Brainoware in numbers
From the 2023 paper and the reports you'll find in the sources below.
What the team built
The paper came out in Nature Electronics in December 2023, from Hongwei Cai and colleagues in Feng Guo's group at Indiana University Bloomington. They grew a brain organoid from human stem cells and sat it on a high-density multielectrode array, so they could send electrical pulses in and record what the cells did back (you can think of it as a tiny lab-grown chunk of brain tissue on a chip). They called the whole set-up Brainoware.
What I find interesting is that they didn't try to teach the organoid a task directly, which is roughly what the DishBrain team did with Pong. They used it as the reservoir in a reservoir computer instead, and I'll explain what that means if you haven't come across it.
Reservoir computing, briefly
If you haven't met reservoir computing before, I'd think of it in three parts. You turn the input into signals and feed them into a messy network (the reservoir), which reacts in a nonlinear way and remembers what came in a moment ago, though the memory fades. Then a simple readout learns to turn the state of that network into an answer. As far as I know you usually only train the readout and leave the reservoir alone, which is why people have tried all sorts of physical things as reservoirs, from circuits to buckets of water.
(There's a longer explanation, with where the idea came from, in what is reservoir computing.) In Brainoware the organoid is the reservoir. The paper goes a bit further, though. It says the pulses gave the organoid the nonlinear behaviour and fading memory you'd want, and that the organoid rewired some of its own connections while it was being trained (they call that unsupervised learning). I'd be careful with that last claim, since it's one study and I haven't seen it repeated.
The speech test
The team took 240 audio clips of 8 adult Japanese men speaking vowels and turned them into patterns of electrical pulses. The task was to say which of the 8 men was speaking from a single vowel. After two days of training Brainoware got it right 78% of the time, which you'd compare with 12.5% for a random guess.
They also had it predict a Hénon map, a standard chaotic equation that people use to test this kind of system, and it did that after a fairly short training time.
What it doesn't show
I think you should read all this as a proof of concept, and that's how Guo put it himself: "This is just proof-of-concept to show we can do the job." A normal AI system on an ordinary computer did better at the same speech task. Also, the organoid needed a CO2 incubator and other kit to stay alive, so I wouldn't use this paper to say living cells save energy (the brain versus GPU guide has what that comparison can and can't tell you). If you're asking whether the organoid felt anything, see are brain organoids conscious.
Organoids don't last for ever either. You'll find the survival figures in how long brain organoids live, and I'd say a reservoir that dies after a few months is a real problem if you want something that runs for years. If you want the wider field this sits in, have a look at organoid intelligence, and for the silicon route (chips that copy neurons) there's neuromorphic computing vs wetware.
Questions
What is Brainoware?
It's a system from Indiana University, published in Nature Electronics in December 2023, where a human brain organoid on a multielectrode array is used as the reservoir in a reservoir computer. Pulses go in, the organoid's response is recorded, and a simple readout is trained on it.
How accurate was Brainoware at speech recognition?
After two days of training it identified which of 8 Japanese speakers was saying a single vowel 78% of the time, from a set of 240 clips. A conventional AI system did better on the same task.
What is reservoir computing?
It's a way of computing where inputs are fed into a fixed network that reacts nonlinearly and remembers recent inputs for a short time, and only a simple readout layer is trained. In Brainoware that network is a living brain organoid.
Sources
- Cai et al., Brain organoid reservoir computing for artificial intelligence, Nature Electronics, 2023
- A 'brain organoid' biochip displayed serious voice recognition and math skills, Popular Science, 12 December 2023
- Cyborg computer combining AI and human brain cells really works, Freethink, 2024
- Lancaster et al., Cerebral organoids model human brain development and microcephaly, Nature, 2013