model-efficiency

The Very Similar Objects Recognition repository focuses on advancing object recognition through deep learning, inspired by the Chihuahua-Muffin classification challenge.

Neural Importance-based pruning to prune the least “important” neurons from feed forward layers in a network, while maintaining the constraint of minimal impact on the loss / accuracy.

A controlled 180-run study of DeepSeek-inspired MLA, sparse MoE routing, V3-style load balancing, and multi-token prediction under constrained compute.

OpenAI Parameter Golf experiments for parameter-constrained language modeling, ablations, and efficient architecture tradeoffs.

Adaptive inference algorithm for transformers inspired by quantum collapse (SR framework)

Token cost is a design problem, not a billing problem. Most LLM cost overruns come from architectural waste, not model pricing. This tool is a token waste profiler that helps you understand where your tokens are going and which ones are useless.

"Ablation study on CNN depth, data augmentation, and background removal for efficient plant disease classification — 97.8% accuracy on potato, <7.5M params, <55ms inference."

Turn a small dense LLM into a Mixture-of-Experts model, then specialize the experts by distilling from a teacher. Reproducible toolkit + honest benchmarks (Qwen2.5-0.5B demo).

A brain-inspired language model that gets cheaper as it gets bigger. Top-1 spiking experts + an offline "sleep" phase that rewires the network → up to 7.1× less serving energy and ~1/43 the active compute of a dense model its size, while staying quality-competitive.

Reliability-constrained visual-token budgeting for energy-efficient vision-language model inference.

A lightweight CNN efficiency study for cassava leaf disease classification using controlled architecture and training-strategy ablations.

Study the intersection of model efficiency and calibration quality under distribution shift