dev.log
Over the years, ML researchers have burned countless hours training large neural networks. However, its inner workings remain opaque and incomprehensible to humans. There have been many research efforts in the past [1] [2] introspect what happens inside neural networks. Most recently, Sparse Autoencoders have shown to be a promising tool in analysing Large Language Model’s behavior.
This work explores how Sparse Autoencoders (SAE) can be applied to diffusion models and vision encoders. We demonstrate that SAEs can be used to decompose model’s activation space into interpretable, semantic components. Furthermore, to the best of our knowledge, we are the first to show that SAEs trained directly on flow matching models (flux). This work is accompanied by fluxlens, a minimal visualisation tool I’ve built to explore SAE features and multimodal datasets.
This work began early October in collaboration with krea.ai who’ve dragged me out of GPU poverty line. The initial exploration and visualisation tool was built on top of krea’s internal assets. However, to facilitate SAE open research efforts, I’ve retrained SAEs on public multimodal dataset (CC3M). The code for SAE training will be open sourced in the coming weeks.
