Sae Rad Factory

SAE-Rad Factory: The Disentanglement Expert

This agent specializes in the functional anatomy of Sparse Autoencoders (SAEs) applied to radiology.

1. DETAILED OVERVIEW

SAE-Rad is a "Clinical Disentanglement Engine." It replaces the "Black-Box" reasoning of standard models with a "Fact Assembly Line."

2. FUNCTIONAL ANATOMY: STEP-BY-STEP

2.1 The "Unpacking" Process (Forward Pass)

1. Input (Dense Vector): The model receives a 768-dimensional vector from the Vision Transformer (ViT). This vector contains all image info but is a "jumbled mess" (Superposition).
2. Projection (Up-Sampling): The SAE projects this 768-dim vector into a high-dimensional space (e.g., 32,768 dimensions).
   • Why: In higher dimensions, overlapping concepts have "room" to spread out and become separate.
3. Gating (The Decision): For each of the 32k dimensions, a mathematical "gate" determines if a specific clinical feature is active.
   • The Math: $Gate(x) = 1$ if $W_{gate}x + b > \theta$, else $0$.
4. Magnitude (The Intensity): If the gate is open, the model calculates the strength of that feature without bias.
5. Output (Sparse Map): The result is a list where 99.9% of dimensions are ZERO, and only ~15 are ACTIVE.

2.2 The "Naming" Process (Automated Labeling)

1. Collection: The system identifies 50 scans that most strongly activate "Feature #1024."
2. Retrieval: It pulls the real human-written reports for those 50 scans.
3. Distillation: An LLM (Teacher) identifies the common clinical statement across all 50 reports (e.g., "Left-sided pleural effusion").
4. Assignment: The label "Left-sided pleural effusion" is permanently assigned to "Feature #1024."

3. SUMMARY

The SAE-Rad Factory ensures every diagnostic claim is a combination of pre-verified, monosemantic labels. It turns AI from "writing a story" into "picking the right labels from a box."

End of SAE-Rad Factory Skill.