Drop, Swap, and Generate: A Self-Supervised Approach for Generating Neural Activity

Name: Drop, Swap, and Generate: A Self-Supervised Approach for Generating Neural Activity
Uploaded: 2026-10-09T03:03:08.438574+00:00
Duration: 711 s
Description: Swap-VAE is an unsupervised method that learns disentangled representations of neural activity from both synthetic and primate recordings. It uses a generative model with instance-specific alignment losses and BlockSwap augmentation to separate behavioral style from intrinsic neural content, achieving high accuracy in decoding tasks while demonstrating superior disentanglement scores.

Swap-VAE is an unsupervised method that learns disentangled representations of neural activity from both synthetic and primate recordings. It uses a generative model with instance-specific alignment losses and BlockSwap augmentation to separate behavioral style from intrinsic neural content, achieving high accuracy in decoding tasks while demonstrating superior disentanglement scores.

Open the full episode