1/ Here's a critical problem that the #neuroai field is going to have to contend with: Increasingly, it looks like neural networks converge on the same representational structures - regardless of their specific losses and architectures - as long as they're big and trained on real world data. 🧠📈 🧪

If we're lucky, we'll learn how to unknot our own biases as a side benefit. I don't think we'll succeed in this without adopting a direct neuro-anatomy to programmatic alignment. (Even then, that xenophobia is kinda baked in yet...).

An exciting problem to think about. It seems to suggest that the networks are gravitating towards a universal understanding of reality. Is this behavior indicating a natural emergence of Occam's razor principles at play, guiding the networks towards the simplest, most efficient representations?

2/ Ex 1: Platonic representations Whether you train on vision or langauge, big nets have aligned representations with similar structures arxiv.org/abs/2405.07987 So, networks may be extracting some "ideal" representation of reality, but then, how can we distinguish our models with brain data?

The Platonic Representation Hypothesis

We argue that representations in AI models, particularly deep networks, are converging. First, we survey many examples of convergence in the literature: over time and across multiple domains, the...

These are fantastic ideas to think about. Some thing that I think about a lot is whether the issue is that our mathematical frameworks are too similar, & our computational frameworks are not sufficiently varied. In the absence of biological anchoring Many aspects of the maths are translatable.

Isn’t this just a reflection that the data available in the world contains statistical regularities?