If ‘DishBrain’ can learn to play Pong – what does this mean for science?

It sounds like the start of a nerdy joke: What can you do with 800,000 brain cells, a computer chip and Pong? If you’re an Australian commercial startup, the answer is apparently ‘create sentience’.

It’s a pretty big claim, and it will need pretty big evidence to confirm, but the team’s system – which they’ve called DishBrain — does seemingly learn how to play a version of the game Pong, which gives us a perfect chance to look into the ethics of systems like these, as well as understand what a ‘brain-on-a-chip’ even is.

“The beautiful and pioneering aspect of this work rests on equipping the neurons with sensations – the feedback — and crucially the ability to act on their world,” says one of the authors, Professor Karl Friston, a theoretical neuroscientist at University College London.

“Remarkably, the cultures learned how to make their world more predictable by acting upon it. This is remarkable because you cannot teach this kind of self-organisation; simply because – unlike a pet – these mini-brains have no sense of reward and punishment,” he says.

Ezgif. Com gif maker
Gif showing ‘Dishbrain’. Credit: Cortical Labs

DishBrain

Let’s say upfront that this is not a brain. Not even close. The human brain has somewhere in the region of 86 billion neurons. Even a mouse has 70 million. The team suggest DishBrain has a similar number of neurons to a bumblebee.

Even if DishBrain did have 86 billion neurons, we just don’t know enough about the human brain to be able to create anything resembling a human brain.

“The brain has a complexity that cannot currently be replicated in a dish. The shear number of cell types, the dynamics of the synaptic input they receive, and the neuromodulatory processes are not easy to replicate,” explains Associate Professor Lucy Palmer, the head of the Neural Networks Laboratory at the Florey Institute.

“Despite this, recording from neurons on a chip can provide extremely valuable insights into neural interactions and how they alter this activity to generate a desired outcome. This study puts a real-world example (playing Pong) on an otherwise hard to clearly illustrate the impact and refinement of neural output.”

So, with all that said, DishBrain just learnt how to play Pong, so it’s probably worth understanding how it works.

The team took either human neurons called hiPSC, or mouse embryo neurons called E15, and put them onto a high-density multi-electrode array – think of this like a silicon chip inside a medium filled petri dish.

The neurons were then set up to play Pong by inducing zaps of different voltages depending on where the ball was in relation to the paddle. Electrodes on the left or right of one array were fired to tell DishBrain which side the ball was on, while distance from the paddle was indicated by the frequency of signals. There was a higher voltage at random locations if DishBrain didn’t hit the ball back.

All of this created a system where the cells started to ‘learn’ how to play the game better over time. To be fair, it’s not going to be beating a human in a competition any time soon – it hit only slightly more than it missed. But it was better than a system that had received stimulus but no feedback.

“I think if you saw a bumblebee manage to play the game of Pong and it could hit more often than it misses, I think you’d be fairly impressed with the bumblebee,” says lead author Dr Brett Kagan, chief scientific officer at Cortical Labs.

“I think that’s the right sort of approach in considering its performance here.”

Just a note on that – honeybees have been shown to be able to do basic maths like adding and subtracting, understand zero and can link symbols to numbers. Playing an average game of Pong is not necessarily out of their purview.

The team has more research to be released. They’re first investigating what happens if you get DishBrain ‘drunk’, but this is likely only the very start in a system which might one day be able to be used as a new way to study the human brain without animal models.

Brain on a chip

DishBrain isn’t the first time that scientists have mounted neurons on arrays and read their activity. Traditionally these are called a ‘brain on a chip‘. The technology for these is still in its infancy but it’s hoped that one day these might be able to model different parts of the brain. You could use these with less ethical issues than an actual human, but still be able to measure, for example, how drugs affect certain areas of the brain.

“After more than 10 years from its first appearance, the term ‘brain on a chip’ can now be appropriately used,” bioengineers Martina Brofiga and Paolo Massobrio wrote in an opinion piece for the journal Frontiers in Neuroscience back in March.

“Nowadays, we are able to replicate many human neuronal types and peculiar brain regions in the form of engineered neuronal cultures, like neurospheroids or brain organoids, directly from embryonic and human induced pluripotent stem cells (hiPSC), and to couple them to a technological counterpart (i.e., chip).”


Read more: Brain organoids ready for real-time observation


Although scientists have been working on these ‘brain on a chip’ technologies for a while, DishBrain is the first time that the cells have been stimulated and given feedback in a structured way.

The team controversially calls this sentience.

Sentience?

“We didn’t use the word ‘sentience’, willy nilly. It was a lot of internal debate, and a lot of external debate as well,” says Kagan.

“We opted for sentience because it was something that was formally quite a correct definition to explain what was going on.”

Sentience – as the team describe it – is the idea of being able to sense information and respond. They stress that that’s not the same as ‘consciousness’, where something can think and be aware of existence.

Although DishBrain can ‘respond’, it is very unlikely to be able to be aware of anything. Think of it more like a knee being hit with a reflex hammer than a thinking brain.

“For me, this is an exciting development. We usually look at neural code, that is, action potentials, and how they alter throughout learning and/or the activation of certain sensory and feedback inputs,” says Palmer.

“However, we typically do not have a sense of how changes in the neural code through learning can refine outcomes to the degree that playing Pong can. Learned behaviour is usually intermingled with lots of fluctuating brain states that can not be controlled. Here, the environment can be controlled in a manner to directly assess consequences of learning.”

Another paper, published by the team back in March, looks into ethical questions of sentience and consciousness further.

But there are more ethical questions than just sentience. The research is happening not in a university but in a start-up lab. The research has been funded by venture capital money, and the results are patented.

Nikond850 8403049 200303 2914
Dr Brett J. Kagan and Chief Executive Officer, Dr Hon Weng (standing), conducting cell work on multielectrode arrays in a biosafety hood. Credit: Cortical Labs

Kagan told me that they’re trying to be as open as possible, setting up a subreddit and a Discord to try and get people involved.

“Even though we’re a start-up and get venture capital money, we’re not sort of competitive or secret. We’re trying to take the opposite approach to be as transparent as possible, which is why we’ve gone to the large effort of trying to get through the whole peer review process and write all this stuff up,” says Kagan.

“We want people to see the level that we’re at, and it’s going to be too large just for one company to work on.”

Despite all this, Cortical Labs’ goal is to create biological computer chips “to create machines with biological intelligence”.

If teaching a bunch of neurons to learn Pong isn’t ethically dicey, it seems unlikely that it will stay that way for long.

Please login to favourite this article.