Applying The Thousand Brains Theory in AI

It's exciting to see the kick-off of the Thousand Brains project at the Stanford Institute for Human-Centered AI (HAI).

This reflects what we observe in the trends in the LLMs, LMMs: The representations in AI models, especially deep networks, are becoming more similar over time and across different domains. The methods by which various neural networks encode data are getting more consistent. And this similarity is evident across data modalities: as vision models and language models grow larger, they compute distance between datapoints in an increasingly similar way. We propose that this similarity is moving toward a shared statistical model of reality, similar to Plato’s idea of a perfect reality - the Platonic Representation Hypothesis.

The discovery and the implications of these trends give us a stronger sense for continuing and increasing the effort on the Neuroscience and Computer science joint approach, which ultimately reveal how our brain models the world and apply that to our Deep learning frameworks rather than just depend on the transformer approach with the scaling law.

Please refer to my previous post on the One Thousand Brains and Numenta’s announcement.

#neuroscience #ai #artificialintelligence #deeplearning #genericworldmodel