Research map / Architectures and efficiency / RNNs and normalization
Ultrafast On-Chip Online Learning via Spline Locality in Kolmogorov–Arnold Networks
Duc Hoang, Aarush Gupta, Philip Harris
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 |
|---|---|---|---|
| use KAN instead of parameter-matched MLP for drifting regression under fixed-point | online learner on FPGA · Adaptive function approximation (<6,2>) | cumulative regret: 97.6 -> 13.2 | Improved |
| use KAN instead of parameter-matched MLP for qubit readout under fixed-point | online learner on FPGA · Single-shot qubit readout (<7,3>) | accuracy (%): 50.4% -> 92.8% | Improved |
| scale MLP up to MLP-L on qubit readout (ablation) | MLP · Single-shot qubit readout (<7,3>) | accuracy (%): 50.4% -> 49.0% | No clear effect |
| scale MLP up to MLP-L on drifting regression (ablation) | MLP · Adaptive function approximation (<6,2>) | cumulative regret: 97.6 -> 48.3 | Improved |
| use KAN actor-critic instead of MLP actor-critic on non-stationary Acrobot | online actor-critic · Non-stationary Acrobot control | episodic return: -500 -> -104 | Improved |
| use online DQN update with a large MLP instead of actor-critic (ablation) | MLP on Acrobot · Non-stationary Acrobot control | episodic return: -500 -> -298 | Improved |
| add fixed-point quantization at low bitwidth (ablation) | MLP-L · Adaptive function approximation (2 integer bits) | final cumulative regret | Hurt |
| add fixed-point quantization at low bitwidth (ablation) | KAN · Adaptive function approximation (2 integer bits) | final cumulative regret | No clear effect |
| increase KAN grid size G (ablation) | KAN · Qubit readout (7-bit fixed-point) | running accuracy | Improved |
| increase MLP parameter count under fixed-point (ablation) | MLP · Qubit readout (10-bit fixed-point) | running accuracy | No clear effect |
| add LayerNorm or LSUV initialization to MLP (ablation) | MLP · fixed-point online learning | performance gap to KAN | No clear effect |
| keep updating online during angular drift instead of freezing the learner (ablation) | KAN digit classifier · Digit classification with angular drift | running accuracy: about 20% -> 80% | Improved |
Closest papers by meaning
- Analog Foundation Models (NeurIPS 2025 · similarity 0.84)
- Towards training digitally-tied analog blocks via hybrid gradient computation (NeurIPS 2024 · similarity 0.84)
- PowerMLP: An Efficient Version of KAN (AAAI 2025 · similarity 0.84)
- KAN: Kolmogorov–Arnold Networks (ICLR 2025 · similarity 0.84)
- Exponential Hardness of Optimization from the Locality in Quantum Neural Networks (AAAI 2024 · similarity 0.84)
- TITAN: A Trajectory-Informed Technique for Adaptive Parameter Freezing in Large-Scale VQE (NeurIPS 2025 · similarity 0.83)
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.
