Neurological Biocompute field building

Seeking Fundingactive

Develop long-term learning and memory retention in neuron-culture biocomputing via multi-day training protocols on Cortical Labs’ platform and an open-source light-microscope scanner to track structural changes.

Project Details

Current neuron-based biocomputing is able to replicate a limited amount of current AI capabilities. However, the primary hole in the field is long-term learning in cultures. Experiments can train neurons to perform both reinforcement learning tasks, such as "Doom 93" or Pong. We also have some classification work, such as MNIST (Cortical Labs, currently unpublished). However, these results are over 30-1 hour, and the learning is ephemeral. We seek to address this in two ways.
Cortical Labs' online platform enables multiple-day, month-long experiments with biological neurons. The main advantage of the platform is that experiments can be done with relatively low time investment. We are seeking to induce long-term learning and experiment with testing residual skill retention. Examples include a pseudo-sleep protocol to help encode skills over multiple days, and label permutation experiments to assess if learning is fully lost or not. Aka, if you swap the labels around after a cooling-off period, does that old "memory" make it hard to learn the new labels? The output here is an academic paper or bioarxiv note.
In-house, we have our own neurons and hardware. We are asking for additional funding to support the work of building out a "Celpha"-style open-source light microscope scanner for coarse connectomics. While true synaptic changes are below light microscope resolutions, gross structural changes are visible (neurite outgrowth) before they are measurable with our microelectrode arrays. Thus, we would gain deeper access to long-term changes in cell culture, and combined with our more flexible hardware, we are able to test long-term learning better.

Theory of Impact

Biological computing (With neurons) has a fundamentally different safety profile than Silicon AI. Biological systems need to be grown, need tighter networking timing to maximize performance 10ms or less, have natural intelligence analogs, and improving neuronal computing directly translates into human enhancement. Lastly, because it is inherently more "gross," it makes the involved regulation and treaties more likely. The Silicon AI replacement thesis is driven by profit-motivated actors via the extreme energy efficiency of neurons vs current AI systems/data centers.
Networking requirements for separate biological computers need to be on the order of magnitude of 10ms or less, or else one loses the advantage of individual neurons. So there is a reduced risk of Bio AI getting out of the box. It can't network its way out or be globally distributed. If you don't feed it, it dies.
Improving neurological computing promotes human enhancement in two directions. A better understanding of neurons and their long-term learning systems enables better mind emulation. Obviously, there are risks with enabling better coarse-grain simulations, but there are clearly vast safety upsides if the simulations are human-like. Second, biological computing focused on neurons will eventually work with synthetic biology approaches to improve the function of neurons/brain tissue. These improvements would be directly beneficial to humans and would reduce the overall intelligence gap between AI systems and humans while still providing the advantages of advanced AI.

People

Grant Getzelman

Team Member

Grants Received– no grants recorded

Add grant

Discussion

Private comments in this thread are only shown to approved funders and grant reviewers. Email hi@grantmaking.ai or to request access.

Matt Brooks2mo

Are you happy with your AI generated one-liner, or would you like to replace it with something else?

GGrant Getzelman2mo

The AI-generated one-liner is fine. The compression ratio is acceptable, but misses the why.

Grant ReviewerPrivate2mo?