient temporal DNNs while preserving accuracy and dynamic cycles (abstract).
• MEET-M (mixed) reaches 86.47% PCK@0.5 with 6.8 MB state and 12.8 MB total memory, within 36 MB SRAM of GrAI-VIP, with 2.7x energy reduction vs baseline.
• On UA-DETRAC, MEET-Full achieves 51.50% mAP@0.5 vs 50.15% for MobileNetV2, with total memory cut from 86.4 MB to 26.7 MB.

Measured effects reported in the paper

Surprising findings

• Mixed spatial-temporal execution outperformed full temporal execution in dynamic cycles for semantic segmentation (66M vs 60M), contrary to the object detection trend.
• Increasing static computation (FuseV2 MB, CSM-NAS) is largely eliminated by temporal sparsity, so dynamic cycles stay similar despite higher static cycles.

Limitations

• Mixed execution lags full temporal execution on object detection due to high temporal redundancy in surveillance video, so λ must be tuned per deployment.
• Evaluation relies on a performance estimator (GrAIFlow) for event-driven hardware rather than physical measurement.
• Weight compression beyond pruning is only noted as future integration; compression of weights is not fully explored.

Experimental setup

Datasets: MPII UA-DETRAC Cityscapes

Benchmarks: MPII UA-DETRAC Cityscapes

Metrics: PCK@0.5 mAP@0.5 mIoU dynamic cycles state memory total memory normalized active energy

Baselines: MobileNetV2 EfficientNet-Lite2 EfficientNet-Lite4 ResNet18 VGGBN16 Hoyer STAR EvNet CATS TempDiff RRM ELSE

Builds on: EvNet EfficientNet-Lite MobileNetV2 NSGA-III SimpleBaselines

Closest papers by meaning

• Sparse Delta Memory: Scaling the State of Linear RNNs through Sparsity (NeurIPS 2026 · similarity 0.87)
• Exploiting Symmetric Temporally Sparse BPTT for Efficient RNN Training (AAAI 2024 · similarity 0.86)
• Accelerating Linear Recurrent Neural Networks for the Edge with Unstructured Sparsity (ICML 2025 · similarity 0.85)
• One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.85)
• Towards Practical Real-Time Neural Video Compression (CVPR 2025 · similarity 0.85)
• Auto-Compressing Networks (NeurIPS 2025 · similarity 0.85)