Guided by millions of years of evolution, our own brain cells know a thing or two about moving through a chaotic world where learning the right move is a matter of life and death.
Biohybrid computing is an emerging field of technology that combines living neural tissue with conventional electronics and software. Rather than simply imitating the brain with silicon chips, the approach uses actual neurons to perform computational work.
The potential is considerable. Biological neural networks can adapt quickly, processing information in parallel using little energy. Harnessing these capabilities could eventually help computers and robots learn from experience without the enormous computational resources demanded by other computational systems.
But turning living cells into reliable computational components remains a serious engineering challenge.
In a study published this month in Frontiers in Artificial Intelligence, researchers at the University of Bristol investigated whether tiny, lab-grown models of the human forebrain known as organoids could process tactile information represented in Braille.
The millimeter-wide arrangements of around 10,000 nerves and supporting cells sat on microelectrode arrays that let researchers stimulate the tissue and record its activity. A tactile sensor scanned Braille characters and converted the measurements into electrical stimulation patterns. The sensor sent signals to the organoids, and researchers then analyzed their responses using a conventional machine-learning algorithm.
The algorithm could identify the correct letter with 61% accuracy based on a single organoid’s response. Combining the processing of three organoids increased the system’s accuracy to an impressive 83%. The combined approach also proved more robust against noise that the team introduced artificially.
The findings suggest living tissue can transform tactile information in ways that preserve useful patterns.
The organoids weren’t taught to recognize Braille through experience. They used a fixed method to encode the letters and a computer algorithm to classify neural responses. Yet other researchers are discovering that biohybrid systems have significant potential when it comes to learning through stimuli.
Melbourne-based biological computing company Cortical Labs, working with Honda Research Institute Europe, recently showed that cultured neurons can improve their performance through interactions with a simulated environment.
Biological computing faces a major obstacle: neurons don’t respond to electrical inputs as predictably as conventional digital hardware. This makes sense, given that our meat circuits are messy things that depend on complex, changing biological dynamics. The Embodied Neurocomputation Framework treats this as an engineering problem, systematically optimizing information entering the neural network.
Research to be presented at this year’s NeurIPS conference in Sydney, Australia, demonstrates the benefits of a process for making biological computing more systematic known as an Embodied Neurocomputation Framework.
Using Cortical Labs' CL1 biological computers, the collaboration connected living neural networks to a virtual agent navigating a grid-like world.
The agent had to move towards a target guided by a simulated odour gradient, choosing whether to move forward, turn left, or turn right to find a goal of “food”. The researchers tested approximately 1,300 configurations of the electrical signals used to communicate with the neurons, accumulating more than 4,000 hours of real-time interaction across dozens of biological computing units.
In the end, 12 configurations consistently demonstrated a suitable level of learning across multiple episodes. Under comparable interactions, this handful of optimized biological agents outperformed conventional reinforcement-learning benchmarks in navigation.
So why bother using biology when silicon doesn’t even require such optimization?
The appeal of biohybrid computing lies partly in our brain's extraordinary energy efficiency. Your gray matter performs astonishing feats of information processing using little more than the equivalent of a relatively shiny light globe.
Neurons also adapt their connections and activity in response to experience. This plasticity provides an alternative pathway to learning, one that differs from the approaches used by neural networks that rely on extensive number crunching.
If researchers can reliably take advantage of these characteristics, biological computing could one day be the prime choice in applications where systems need to adapt to unfamiliar conditions in a field far from home.
In a separate demonstration by Cortical Labs, the frequency at which cells in a biological neural network fired was used to direct a robot dog's movements in a small space.
Bio-computing guides robot dog
While a waltzing dog falls a little short of a canine unit befitting Robocop, the neural choreography represents a critical link between adaptable wetware and the programmable actuators, sensors, and gears we increasingly pin our hopes on for reliable, real-world drudgery.
“Robots that are becoming remarkably capable machines at their physical control still struggle with the higher-order control required to learn quickly or adapt when conditions change,” says Brett Kagan, co-founder and CEO of Cortical Labs.
“Walking, running, and doing acrobatics is amazing, but determining where and when to do actions remains a challenge with traditional methods.”
It's far from the first example of neurology being used to drive the movements of a robot. Yet every experiment brings us closer to a system that can learn about novel changes in its surroundings and make informed choices on what action to take next.
Despite the progress, none of these demonstrations suggest that biological computers will soon serve in Amazon warehouses or chauffeur you around town in a Waymo. Keeping the tissue alive is one problem yet to be solved. Cultured brain cells require carefully controlled conditions and a continuing supply of nutrients. Though energy efficient themselves, all of that supporting hardware chews up energy. The total energy cost of a practical system should be assessed across the entire process, not just the neurons.
Still, the field is moving beyond whimsical demonstrations of organoids playing computer games. Companies are developing frameworks for measuring performance, optimizing interfaces, and testing whether biological computation can deliver consistent results across tasks.
The goal is not to build a computer made entirely of brain tissue. It is to discover whether living neural networks truly can complement silicon. If they succeed, biohybrid computing could offer a way to build systems that learn and adapt efficiently in places our own brains can’t be present.