Another update from the Flyborg laboratory.
I'm spending the weekend performing brain surgery on a digital fruit fly.
A memory prosthesis for the connectome
The reconstructed brain has 127,400 neurons and 14.7 million connections, but lacks much of the biology behind learning.
So I'm building a synthetic memory prosthesis.
The idea is to intercept activity from the fly's learning circuits, translate it into searchable vectors, and let the nervous system retrieve experiences later.
We can capture, store, and feed neural representations back into the connectome. We can even rewind, suppress one memory's feedback, and measure what changes.
Finding a learning signal
Getting the fly to decide what to remember is trickier.
We found neurons whose spike timing changed in response to sugar. Could we identify reinforcement from those patterns?
Nope.
Neutral and bitter stimuli produced similar changes. We tried a more sophisticated receiver comparing neurons.
Also nope.
Time to consult the neuroscience literature.
Real flies use dopamine, octopamine, and receptor dynamics to regulate learning. Our simulation models electrical signals through fixed connections, without that chemistry.
Adding the modeled chemistry
So we found Temur Turayev's open-source synthetic dopamine model, reproduced its experiments in an isolated mushroom-body network, and began adapting the approach to our full-brain simulation.
Now we've got something new.
A passive synthetic chemical layer observes simulated neural activity and produces modeled dopamine and receptor states without altering the original brain.
And sensory signals reach it. Sugar and bitter inputs change the modeled chemical states. Cut the sensory pathways, and those differences disappear.
Here's what caught my attention: neural spikes happen quickly, but chemical states evolve more slowly, sometimes continuing to change after the input stops.
Three kinds of computation
Suddenly we're looking at three kinds of computation:
- Fast electrical activity processing sensory information.
- Slower modeled chemistry that could regulate learning and responsiveness.
- Persistent synthetic memory retaining experiences after the original activity disappears.
They don't have to operate at the same speed.
We haven't solved reinforcement learning for flies... yet.
The chemical readout can't reliably distinguish sugar from neutral or bitter, and we haven't demonstrated learned preferences.
What this architecture makes possible
But the architecture has my brain spinning.
Most agents treat memory as something a model reads or writes. What if a continuously operating neural subsystem managed attention, memory eligibility, and consolidation?
What if we attached an artificial reasoning network to that same neuromodulatory infrastructure?
And because it's digital, we can instrument, modify, and replace individual components.
For now, the fly lives in a little Godot terrarium.
Next: validate the anatomical pathways and connect a meaningful learning signal to memory.
It's a different way to approach machine learning.
More soon. 🪰