DeepSI: Interactive Deep Learning for Semantic Interaction
summary
DeepSI-finetune fine-tunes a pretrained BERT inside the interactive loop so semantic interactions update model weights and produce user- and task-specific document embeddings. Prototype uses BERT from HuggingFace, PyTorch backend, D3.js frontend and MDS with an inverse-MDS loss that backpropagates pairwise projection distances to the model. On COVID-19 papers it forms four clear risk-factor clusters with fewer interactions than a frozen-feature baseline that only separates two clusters. DeepSI-finetune inverse-MDS
AI-written summary · about
related talks
39mAI-powered Semantic Search; A story of broken promises?
48mDeep Learning in the Browser
07mInteractive Debugging and Testing Support for Deep Learning
27mIs This Feature Interesting? Interfeat: AI for Insight Discovery
01h 00mData, visualization, and designing AI
32mBuilding semantic search using Deep Learning
02h 10mIntro to Hugging Face: Fine-tuning BERT for NLP tasks
01h 03mIt's All About AI