Future robots will be powered by human brain cells. Companies will use them as conscious slaves and they'll get around slavery laws by saying they're not human.
Personally, dislike this direction a lot. I don't like that they're using a killing game (I understand the trope, doesn't make me like it any less) and the general idea of this whole thing makes me quite uneasy.
> The neurons serve as a biological filter: the training system translates screen pixels and ray-cast distances into electrical zaps, the living cells fire spikes, and those counts feed straight into a PyTorch decoder that maps them to Doom actions. The PPO agent, CNN encoder and entire reward loop run on ordinary silicon elsewhere. Cole’s ablation modes make the split testable, set decoder output to random or zero and the game still plays. The CL1 hardware interface works exactly as advertised. What remains unproven is whether 200,000 human neurons can ever carry the policy instead of just riding along.
Yeah… That’s quite the smoking gun.
So it’s quite likely then that the neurons are just acting as a bad conductor. The electrodes read a noisy version of the signals that go into the neurons, and they just train a CNN with PPO to remove that noise, get the proper inputs, and learn a half-decent policy for playing the game.
If this worked as advertised they shouldn’t need a CNN decoder at all! The raw neuron readout should be interpreted as game inputs directly.
Besides, they are not streaming the video into the neurons at all. Just the horizontal position of the enemies and the distance, or some variant of that. In that sense it’s barely more than pong isn’t it? If enemy left, rotate left, if enemy right, rotate right, if enemy center shoot. At a stretch, if enemy far, go forward, if enemy close, go back. The rest of the time just move randomly. Indeed, the behavior in the video is essentially that…
While we are at it, the encoded input signal itself is already pretty close to a decent policy if mapped directly to the keys (how much enemy left, center, right), even without any CNN, PPO or neurons.
EDIT: It seems like the readme does address these concerns, and the described setup differs significantly from the description in the critical blogpost. Still not entirely convincing to me, a lot of weights being trained in silicon around the neurons, but it sounds better. I don’t have time right now to look deeper into it. They outline some interesting details though.
No, this is precisely why there are ablations. The footage you see in the video was taken using a 0-bias full linear readout decoder, meaning that the action selected is a linear function of the output spikes from the CL1; the CL1 is doing the learning. There is a noticeable difference when using the ablation (both random and 0 spikes result in zero learning) versus actual CL1 spikes.
Isn't the encoder/PPO doing all the learning?
This question largely assumes that the cells are static, which is incorrect; it is not a memory-less feed X in get Y machine. Both the policy and the cells are dynamical systems; biological neurons have an internal state (membrane potential, synaptic weights, adaptation currents). The same stimulation delivered at different points in training will produce different spike patterns, because the neurons have been conditioned by prior feedback. During testing, we froze encoder weights and still observed improvements in the reward.
How is DOOM converted to electrical signals?
We train an encoder in our PPO policy that dictates the stimulation pattern (frequency, amplitude, pulses, and even which channels to stimulate). Because the CL1 spikes are non-differentiable, the encoder is trained through PPO policy gradients using the log-likelihood trick (REINFORCE-style), i.e., by including the encoder’s sampled stimulation log-probs in the PPO objective rather than backpropagating through spikes.
Be sure to dig into the details before taking this at face value. There once was a story "Rat brain flies plane" a couple decades ago, and it turned out to be bogus. But to find that out, you had to read the paper and reverse engineer that nothing substantial was actually going on. It's tempting to be charitable, but you can't really know whether headlines like this are legit till you understand exactly what they did.
(The rat brain guys repeated the experiment until the plane stopped crashing, but no "learning" was happening; it was expected that when the neuron's range reached so-and-so, that the plane would fly level. So they started with a neuron outside that range, showed that it crashed, then adjusted the neuron until it flew level. But that's not what "rat brain flies plane" implies.)
