NOW OUT in Scientific Reports! the latest from @mehdiorouji.bsky.social, me, and many others! Task relevant autoencoding enhances machine learning for human neuroscience www.nature.com/articles/s41... thread 🧵👇 - ThreadSky

meganakpeters.bsky.social • 49 days ago

NOW OUT in Scientific Reports! the latest from @mehdiorouji.bsky.social, me, and many others!

Task relevant autoencoding enhances machine learning for human neuroscience
https://www.nature.com/articles/s41598-024-83867-6

thread 🧵👇

Comments

meganakpeters.bsky.social•49 days ago

In human neuro--here, focused on fMRI--we often use machine learning techniques to help reveal neural representations relevant to behavior. But those large ML models take TONS of data to train... and fMRI datasets are feature-rich but sample-poor. So lots of ML models can be prone to overfitting.

meganakpeters.bsky.social•49 days ago

We wanted to see if a small change to a 'simple' ML model could help clean up representations *even under really non-ideal conditions*. We hooked a classifier to the bottleneck of a fully-connected autoencoder & trained it under SEVERE data paucity on benchmark datasets: MNIST & Fashion MNIST.

meganakpeters.bsky.social•49 days ago

This Task-Relevant Autoencoder via Classifier Enhancement (TRACE) outperformed a number of comparison models on 4 metrics, even when there was SO LITTLE data to train it. So next, we tested TRACE on fMRI data from 59 subjects who observed animals and objects.

meganakpeters.bsky.social•49 days ago

TRACE outperformed alternative models nearly unilaterally, showing up to 12% increased classification accuracy and up to 56% improvement in discovering “cleaner”, task-relevant representations. These results showcase TRACE’s potential for a wide variety of data related to human behavior.

meganakpeters.bsky.social•49 days ago

TRACE is simple by design, definitely not meant to 'keep up with the Jonses' w.r.t fancy state-of-the-art ML/AI models. But this means it has fewer parameters and is less prone to overfitting in human fMRI.

meganakpeters.bsky.social•49 days ago

So perhaps TRACE will be useful in your work! With lots of efforts not only discovering and characterizing neural 'representations' (a loaded word for the philosophers out there) but also manipulating them noninvasively with Decoded Neurofeedback, TRACE may be a useful added tool in your toolbox!

Comments

Posting Rules

Reply