Machine Perception & Interpretability
Research notes, visual probes, and interactive maps of neural network internals by Tom White (drib.net · github.com/dribnet).
-
Abstract neurons in vision towers: from Rosetta neurons to Motam
Porting native neuron communities into a vision-language model's vision tower (Qwen3.8-27B) surfaces cross-community units encoding abstract relations such as "moving through a medium". Includes ten-line differentiable fluid-streamline probes, switch-off ablations, a 15-neuron confusion matrix, a 6,000-photo neutral-pool Rosetta scan, and initial Jacobian-lens readouts.
-
Show of Hands
Moving away from sparse autoencoders to raw native units: a two-ink screen-print evolved to activate a single Rosetta neuron (Dravid et al.) for raised hands shared across twelve independently trained vision models (~90M to 7B parameters), transferring to held-out networks including one trained exclusively on plants and animals.
-
Basis of Attraction
The working series behind CVPR 2026's Synesthetic Visions: aligning SigLIP, OpenAI CLIP, and LAION CLIP into a shared embedding space via Procrustes rotation and a joint TopK sparse autoencoder to surface unnamed perceptual gestalts, verified by retrieval across 2.9M held-out CC3M photographs.
-
Atlas of Perception
Mapping the elementary visual vocabulary of SigLIP2 with a sparse autoencoder and a radially symmetric diffusion visualiser, rendering individual machine-learned concepts as navigable visual studies and video loops.
-
Maximum Activations
Interactive cartographic maps of linear steering directions inside large language models—including refusal, untruth, and concept vectors—clustered from hundreds of maximally activating prompts.
-
Sparse Latents
Cartographic exploration of twelve Gemma Scope sparse autoencoder features inside Gemma 2 2B, clustering and rendering thousands of maximally activating prompts into navigable semantic maps.
Earlier & Related
- swarm.art — live multi-agent collective drawing venue (2026)
- Shared Visual Abstractions (arXiv:1912.04217) & Perception Engines
- Sampling Generative Networks (arXiv:1609.04468)
- Generative Adversarial Networks: An Overview (IEEE Signal Processing Magazine, 2018)