Comment by toisanji | Hacker News Reader

Comment by toisanji | Hacker News Reader

I've written about the cognitive aspect the author talks about before:

https://blog.jtoy.net/examining-concepts-through-different-f...

https://blog.jtoy.net/on-the-wondrous-human-simulation-engin...

https://arxiv.org/abs/2210.12068

The human brain has at least over 12 different coordinate systems. I like to reference this image: https://cln.sh/gpqhwrw2Hg92MQz62TY6

Humans have the ability to inspect anything from any point of view that we want, like a real time video game simulation system in our minds. I suspect that artificial neural networks are missing this core ability to encode and translate its point of view when processing data. Just look at basic LLMs failing at the reversal curse: https://arxiv.org/abs/2309.12288 No amount of more GPUs and more training data will fix it.

The entorhinal cortex in humans has a special kind of neuron called grid cells that works with hippocampus cells to do generalized coordinate translations. The discovery of grid cells won a nobel peace prize ( https://www.nobelprize.org/prizes/medicine/2014/advanced-inf... ). It is a prime suspect to incorporate for a coordinate transformation layer in my opinion.

I want to build an ANN that incorporates this "coordinate transformation system", but I'm not sure how to translate it into code, this is my dream project.

The problem is an ANN or LLM has no physical location or “point of view”, which has led ML practitioners toward the concept of embodiment: where algorithms and agents no longer learn from datasets of images, videos or text curated primarily from the internet - instead they learn through interactions with their environments from an egocentric perception similar to humans.

Curiously I didn’t see any mention of this term in your paper. But you did touch on the contrast between egocentric and allocentric representations, which lies at the heart of the issue.

I think the grand challenge of AGI will be unifying these two to reap the benefits of both while minimizing the dangers of imbalance. A bit like physics wrt gravity and quantum mechanics.

We’ll need allocentric perspective for globally optimal decisions (superorganism), but we’ll also need egocentric reasoning to enable robust autonomy in unique local environments.

Also worth mentioning that multimodal representation learning seems like it essentially achieves the sort of cross modality behavior of grid cells, although physical location (not positional encoding) is not an input in any of the research I’m aware of.

Are there groups actively attempting this? I was thinking that this type of data gathering is a simple solution to LLMs “running out of data” to train on. Just put a camera in a room and let the model learn by exploring its environment.

You’re looking for group equivariance - there is some work in group equivariant neural networks. You might find the very recent paper on Clifford Group-Equivariant networks interesting - it applies to not only Euclidean spacetime transformations but even Minkowskian spacetime.

You might also be interested in implementations for novel view synthesis such as instantngp and other NERF-type models. InstantNGP is actually ludicrously simple relative to most recent neural architectures, imo.

With respect to the reversal problem, I wonder how strongly this would still hold for eg encoder-decoder (eg T5) or “order-free autoregressive” models (eg sigma-GPT from last week)

Positional embeddings used in transformers are actually suspiciously similar to grid cells. They allow for a very similar kind of coordinate transform.

You might enjoy this paper[0] which shows that recurrent position encodings recover grid cell representations and maps to path integration found in a popular model of the hippocampus. This isn't terribly surprising since RNNs have shown this before[1, 2] but its an interesting connection.

[0] https://arxiv.org/abs/2112.04035

[1] https://arxiv.org/abs/1803.07770

[2] https://www.nature.com/articles/s41586-018-0102-6

Reply on news.ycombinator.com