Perception seems to be one of the main constraints on LLMs that not much progress has been made on. Perhaps not surprising, given perception is something evolution has worked on since the inception of life itself. Likely much, much more expensive computationally than it receives credit for.
I strongly suspect it's a tokenization problem. Text and symbols fit nicely in tokens, but having something like a single "dog leg" token is a tough problem to solve.
The neural network in the retina actually pre-processes visual information into something akin to "tokens". Basic shapes that are probably somewhat evolutionarily preserved. I wonder if we could somehow mimic those for tokenization purposes. Most likely there's someone out there already trying.
AFAIK this is actually a separate mechanism, which is part of the visual cortex and not the retina. Essentially recognizing even a single object requires the complete attention of pretty much your entire brain in the moment of recognition.
What I am referring to is a much more basic form of shape recognition that goes on at the level of the neural networks in the retina.
I think in this case, tokenization and percpetion are somewhat analogous. I think it is probably the case our current tokenization schemes are really simplistic compared to what nature is working with. If you allow the analogy.
Why should it have to be expensive computationally? How do brains do it with such a low amount of energy? I think catching the brain abilities even of a bug might be very hard, but that does not mean that there isn't a way to do it with little computational power. It requires having the correct structures/models/algorithms or whatever is the precise jargon.
> How do brains do it with such a low amount of energy?
Physical analog chemical circuits whose physical structure directly is the network, and use chemistry/physics directly for the computations. For example, a sum is usually represented as the number of physical ions present within a space, not some ALU that takes in two binary numbers, each with some large number of bits, requiring shifting electrons to and from buckets, with a bunch of clocked logic operations.
There are a few companies working on more "direct" implementations of inference, like Etched AI [1] and IBM [2], for massive power savings.
This is the million dollar question. I'm not qualified to answer it, and I don't really think anyone out there has the answer yet.
My armchair take would be that watt usage probably isn't a good proxy for computational complexity in biological systems. A good piece of evidence for this is from the C. elegans research that has found that the configuration of ions within a neuron--not just the electrical charge on the membrane--record computationally-relevant information about a stimulus. There are probably many more hacks like this that allow the brain to handle enormous complexity without it showing up in our measurements of its power consumption.
My armchair is equally comfy, and I have an actual paper to point to:
Jaxley: Differentiable simulation enables large-scale training of detailed biophysical models of neural dynamics [1]
They basically created sofware to simulate real neurons and ran some realistic models to replicate typical AI learning tasks:
"The model had nine different channels in the apical and basal dendrite, the soma, and the axon [39], with a total of 19 free parameters, including maximal channel conductances and dynamics of the calcium pumps."
So yeah, real neurons are a bit more complex then ReLU or Sigmoid.
My whole point is that it maybe possible to do perception using a lot of computational power, or alternatively, there could be another kind of smart ideas that allows to do it in a diferent way with much less computation. It is not clear it requires it.
There could definitely be a chance. I was just responding to what in your comment sounded like a question.
That said, I think there is a good reason to be skeptical that it is a good chance. The consistent trend of finding higher complexity than expected in biological intelligences (like in C. Elegans), combined with the fact that the physical nature of digital architectures versus biological architectures are very different, is a good reason to bet on it being really complex to emulate with our current computing systems.
Obviously there is a way to do it physically--biological systems are physical after all--but we just don't understand enough to have the grounds to say it is "likely" doable digitally. Stuff like the Universal Approximation Theorem implies that in theory it may be possible, but that doesn't say anything about whether it is feasible. Same thing with Turing completeness too. All that these theorems say is our digital hardware can emulate anything that is a step-by-step process (computation), but not how challenging it is to emulate it or even that it is realistic to do so. It could turn out that something like human mind emulation is possible but it would take longer than the age of the universe to do it. Far simpler problems turn out to have similar issues (like calculating the optimal Go move without heuristics).
This is all to say that there could be plenty of smart ideas out there that break our current understandings in all sorts of ways. Which way the cards will land isn't really predictable, so all we can do is point to things that suggest skepticism, in one direction or another.
Following the trend of discovering smaller and smaller phenomena that our brains use for processing, it would not be surprising if we eventually find that our brains are very nearly "room temperature" quantum computers.
