IS-SNN removes activation normalization from deep SNNs via topology-aware weight standardization folded into static weights, matching dynamic BN accuracy on ImageNet (68.05%) while cutting FPGA neuron LUT usage by 96.4%.
Micro-batch training with batch-channel normalization and weight standardization
5 Pith papers cite this work. Polarity classification is still indexing.
5
Pith papers citing it
citation-role summary
background 1
citation-polarity summary
years
2026 5roles
background 1polarities
background 1representative citing papers
SpikON introduces learnable threshold and weight techniques plus a dual-parallel SNN accelerator that cut training latency by 32% and energy by 35% while delivering 7-27x throughput gains over Apple M4 GPU and TPU-like designs.
Manifold constraints via the new MACRO optimizer independently bound activation scales and enforce rotational equilibrium in LLM pre-training, subsuming RMS normalization and decoupled weight decay while delivering competitive performance with convergence guarantees.
Splitting weight matrices into a fixed-norm direction and learnable per-row/column magnitudes improves LLM training over AdamW/Muon, removes weight decay and warmup, and transfers the optimal LR across width.
A polynomial preconditioning layer controls singular value spectra of transformer weights to stabilize pre-training, shown effective on Llama-1B and supported by convergence theory for deep linear networks.
citing papers explorer
-
Intrinsically Stable Spiking Neural Networks: Overcoming the Performance Barrier in the Absence of Batch Normalization
IS-SNN removes activation normalization from deep SNNs via topology-aware weight standardization folded into static weights, matching dynamic BN accuracy on ImageNet (68.05%) while cutting FPGA neuron LUT usage by 96.4%.
-
SpikON: A Dual-Parallel and Efficient Accelerator for Online Spiking Neural Networks Learning
SpikON introduces learnable threshold and weight techniques plus a dual-parallel SNN accelerator that cut training latency by 32% and energy by 35% while delivering 7-27x throughput gains over Apple M4 GPU and TPU-like designs.
-
Demystifying Manifold Constraints in LLM Pre-training
Manifold constraints via the new MACRO optimizer independently bound activation scales and enforce rotational equilibrium in LLM pre-training, subsuming RMS normalization and decoupled weight decay while delivering competitive performance with convergence guarantees.
-
Improving Neural Network Training by Decoupling the Magnitude and Direction of Weight Vectors
Splitting weight matrices into a fixed-norm direction and learnable per-row/column magnitudes improves LLM training over AdamW/Muon, removes weight decay and warmup, and transfers the optimal LR across width.
-
PC Layer: Polynomial Weight Preconditioning for Improving LLM Pre-Training
A polynomial preconditioning layer controls singular value spectra of transformer weights to stabilize pre-training, shown effective on Llama-1B and supported by convergence theory for deep linear networks.