Matched-Learning-Rate Analysis of Attention Drift and Transfer Retention in Fine-Tuned CLIP
An 80-run matched-learning-rate study of CLIP attention drift, LoRA, and zero-shot transfer retention.
Work
Results
These pages are the work that follows the ideas: access, compute limits, and visible adaptation. Venue and preprint links live on each paper page, as publication details rather than as the point of the site.
Papers
Real-Time On-Device Diffusion: Practical Acceleration via Fused Low-Bit Kernels
A systems paper on accelerating diffusion inference with fused low-bit kernels and cache-update fusion.
Value-Aligned Evaluation for Everyday AI
A framework for evaluating AI by explanation, contestability, accessibility, and fit—not just accuracy.
Human Judgment in the Loop
Why meaningful human oversight requires more than rubber-stamping machine output.
Text2Sign: A Single-GPU Diffusion Baseline for Text-to-Sign Language Video Generation
Peer-reviewed IEEE Access article on a single-GPU diffusion baseline for text-to-sign language video generation.