Breakthrough Technique: Meta-learning for Compositionality
Original :
https://www.nature.com/articles/s41586-023-06668-3
Vulgarization :
https://scitechdaily.com/the-future-of-machine-learning-a-new-breakthrough-technique/
How MLC Works
In exploring the possibility of bolstering compositional learning in neural networks, the researchers created MLC, a novel learning procedure in which a neural network is continuously updated to improve its skills over a series of episodes. In an episode, MLC receives a new word and is asked to use it compositionally—for instance, to take the word “jump” and then create new word combinations, such as “jump twice” or “jump around right twice.” MLC then receives a new episode that features a different word, and so on, each time improving the network’s compositional skills.
While in not in the field either, I do know that it is quite unusual in computer science academics to publish in actual peer reviewed journals. This is because it can be a long process, and the field is very fast moving, so your results would be outdated by the time you publish. Thus, a paper is typically synonymous with a conference proceeding, and can be found on arxiv. I found this Paper on the arxiv from 2017/2018 which seems to be when this paper was originally published for the scientific community and presented at a very “good” (if I had to guess) conference. Google scholar says this paper has 650 citations, so it probably has had quite some impact. However, I would guess this method is well known and is already implemented in many models, if it was truly disruptive.
To be clear, the papers at conferences undergo a peer review process as well. There are journal publications in CS, but a lot of publishing is done through conferences. Arxiv, while a great resource, has little to do with the conferences and it is worth noting that the papers on arxiv do not go through a peer review process (but are often published at conferences where the paper has gone under peer review — some papers on arxiv may be preprint versions from before the peer review process).
For reference, ICML is one of the most prestigious machine learning conferences alongside ICLR and NeurIPS.
Good to know, Thanks.
Yours is the type of comment I was really hoping to read here.You are right : it’s the same authors (Brenden M. Lake & Marco Baroni) with mostly the same content.
But, they also write (in nature) that modern systems (GPT-4) do not yet incorporate these abilities :
Preliminary experiments reported in Supplementary Information 3 suggest that systematicity is still a challenge, or at the very least an open question, even for recent large language models such as GPT-4.
This certainly could be part of the motivation for publishing it this way, to make themselves more noticed by the big players. Btw, publishing in open source nature is expensive, it’s like 6-8000 euro for the big ones, so there definitely is a reason.
Edit : Please read @DigitalMus@feddit.dk’s comment before mine.
Hey folks, I believe this is really big.
Traditional deep neural network’s training requires millions of example and so, despite its great success, is immensely inefficient.
Now what if learning of these machines was as fast or faster than a human’s ? Well, it seems this is it.
Look at how large language models are disruptive for many sectors of society. This new technology could accelerate the process exponentially.
Traditional deep neural network’s training requires millions of example and so, despite its great success, is immensely inefficient.
Is this a limited advancement in training techniques? Right now I’m working on several types of image classification models. How would this be able to help me?
Sorry, I just read a lot… but I don’t work in this field.
If you don’t really understand it, why do you believe this is so big?
I am not sure what “image classification models” incompasses. I would have to read more and understand and I don’t have enough time and energy.
Yet in the past I have read and understand a few books about neural networks and this new article in nature is something else : it’s clear when reading it.
( also to @TropicalDingdong@lemmy.world )I mean is this any different than standard gradient descent with something like Adam as optimiser.
That’s my assumption based on the headline. But the quick skim I gave the article seemed to only discuss it in the context of NLP. Not exactly my field of study.
Based on your reading, swag?
