Neuroscientists often study sequential neuronal activations—neurons firing one after another—to explain short-term memory.
We show that two geometric mechanisms can drive this:
🌐 Slow-point manifolds,
🌐 Limit cycles.
But which one is used? 🤔
We show that two geometric mechanisms can drive this:
🌐 Slow-point manifolds,
🌐 Limit cycles.
But which one is used? 🤔
Comments
🤖 the task structure and
🤖 the learning rate of the RNN.
We further derive scaling laws for how long RNNs can store info before failing.
📖 https://arxiv.org/abs/2502.17433
@geometric-intel.bsky.social @neurreps.bsky.social @ai-ucsb.bsky.social @ucsbengineering.bsky.social @ucsbece.bsky.social