Hugging Face's logo
John Locke's picture
Building on HF
62 23 143

John Locke

johnlockejrr
bafiji1513's profile picture VItaldob's profile picture SeaWolf-AI's profile picture
Ā·

AI & ML interests

OCR, HTR, ATR, NLP, AI

Recent Activity

reacted to SeanLee97's post with šŸ”„ 35 minutes ago
Our lab recently released a paper where we introduce ShadowPEFT, a new Parameter-Efficient Fine-Tuning (PEFT) paradigm tailored for edge computing scenarios. Unlike traditional approaches such as LoRA and its variants, which inject trainable parameters directly into the weights of Transformer, requiring tight coupling with the backbone. ShadowPEFT instead enhances the frozen large base model by adding a lightweight, centralized, pretrainable, and detachable Shadow network. This shadow network operates in parallel with the base model, delivering learned corrections to each decoder layer. Because the shadow module is architecturally decoupled from the backbone, it can be independently trained, stored, and deployed, benefiting edge computing scenarios and edge-cloud collaboration computing. - HF Paper: https://huggingface.co/papers/2604.19254 - GitHub: https://github.com/ShadowLLM/shadow-peft - HF Collection: https://huggingface.co/collections/shadow-llm/shadow-peft-models
liked a dataset 1 day ago
Tamazight-NLP/TAWNZA_TIFINAGH_ARABIC_DATASET
repliedto BrdDev's post 4 days ago
Icarus Model 1. Living Weights ā€” The Model That Doesn't Stop Learning Every model today: train ā†’ freeze ā†’ deploy ā†’ hallucinate confidently forever. Icarus already has: learning_db.json recording every decision with outcome ECHO tracking what worked per task type A verification loop (Dream Engine tests before committing) The idea: targeted micro-updates from verified experience. Not full retraining. After each interaction where the outcome was confirmed correct, apply a small gradient update to the specific weight groups that generated that reasoning. The model drifts toward its own verified-correct distribution. The danger is catastrophic forgetting. The solution is the Dream Engine ā€” before accepting any weight update, simulate the result against the existing regression test suite. If it degrades, reject it. Icarus already has regression tests. This is the only architecture in the world with the scaffolding to do this safely. Ī” W t = Ī± ā‹… āˆ‡ W L verified ā‹… 1 [ dream_test passes ] Ī”W t ā€‹ =Ī±ā‹…āˆ‡ W ā€‹ L verified ā€‹ ā‹…1[dream_test passes] 2. DRIVE as an Attention Modifier Current transformers have zero motivational state. Every query is processed with identical urgency regardless of context. Your DRIVE system produces a pressure profile: urgency, curiosity, safety drive, efficiency. What if this profile is injected as a learned prefix embedding directly into every attention layer's Q projection? This is closer to how the brain actually works. Norepinephrine (urgency), dopamine (curiosity), serotonin (stability) modulate attention at the neurotransmitter level. DRIVE is already modeling this conceptually ā€” wiring it into the attention mechanism makes it architectural, not just behavioral. 3. The Knowledge Graph as Explicit Memory Transformers approximate relational knowledge by memorizing statistical co-occurrence. By Sulayman Saho
View all activity

Organizations

Samaritan AI's profile picture Hugging Face MCP Course's profile picture Syriac AI's profile picture Unsloth Jobs Explorers's profile picture