It’s the first time I’ve heard about this company, and of course I haven’t taken the time to check how real their product is, but honestly, for me it’s very difficult to believe we currently have the technology to correctly integrate a living neuron into a chip, let alone compute anything meaningful with it.
From what I’ve read elsewhere, our understanding of neurons is still very basic, and we need a lot more fundamental research before reaching results like these. We still don’t even properly know how migraines work, nor can we cure paraplegia, yet somehow we supposedly have the capacity to grow second brains and program them on top of that.
From their video it just comes across as they stimulate different left/right neurons depending on where the enemy is on screen and then listen to some output that also says left/right. Shooting looks completely random, to be frank.
If you connected electrodes to two different fish, shocked them and interpreted twitching as intelligent output, fish could also play Doom. The interface is doing all the work.
It doesn't sound like the neurons have any concept of the game other than "left input means left output", which is a rather trivial result... It's effectively no different than the pong example.
They don't say anything on how much training is required for this to happen, or if there's any "learning" going on at all. The learning part is "next".
Hard to tell If the neurons actually learned to play doom or if its just the decoder that learned from the neuron responses. The disease modeling for this system is a very cool usecase though.
There's a number of "immortal" human cell lines dating back to as far as the 1950s (you may have heard of Henrietta Lacks? [1] and the immortal HeLa cell line).
Today there are several
immortalized neuron cell lines used in research to model neuronal function, like HeLa but of neuron type obviously, that are also typically derived from tumours (e.g., SH-SY5Y, PC12) or immortalized via genetic modification (e.g., v-myc) like CTX0E03 [2] which was designed to allow for continuous growth in the presence of particular reagents.
Is there a reason they're using human brain cells specifically? This seems like it would also work with neurons from other creatures.
I was under the impression that the relative intelligence of humans versus other animals was largely a function of brain cell quantity, not quality. Can 200k human brain cells really learn faster than 200k mouse brain cells?
A more cynical take is that they're just using human brain cells for shock value. They chose DOOM because of the "can it run DOOM" meme, so they clearly value publicity a lot.
If this can be taken at face value... it's creepy.
I get that they're doing it for the meme. But perhaps something getting close to human intelligence, made out of human cells, shouldn't be forced to play a violent video game without any alternative options? Does 'the meme' justify that?
I dunno. Nothing against violent games myself. Just feels like it's starting to get quite questionable, ethically speaking.
This seems very very far fetched. If I understand correctly, these cell brains just respond to some stimuli, it does not seem more intelligent than any automat to me, just creepier.
74 comments
[ 2.4 ms ] story [ 92.9 ms ] threadThere is a line somewhere here that I personally feel we should not cross.
If they can get Doom to run on a pregnancy test surely they could get it running on human brain cells?
Personally, dislike this direction a lot. I don't like that they're using a killing game (I understand the trope, doesn't make me like it any less) and the general idea of this whole thing makes me quite uneasy.
Yeah… That’s quite the smoking gun.
So it’s quite likely then that the neurons are just acting as a bad conductor. The electrodes read a noisy version of the signals that go into the neurons, and they just train a CNN with PPO to remove that noise, get the proper inputs, and learn a half-decent policy for playing the game.
If this worked as advertised they shouldn’t need a CNN decoder at all! The raw neuron readout should be interpreted as game inputs directly.
Besides, they are not streaming the video into the neurons at all. Just the horizontal position of the enemies and the distance, or some variant of that. In that sense it’s barely more than pong isn’t it? If enemy left, rotate left, if enemy right, rotate right, if enemy center shoot. At a stretch, if enemy far, go forward, if enemy close, go back. The rest of the time just move randomly. Indeed, the behavior in the video is essentially that…
While we are at it, the encoded input signal itself is already pretty close to a decent policy if mapped directly to the keys (how much enemy left, center, right), even without any CNN, PPO or neurons.
EDIT: It seems like the readme does address these concerns, and the described setup differs significantly from the description in the critical blogpost. Still not entirely convincing to me, a lot of weights being trained in silicon around the neurons, but it sounds better. I don’t have time right now to look deeper into it. They outline some interesting details though.
