Research map / Architectures and efficiency / RNNs and normalization
Two-Stage Unit Tying for Simplifying Differentiable Logic Gate Networks
Seungheon Lee, Jeongmin Sun, Jaeyong Chung
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add Stage-B Binary Split refinement to Stage-A screening (ablation) | Ours · CIFAR-10(M) 50% tied | accuracy (%): 59.22 -> 61.44 | Improved |
| add Stage-B Binary Split refinement to Stage-A screening (ablation) | Ours · MNIST(M) 50% tied | accuracy (%): 91.71 -> 96.20 | Improved |
| add post-tying fine-tuning (ablation) | Ours · CIFAR-10(M) 50% tied | accuracy (%): 61.44 -> 69.22 | Improved |
| use Binary Split refinement instead of Tail Rescore (ablation) | Ours · CIFAR-10(M) 30% tied | accuracy (%): 66.33 -> 67.30 | Improved |
| use Binary Split refinement instead of Full-set Rescore (ablation) | Ours · CIFAR-10(M) 30% tied | accuracy (%): 66.88 -> 67.30 | Improved |
| use logit-level MSE instead of KL with tau=1 (ablation) | Ours · CIFAR-10(M) 50% tied, Stage-A | accuracy (%): 55.88 -> 59.22 | Improved |
| raise KL temperature from 1 to 20 (ablation) | Ours · CIFAR-10(M) 50% tied, Stage-A | accuracy (%): 55.88 -> 59.14 | Improved |
| Gauss-Newton screening plus refinement vs FD-DL (N=128) | Ours · CIFAR-10(M) 50% tied | accuracy (%): 56.05 -> 61.44 | Improved |
| Ours vs Fisher-TL saliency | Ours · CIFAR-10(M) 50% tied | accuracy (%): 53.97 -> 61.44 | Improved |
| Ours vs ConstProb saliency | Ours · MNIST(M) 50% tied | accuracy (%): 31.18 -> 96.20 | Improved |
| increase FD-DL calibration size N from 16 to 64 (ablation) | FD-DL · CIFAR-10(M) 30% tied | accuracy (%): 66.38 -> 67.30 | Improved |
| use Stage-A only vs FD-DL at 30% tied | Ours (Stage-A only) · CIFAR-10(M) 30% tied | accuracy (%): 66.49 -> 67.19 | Mixed |
Closest papers by meaning
- Mind the Gap: Removing the Discretization Gap in Differentiable Logic Gate Networks (NeurIPS 2025 · similarity 0.87)
- Convolutional Differentiable Logic Gate Networks (NeurIPS 2024 · similarity 0.87)
- Towards Next-Generation Logic Synthesis: A Scalable Neural Circuit Generation Framework (NeurIPS 2024 · similarity 0.85)
- Retrieval-Guided Reinforcement Learning for Boolean Circuit Minimization (ICLR 2024 · similarity 0.85)
- Rethinking Circuit Completeness in Language Models: AND, OR, and ADDER Gates (NeurIPS 2025 · similarity 0.85)
- Circuit Representation Learning with Masked Gate Modeling and Verilog-AIG Alignment (ICLR 2025 · similarity 0.84)
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.
