A recurrent neural network model of memory

Wander · Cache · Recall

When chickadees stash their seeds, the same part of their brain that tells them where they are forms a barcode that's linked to the storage site, so they can later recall it from the place cue alone. Click through the three steps below to run the network yourself.

Pick a spot on the track
Click anywhere on the track to place the bird and run the network.
0← click to place →100
neuron activity
← 800 RNN neurons, sorted by preferred location →

Each bar is one neuron's activity. A smooth hump means a place code; a jagged, sparse pattern is a barcode.

place code barcode recalled
✓
The network, as a web of neurons

The same 800 neurons drawn as a network — brighter dots are more active. Once you've cached, show stored replays the barcode held in memory.

Live model from chingf/barcodes (Fang et al., 2025, eLife): divisive-normalized RNN dynamics with one-shot Hebbian storage. 800 neurons, 100-state circular track — scaled down from the paper's 5000 to run in-browser.

Adapted by Chenguang Li for Dendrite @ Substack.