Research map / Architectures and efficiency / RNNs and normalization
Phased LSTM: Accelerating Recurrent Network Training for Long or Event-based Sequences
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace LSTM units with Phased LSTM (time-gated updates) | RNN on N-MNIST event streams · N-MNIST | accuracy at epoch 1: 90.32% vs 40.87% for BN-LSTM | Helped |
| Replace LSTM units with Phased LSTM (time-gated updates) | RNN on N-MNIST event streams · N-MNIST test rho=0.4 and rho=1.0 | accuracy: 95.11% vs 94.79% (rho=0.4); 97.27% vs 96.55% (rho=1.0) | Helped |
| Replace frame-based CNN with event-stream Phased LSTM | N-MNIST · N-MNIST test rho=0.4 | accuracy: 95.11% vs 90.67% for CNN | Helped |
| Use Phased LSTM time gating (open ratio 5%) instead of full BN-LSTM updates | N-MNIST · N-MNIST | LSTM updates per neuron: 159 vs 3153 per neuron (~20x fewer) | Helped |
| Oversampling input 10x (0.1 ms vs 1 ms) with Phased LSTM vs LSTM/BN-LSTM | Frequency discrimination task · oversampled condition | accuracy | Mixed |
| Add Phased LSTM time gate (learned tau, s) instead of timestamp as input to LSTM | Superimposed frequencies task · Superimposed sine waves | training convergence speed and variance | Helped |
| Use longer oscillation period tau for the time gate (ablation) | Adding task · sequence length 490-510 | training convergence | Helped |
| Phased LSTM sampling-aligned time gates in lip reading audio/video layers | Lip reading (GRID) video-only · GRID video-only, low and high resolution | test loss convergence | Helped |
| Increase audio MFCC sampling rate from 25 fps-matched to 100 Hz (ablation) | Lip reading (GRID) · GRID video-only | accuracy | No clear effect |
Closest papers by meaning
- Efficient recurrent architectures through activity sparsity and sparse back-propagation through time (ICLR 2023 · similarity 0.89)
- Dynamical Isometry and a Mean Field Theory of RNNs: Gating Enables Signal Propagation in Recurrent Neural Networks (ICML 2018 · similarity 0.88)
- AntisymmetricRNN: A Dynamical System View on Recurrent Neural Networks (ICLR 2019 · similarity 0.88)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.88)
- Exploiting Symmetric Temporally Sparse BPTT for Efficient RNN Training (AAAI 2024 · similarity 0.87)
- xLSTM: Extended Long Short-Term Memory (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.