how these brain cells in a dish learned to play pong!

@nasjaq

921.0K followers · usually 75.8K views a post

Views

420.2K

Likes

44.2K

Comments

614

Shares

1.7K

SoundTikTok sound · original sound - nasjaq

Promoted product

Cortical Labs

We've combined lab-grown neurons with silicon chips and made it available to anyone, for first time ever.

Why it worked

The video likely performed well due to its novel and intriguing concept of training brain cells to play a video game, tapping into curiosity about AI, neuroscience, and the future of computing. The use of clear visuals and explanations, combined with a relatable game like Pong, makes the complex topic accessible and engaging.

Summary

This video explains how brain cells cultured in a dish were trained to play the game Pong using a specialized computer chip. The experiment demonstrates that these biological neural networks can learn and adapt, offering insights into potential future applications in AI and robotics.

Structure

1Introduction to brain cells playing Pong
2Explanation of the technology and training process
3Discussion of the free-energy principle
4Comparison to traditional neural networks

Original caption

how these brain cells in a dish learned to play pong!

More from @nasjaq

Product placement

Cortical Labs' technology is showcased as the subject of the video, demonstrating how their chip and cultured brain cells can learn to play Pong.

Call to action

If you want to learn more, I just posted a podcast with the founder.

On-screen text

GAME VISUALISER HD-MEA Chip Data In/Out In vitro neurons on silicon chip Simulated Environment: Pong The free-energy principle Sensory states (perception) External states Markov blanket brain Internal states Active states (action) F = Energy - Entropy = - (ln p(s,n) + (ln q(n))g Action to minimise a bound on surprise F = Complexity - Accuracy D(q || p(s|q)) - (ln p(s|q))g a = argmax Accuracy Perception to optimise the bound F = Divergence + Surprise D(q || p(s|q)) + (ln p(s|q))g μ = argmin Divergence STIM 1 Low latency closed-loop feedback system Unpredictable random sensory input ball returns Predictable sensory feedback 2 SIL It hurt itself in its confusion? MaxOne 4252 CORTICAL LABS

Transcript

So these brain cells in a dish learned to play pong and supposedly they learned faster than a traditional neural net. So how does this work? It works by slathering brain cells on top of a computer chip which then acts as a communication layer between the brain cells and the software of the pong game. Then they train the brain cells using something called the free-energy principle. Basically, brain cells don't like environments where they don't understand something, so they work to minimize the random noise. This taught them how to play pong. It turns out they learn more efficiently than traditional neural nets, but the neural nets were just learning in real time, so it wasn't technically faster. Now, the startup behind this envisions racks of these brain cells linked together in the cloud, training on data similar to Amazon Web Services. And even crazier ideas are that since brains learn by movement in their environment, there's even talk about connecting these brain cells to like a Tesla robot to see if they could learn better. Anyways, if you want to learn more, I just posted a podcast with the founder.