Today, Goodfire launched Silico 🔥 I’ve had a lot enjoyable working with this wonderful platform for AI analysis…. and I figured I would share what that appears like. General-purpose human movement AI…

Today, Goodfire launched Silico 🔥 I’ve had a lot enjoyable working with this wonderful platform for AI analysis…. and I figured I would share what that appears like.
General-purpose human movement AI fashions can meaningfully detect Parkinson’s gait impairment. But how do they do it?? I used Silico to interrupt these fashions open and determine what about somebody’s stroll truly drives their predictions.

➡️ Silico finds their embeddings encode a bunch of information we count on. Probe a frozen movement mannequin embedding for fundamental gait measurements. You’ll discover trunk lean, gait pace, arm-swing amplitude, stride size…
➡️ So what information does the mannequin acc depend on for predicting Parkinson’s severity? Silico skilled a sparse dictionary (SAE) on the activations and constructed a dashboard to point out their prime samples. So for every prime function, we are able to take a look at the way it reveals up in information AND in predictions.
➡️ *But* function contribution /= function necessity. The finest “walk speed” function may be eliminated with 0 affect to the mannequin’s prediction. But the other is true for stride amplitude. Reducing a stroll’s stride by 30% will increase grade severity by virtually 40%.

We know fashions usually do not cause the best way we guess. Figuring out what mechanisms they DO depend on is time intensive. Silico from Goodfire hurries up that work in a manner that I can 1. perceive and a couple of. monitor. It’s a lot enjoyable!

…extra