I fail to understand why we would lack data. Sure, there is limited (historical) text, but if we just open up all available video, and send out interactive robots into the world, we'll drown in data. Then there is simulated data, and tons of sensors that can capture vast amounts of even more data.
Edit: from the source [1], this quote pretty much sums it all up: "Our 2022 paper predicted that high-quality text data would be fully used by 2024, whereas our new results indicate that might not happen until 2028."
This paper was just too overhyped by the authors. Also, the initial evals were very limited and very strange. This blog post does a much better job at a similar observation -- goes into details and does proper evaluation (also better attribution): https://jinjieni.notion.site/Diffusion-Language-Models-are-S...
> This paper addresses the challenge by asking: how can we trade off more compute for less data?
Autoregressive models are not matched by compute and this is the major drawback.
There is evidence that training RNN models that compute several steps with same input and coefficients (but different state) lead to better performance. It was shown in a followup to [1] that performed ablation study.
It has already been proven that deep equilibrium models with a single layer are equivalent to models with a finite number of layers and the converse as well. That you can get the performance of a DEQ using a finite number of layers.
The fixed point nature of DEQs means that they inherently have a concept of self assessment how close they are to the solution. If they are at the solution, they will simply stop changing it. If not, they will keep performing calculations.
7 comments
[ 5.5 ms ] story [ 28.8 ms ] threadEdit: from the source [1], this quote pretty much sums it all up: "Our 2022 paper predicted that high-quality text data would be fully used by 2024, whereas our new results indicate that might not happen until 2028."
[1] https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-...
There is also the problem that on-device learning is not yet practical.
There is evidence that training RNN models that compute several steps with same input and coefficients (but different state) lead to better performance. It was shown in a followup to [1] that performed ablation study.
[1] https://arxiv.org/abs/1611.06188
They fixed number of time steps instead of varying it, and got better results.
Unfortunately, I forgot the title of that ablation paper.
The fixed point nature of DEQs means that they inherently have a concept of self assessment how close they are to the solution. If they are at the solution, they will simply stop changing it. If not, they will keep performing calculations.