The aim of this paper is to introduce a new learning procedure for neural networks and to demonstrate that it works well enough on a few small problems to be worth further investigation. The Forward-Forward algorithm replaces the forward and backward passes of backpropagation by two forward passes, one with positive (i.e. real) data and the other with negative data which could be generated by the network itself. Each layer has its own objective function which is simply to have high goodness for positive data and low goodness for negative data. The sum of the squared activities in a layer can be used as the goodness but there are many other possibilities, including minus the sum of the squared activities. If the positive and negative passes could be separated in time, the negative passes could be done offline, which would make the learning much simpler in the positive pass and allow video to be pipelined through the network without ever storing activities or stopping to propagate derivatives.
Metadata: arXiv. Abstracts retain their original authors’ rights.
In the conversation
Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels
While there are some similarities, yes, our brains definitely don't have anything akin to backpropagation, which is the critical mechanism for how current AI models learn.
Hinton has some research on a forward-forward learning paradigm [1], which might be closer to how our brains learn (but the artificial implementations are not great yet). He also posits that maybe the purpose of humans' dreams are generating negative data for such a contrastive forward-forward learning mechanism.
[1] https://arxiv.org/abs/2212.13345
Here's what I envision (note: impossible with current state of the art)
A model that can be incrementally trained (this is the bit we're missing) hosted by a nonprofit, belonging to "we the people" (like wikipedia).
The training process could be done a little like wikipedia talk pages are now - datasets are proposed and discussed out in the open and once generally approved, trained into the model.
Because training currently involves backpropagation, this isn't possible. Hinton was working on a structure called "forward-forward" that would have overcome this (if it worked) before he decided humanity couldn't be trusted [1]. It is my hope that someone smarter than me picks…
Read the full comment
up this thread of research - although in the spirit of personal responsibility I've started picking up my old math books to try and get to a point where I grok the implementation enough to experiment myself (I'm not super confident I'm gonna get there but you can't win if you don't play, right?)
It's hard to tell when (if?) we're ever going to have this - if it does happen, it'll be because a lot of people do a lot of really smart unpaid work (after seeing OpenAI do what it did, I don't have a ton of faith that even non-profit orgs have the will or the structure to pull it off. Please prove me wrong.)
1 - https://arxiv.org/abs/2212.13345
sounds cool,
i wonder if its possible to do similar with "The Forward-Forward Algorithm" by Geoff Hilton maybe on a few add on layers like with lora
[0]: https://arxiv.org/pdf/2212.13345.pdf
The article links to an old draft of the paper (it seems that the results in 4.1 couldn't be replicated). The arxiv has a more recent one: https://arxiv.org/abs/2212.13345
> In May 2023, a colonel in the US Air Force revealed a simulation in which an AI-enabled drone, tasked with attacking surface-to-air missiles but acting under the authority of a human operator, attacked the operator instead. (The US Air Force later denied the story.)
This casts a shadow on the reliability of the whole article. While the retraction (they later stated that this was a thought experiment and NOT a simulation) may or may not be true, the current state of AI would suggest that this kind of chain-of-logic problem-solving is likely out of reach. To put this in here without such a caveat and instead use the 'later denied' language is a strong indicator that we're…
Read the full comment
being sold a story, and not the whole one.
The quotes and other horror-scenario errata also come off a little thin. If Hinton really, genuinely believes the world is ending, then why is he working on the generational successor to backpropagation [1]?
Good way to sell 'news', though. Keep 'em scared, keep 'em reading, I guess.
1 - https://arxiv.org/abs/2212.13345
Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task.
Despite their massive size, successful deep artificial neural networks can exhibit a remarkably small difference between training and test performance.
Fast compilation is important when compilation occurs at runtime, such as query compilers in modern database systems and WebAssembly virtual machines in modern browsers.