Research map / Reinforcement learning / Visual and representation RL
ED-Batch: Efficient Automatic Batching of Dynamic Neural Networks via Learned Finite State Machines
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 agenda/depth-based batching with FSM-based batching | Dynamic DNNs · Tree-based models | number of batches: 37% | Helped |
| Replace agenda/depth-based batching with FSM-based batching | Dynamic DNNs · Lattice-based models | number of batches: up to 3.27 times | Helped |
| Use FSM-based batching and PQ tree memory planning instead of Cavs DyNet | ED-Batch · Chain-based models | inference throughput speedup: 1.20x, 1.11x speedup on CPU and 1.20x, 1.12x on GPU | Helped |
| Use FSM-based batching and PQ tree memory planning instead of Cavs DyNet | ED-Batch · TreeLSTM and TreeGRU | inference throughput speedup: 1.63x, 1.46x speedup on CPU and 1.23x, 1.29x speedup on GPU | Helped |
| Use FSM-based batching and PQ tree memory planning instead of Cavs DyNet | ED-Batch · Lattice-based models | inference throughput speedup: 1.32-2.97x on CPU and 2.54-3.71x on GPU | Helped |
| Use PQ tree memory allocation instead of DyNet's memory allocation (ablation) | Static subgraphs · GRUCell | latency (ms): 0.11 / 0.07 | Helped |
| Use PQ tree memory allocation instead of DyNet's memory allocation (ablation) | Static subgraphs · LSTMCell | latency (ms): 0.2 / 0.13 | Helped |
| Use PQ tree memory allocation instead of DyNet's memory allocation (ablation) | Static subgraphs · MVCell | latency (ms): 0.08 / 0.08 | No clear effect |
| Use PQ tree memory allocation instead of DyNet's memory allocation (ablation) | Static subgraphs · TreeGRU-Internal | latency (ms): 0.24 / 0.15 | Helped |
| Use PQ tree memory allocation instead of DyNet's memory allocation (ablation) | Static subgraphs · TreeGRU-Leaf | latency (ms): 0.09 / 0.07 | Helped |
| Use PQ tree memory allocation instead of DyNet's memory allocation (ablation) | Static subgraphs · TreeLSTM-Internal | latency (ms): 0.19 / 0.12 | Helped |
| Use PQ tree memory allocation instead of DyNet's memory allocation (ablation) | Static subgraphs · TreeLSTM-Leaf | latency (ms): 0.12 / 0.09 | Helped |
Closest papers by meaning
- On-the-fly Operation Batching in Dynamic Computation Graphs (NeurIPS 2017 · similarity 0.91)
- Deep Learning with Dynamic Computation Graphs (ICLR 2017 · similarity 0.88)
- Small batch deep reinforcement learning (NeurIPS 2023 · similarity 0.86)
- Dynamic Deep Neural Networks: Optimizing Accuracy-Efficiency Trade-Offs by Selective Execution (AAAI 2018 · similarity 0.86)
- Small Batch Size Training for Language Models: When Vanilla SGD Works, and Why Gradient Accumulation is Wasteful (NeurIPS 2025 · similarity 0.86)
- Dynamic Sparse Graph for Efficient Deep Learning (ICLR 2019 · similarity 0.86)
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.