The Biological Computing Company (TBC) lives somewhere in between brain research and AI software. Its founders, Alex Ksendzovsky and Jonathan Pomeraniec, are both neuroscientists and practicing neurosurgeons. Its vice president of product and engineering, John Wittig, is among other things an experienced IC designer. Together, they and their powerful research team set out to explore how the human brain functions, from the neurons up.
But their research may have the entirely intended consequence of solving some of the most intractable problems with today’s generative AI models: problems that so far have proved resistant to the massive scaling the industry is pursuing. So what are they actually doing?
To begin with, those AI problems are serious. Generative AI is unable to generate long video clips without the image deteriorating and some objects simply disappearing. When AI models are retrained with additional data to extend their scope or to correct problems, they often forget some of the capabilities they originally had. Models fail on inputs very different from their training sets. And so far, AI models must be laboriously retrained to expand their capabilities: continuous learning from the environment has proved an elusive goal.
Yet none of these problems is an issue for a human brain, or even a rat’s brain. Something about the AI models is missing, and it clearly isn’t just a matter of scaling into earth orbit.
These issues have led some researchers to move beyond massive, static models like LLMs, LVMs, and other such transformers. In recurrent networks, reservoir computing, and beyond, they are searching for ways to augment or replace today’s models. They want to make possible for AI models what comes naturally to biological brains.
Enter brain research
The TBC founders love to point out that evolution has already solved these problems somehow, and encoded the solutions in the biology of our brains. So neurological research might be a good place to look for AI solutions. At least that was the thinking that motivated the substantial team of neurological researchers and AI developers at TBC to move into the undefined space between the brain and the software AI model.
But what exactly are they doing?
The company’s original work was to research brain organization and how networks of neurons actually process information, at the micro level. To do this, they grew actual neuron cell cultures, obtained from rat brains and by growing inducible stem cells into neurons, on top of an array of thousands of electrodes exposed on the surface of an IC. Using the electrodes, the researchers could both stimulate the neurons and observe their responses. In principle, such experiments, and a lot of patience, should reveal something about how brain activity is organized at the cellular level, and how that organization results in learning.
The culture on the IC is not a miniature brain, the TBC founders explained. Nor is it a predesigned neural network with a defined topology. TBC starts out with a mixed cortical culture—a small subset of the thousands of identified brain-neuron subtypes. By tuning the culture, stimulating the neurons with a chosen pattern, observing the resulting neural activity, and then stimulating again with a reinforcing pattern, the researchers are able to direct the neurons to self-organize into a network that responds to a particular type of pattern.
This process is not deterministic. Every culture organizes itself in its own way in response to this training. TBC has learned that by manipulating the source and maturity of the cells in the culture and how densely they are cultured, the researchers can tune the culture to a particular category of problem. Then in response to the training stimuli, the neurons self-organize into distributed recurrent networks adapted to the problem. What TBC discovered is that if you repeat this process often enough with different cultures, stable patterns emerge, not at the cellular level, but at the aggregate level.
Every culture responds differently. But by observing the responses of the neurons across many cultures, TBC found relatively stable patterns in the intensity, persistence, and spread of neuron activity. These patterns reflected the aggregated activity of individual neurons, but they were relatively independent of the exact topology of the individual neural network. It was rather like observing stable patterns in the flow of traffic through Manhattan even though the trajectories of individual cars varies enormously.
Culture to software model
The emergence of these patterns raised an interesting question. If the patterns represented the way cortical neurons learned, could that kind of organization be useful as a layer inside of an artificial neural network to improve its behavior? Research in reservoir computing, where high-order, recursive, but randomly organized networks are inserted between layers of conventional network fabric, have shown good results on some problems. So perhaps …
The TBC researchers implemented software models of the abstracted patterns they had observed in the biological neuron cultures. They formatted the code as adapters, and added to each a lightweight, trainable layer to interface the adapter to existing open AI models. Then they tested the modified models on a number of problems, from simple character recognition to generation of long video clips. And the results were gratifying. The new models generally outperformed the original models by significant margins.
But, skeptics might observe, that is not a surprise. When you make a model bigger, and particularly when you insert something suspiciously like a reservoir into it, you would expect it to improve somewhat. Was the biological research actually contributing anything?
TBC researched this question by comparing the biologically-informed models against reservoir-based and conventionally fine-tuned models. In static image reconstruction tasks, the models with biology-derived adapters outperformed parameter-matched frozen reservoirs. On video generation, models with biology-based adapters were not only faster and less expensive than conventional models, but they did better at achieving long-horizon stability of the generated video than did conventional models subjected to parameter-matched fine tuning and low-rank adaptation (LoRA). So yes, it appeared that the cortical neurons were onto something.
Moving forward
In addition to productizing their adapters for specific AI open models in specific applications, TBC intends to continue using their live-neuron research as a discovery mechanism for new adapters. And perhaps whole new machine-learning architectures will emerge, reducing the vast resource requirements of today’s transformers. Those eons of evolution have stored up a treasure chest of experience on how to learn and act in the real world, in real time. The lid may be open a crack, but much more must wait inside.
Editor’s note: Ron Wilson, a former colleague of Junko Yoshida when they both worked at EE Times, is today an independent technology analyst, writer and contributor to both Junko’s Tech Probe on Substack and Junko’s Talk to Us YouTube channel.
Related stories:
Brain-Derived Computing Zeros in on AI Software
Brain-inspired computing is a long-established engineering field, going as far back as the 1940s. The idea of emulating the human brain’s architecture, or at least the actions of individual neurons, has attracted neuroscie…
Updates: TBC Releases Proof of Concept for Neurally-Optimized AI Software
Click here for demo.
TBC’s press release on its neurally-optimized AI software.
