Pourquoi des neurones humains en réseau gagnent en stabilité face aux variations
Cortical Labs wired 15 CL1 biocomputers into one network called HIVE, and in a grid test the group outscored a conventional Deep Q Network by a wide margin.
By
Jenna Kowalski
In brief
- HIVE links 15 CL1 biological computers powered by human neurons, built by Australia’s Cortical Labs.
- The system was shown at the first International Conference on Organoid Intelligence in Washington, D.C.
- The team reports that the group is more stable than any single culture of neurons.
- Cortical Labs describes the results as preliminary and tied to its own experimental design.
Fifteen lab-grown clusters of human neurons, wired together, beat a conventional AI model at a grid-navigation task. That is the headline result from HIVE, a system built by Cortical Labs, the Australian maker of the CL1 biological computer, as reported by New Atlas.
What HIVE is and where it appeared
HIVE connects 15 CL1 machines, each powered by human neurons. The company presented it at the first International Conference on Organoid Intelligence (ICOI), held in Washington, D.C., on October 26 and 27. New Atlas describes Cortical Labs as the maker of the world’s first biological computer, the CL1.
The premise is simple. One machine driven by a group of brain cells can solve tasks, but teamwork between several machines might lift performance further. The team wanted to find out whether a collective would do better than any unit working alone.
How the tests compare
Two experiments carry the numbers. The first was symbol recognition. The second sent a virtual character across a grid to reach a specific target, and the team measured HIVE against a Deep Q Network (DQN), a conventional AI that relies on algorithms rather than brain cells.
| Grid test setup | Approximate share of maximum score |
|---|---|
| Best single Cortical computer | 40% |
| Several biological computers voting on moves | 70% |
The jump from a single machine to a pooled decision is the part worth noting: voting on moves nearly doubled the best individual result, and specialisation added another step.
Why a group behaves differently
The CL1 machines appeared to split up different aspects of the task and settle on the most efficient route to the end point. The scientists could not teach the neurons to do this. The behaviour emerged once the cultures became part of a larger network.
Stability is the second argument. Living neurons may perform well one day and poorly the next, so combining groups spreads the load. The combined system can also keep running if one group of neurons stops responding for any reason.
Johnson Zhou, a scientist at Cortical Labs, put the variability problem this way: “One culture might perform particularly well on a given day while another does not.” He added: “When cultures work together, those individual differences have less influence over the final result.”
Dr Brett Kagan, co-founder and CEO of Cortical Labs, framed the stakes: “Biological computing has enormous potential, but useful systems cannot depend on a single culture performing perfectly every time.”
Limits and next steps
The comparison also rests on one reported source, with figures supplied by Cortical Labs. The team’s next goals are to work out how neurons signal and process feedback across the connected machines, and to test whether adding more connections could improve results further.
Photo: Anna Shvets / Pexels
