parameter-efficiency

This repository contains the source code and results for the experiments presented in Evaluating Parameter-Based Training Performance of Neural Networks and Variational Quantum Circuits.

Zero-hidden neural networks that solve non-linear problems through temporal depth, not spatial layers. 90.14% MNIST with 480 parameters. Intelligence is not depth — it's resonance. Time is the ultimate hidden layer. OdyssNet proves it.

Layer-wise ablation of Whisper ASR models. How to cut 15% of model size without losing accuracy

Does the Universal/Looped-Transformer trick (reuse one block N times instead of stacking N) work on Mamba? Weight-shared depth in a state-space model, with honest per-seed results.

CRS-LM: Structure-aware context reduction for tiny language models under Parameter Golf constraints

Tiny language-model research lab exploring Cellular Neural Network (CeNN) recurrent state cores as alternatives to Transformer layers, with distillation, shared-weight iteration, and efficiency benchmarks.

This repository contains the source code and results for the experiments presented in The Impact of Parameter Count on Sarcasm Detection Using BERT-Based Models.

🧬 Neuro-Symbolic Activation Discovery: Using Genetic Programming to discover domain-specific activation functions and transfer them across scientific domains. Achieves 18-21% higher parameter efficiency with 5-6× fewer parameters.