Hacker News

points by traverseda 1 day ago | 0 comments

I'd be very interested in how a scratch trained model that started training with this neural circuitry would behave. Especially the storage space.

[–] horacemorace 1 day ago

Along these lines I’m curious how a scratch trained model with a few built-in tools (memory, calculator, ?) available during a single generation would behave. Not tool calls as turns.

[–] low_tech_punk 1 day ago

a single generation would still allow looping in the "CPU" because the instruction set can do it. The model is almost a Turing machine (except for the limit on memory)

[–] qwertyforce 1 day ago

As I understand, the problem of continual learning is that if you continue training, it will "overwrite" previous knowledge and this CPU will be lost

[–] im3w1l 1 day ago

There is nothing in theory preventing you from training while keeping a subset of weights locked.

[–] LoganDark 1 day ago

How do you ensure that subset of the weights will continue to be accessible through the others?

[–] hnewsnemo 19 hours ago

Not a holistic answer but one example: There are specific fine-tuning techniques like Orthogonal Subspace Finetuning (https://huggingface.co/docs/peft/package_reference/osf) that target 'unused' parts of the weights so you can keep highly informative parts of the weights as-is.

[–] im3w1l 1 day ago

I don't understand your question.

[–] LoganDark 21 hours ago

Assuming for simplicity that you're talking about locking layers. Each layer influences the input to the next layer, so locking a subset of layers doesn't necessarily mean that you'll continue to see the same results in the output. Not sure if you meant something more granular or otherwise, but I don't see how it's possible to simply freeze a part of a model in order to retain a particular capability.

[–] im3w1l 16 hours ago

I meant taking all the weights that are part of the cpu and locking those, and only those.

[–] LoganDark 16 hours ago

Ah wait, you mean you'd intend the other layers to be able to evolve to use the CPU internally, so all you need is to keep the CPU functional somewhere in the model? That may work, dunno.