Circuit Tracing: Finding Medical Features in Gemma 3Gemma 3 uses Neuronpedia's tool to organize and activate medical knowledge contextually with prompts4 min read·thedatasense.comFeb 4
What Does Medical VLM Actually See? Experiments with MedGemma and Sparse AutoencodersSparse Autoencoders improve medical image interpretability in MedGemma, affecting model responses based on question phrasing5 min read·thedatasense.comJan 24
Fluent But Wrong: LLM and HealthcareDiscusses limitations of deploying language models in healthcare, stressing the need for human verification and system improvements for safety9 min read·thedatasense.comJan 20
Opening the Black Box: How to See What Your Vision Language Model is Actually Looking AtUse Chefer et al.'s techniques to visualize model attention, enhancing trust in AI diagnostics through vision-language model explainability19 min read·thedatasense.comJan 16
Data Generating ProcessUnderstand the Data Generating Process (DGP) to explore mechanisms behind data, enabling better insight into uncertainty and statistical methods7 min read·thedatasense.comJan 14