> Quote from: https://raw.githubusercontent.com/SeanCole02/doom-neuron/mai...
Isn't the decoder/PPO doing all the learning?
No, this is precisely why there are ablations. The footage you see in the video was taken using a 0-bias full linear readout decoder, meaning that the action selected is a linear function of the output spikes from the CL1; the CL1 is doing the learning. There is a noticeable difference when using the ablation (both random and 0 spikes result in zero learning) versus actual CL1 spikes.
Isn't the encoder/PPO doing all the learning?
This question largely assumes that the cells are static, which is incorrect; it is not a memory-less feed X in get Y machine. Both the policy and the cells are dynamical systems; biological neurons have an internal state (membrane potential, synaptic weights, adaptation currents). The same stimulation delivered at different points in training will produce different spike patterns, because the neurons have been conditioned by prior feedback. During testing, we froze encoder weights and still observed improvements in the reward.
How is DOOM converted to electrical signals?
We train an encoder in our PPO policy that dictates the stimulation pattern (frequency, amplitude, pulses, and even which channels to stimulate). Because the CL1 spikes are non-differentiable, the encoder is trained through PPO policy gradients using the log-likelihood trick (REINFORCE-style), i.e., by including the encoder’s sampled stimulation log-probs in the PPO objective rather than backpropagating through spikes.
(The rat brain guys repeated the experiment until the plane stopped crashing, but no "learning" was happening; it was expected that when the neuron's range reached so-and-so, that the plane would fly level. So they started with a neuron outside that range, showed that it crashed, then adjusted the neuron until it flew level. But that's not what "rat brain flies plane" implies.)
What would be surprising is for dead human cells to play anything at all.
From what I’ve read elsewhere, our understanding of neurons is still very basic, and we need a lot more fundamental research before reaching results like these. We still don’t even properly know how migraines work, nor can we cure paraplegia, yet somehow we supposedly have the capacity to grow second brains and program them on top of that.
If you connected electrodes to two different fish, shocked them and interpreted twitching as intelligent output, fish could also play Doom. The interface is doing all the work.
It doesn't sound like the neurons have any concept of the game other than "left input means left output", which is a rather trivial result... It's effectively no different than the pong example.
They don't say anything on how much training is required for this to happen, or if there's any "learning" going on at all. The learning part is "next".
https://corticallabs.com/research
Does anyone have insight into how you would even start to source or grow/create the cells?
Also the machines look very organic and clearly have to keep the cells alive. Do they have to change them out every so often?
Today there are several immortalized neuron cell lines used in research to model neuronal function, like HeLa but of neuron type obviously, that are also typically derived from tumours (e.g., SH-SY5Y, PC12) or immortalized via genetic modification (e.g., v-myc) like CTX0E03 [2] which was designed to allow for continuous growth in the presence of particular reagents.
1. https://en.wikipedia.org/wiki/Henrietta_Lacks
2. https://www.prnewswire.com/news-releases/reneuron-announces-...
If you're in the US, you can buy human neurons online at sciencellonline.com/en/human-neurons/
I was under the impression that the relative intelligence of humans versus other animals was largely a function of brain cell quantity, not quality. Can 200k human brain cells really learn faster than 200k mouse brain cells?
A more cynical take is that they're just using human brain cells for shock value. They chose DOOM because of the "can it run DOOM" meme, so they clearly value publicity a lot.
I get that they're doing it for the meme. But perhaps something getting close to human intelligence, made out of human cells, shouldn't be forced to play a violent video game without any alternative options? Does 'the meme' justify that?
I dunno. Nothing against violent games myself. Just feels like it's starting to get quite questionable, ethically speaking.
This seems very very far fetched. If I understand correctly, these cell brains just respond to some stimuli, it does not seem more intelligent than any automat to me, just creepier.