Probing Parkinson's Gait Models with SAEs: Stride Amplitude Drives Severity Predictions
mathildepapillo · x · 2026-08-05
This report investigates the underlying mechanisms general AI models rely on to predict Parkinson's disease gait severity using the CARE-PD benchmark.
The Silico team trained a Sparse Autoencoder (SAE) on model activations and built a visualization dashboard. Key findings include:
- Model embeddings successfully encode expected metrics like trunk lean, gait speed, arm-swing amplitude, and stride length.
- Feature contribution does not equal feature necessity: Removing the highest-contributing 'walk speed' feature has zero impact on the model's predictions.
- Conversely, stride amplitude is a critical necessary feature: reducing a walk's stride by 30% increases the predicted grade severity by almost 40%.
Related event: Silico Tool Decodes AI Models to Detect Parkinson's Gait(4 posts)→
More from Research
- MIT Team Uses AI to Design Novel Solvents, Boosting Sodium-Metal Battery Fast Charging — nordicinst · 2026-08-05
- Agent Harnesses and Prompting Drive Up to 30x Cost Swings, Benchmark Reveals — omarsar0 · 2026-08-05
- MIT Uses AI to Screen 100k Molecules in a Day, Solving Sodium Battery Fast-Charging Bottleneck — MIT News AI · 2026-08-05
- Measuring the Road to AGI: Insights from ARC Prize President — AnneliesGamble · 2026-08-05
- Training Models on Office Work Improves Coding by 5.8% — echen · 2026-08-05
- Pain Point in AI Academia: Difficulty Adding Authors After Conference Acceptance — liuzhuang1234 · 2026-08-05