Name of Project: ReAct  Recurrence for Attention (Yes I’m looking for cooler names, suggestions welcome! )
Proposal in one sentence: Attempting to explore how recurrence as a prior can help Transformers extrapolate to unseen datapoints during inference/deployment
Description of the project and what problem it is solving: Current DL/AI models can’t extrapolate well to datapoints unseen during training, a capability which arises from scale. By imbuing recurrent priors, I aim to replicate the same ability for much cheaper and smaller models which can be deployed in a practical setting and thus are more feasible for consumers to train and research on without requiring a cluster of expensive GPUs.
It is highly experimental and technical, so please let me know in the comments what parts I should elaborate on  otherwise I’d be filling pages here with papers and explanations.
Grant Deliverables:
choose deliverables you can complete in a month’s worth of parttime work
 PoC of extrapolation on Arithmetic task (simple addition) outside 12 digits from its training dataset.
 In 34 months, being able to scale the above and empirically show model can extrapolate to distinct OOD samples
Spread the Love
If you plan to use some of the funds to reward contributions from other community members, please describe your desired roles/skillsets e.g. looking for a data scientist to work with me to develop a machine learning model. If successful, this role will be advertised on the Algovera opportunities board and DeWork.
Squad
Squad Lead:
 Twitter handle: N/A
 discord handle: Awesome_Ruler_007#7922
Additional notes for proposals

Here’s the WandB report for progress tracking: WandB Notes Dashboard
→ All Code, Configs, Notes, Ideas, Model Checkpoints, Logs and performance results are synced on WandB automatically 
Grant Money would go for hardware and GPUhours for training models. Since every iteration is expensive, I try to work as effectively as possible but there’s a limit here. I believe the grant can help be accelerate my research and hopefully get a PoC out to confirm whether this direction is worth pursuing.
I have a brief writeup here that is semitechnical. As again, I can go indepth in explanations as warranted  lmk below.
Basically, its exploring how recurrence as a prior can aid in OOD generalization and allowing for dynamic compute and memory at inference time (dataset ratios and tokenization matter too but broadly these are key areas)
I’m basing my exploration/experimentation off this interesting result: https://twitter.com/tomgoldsteincs/status/1596210043019722752?cxt=HHwWgIDQkfek8KYsAAAA
This paper demonstrated how CNNs can OOD extrapolate at inference time just by scaling iterations. They leverage a CNN trained to solve 9x9 mazes, which extrapolates to 801x801 mazes even  while conventional convnets barely generalize to 13x13 sized ones. (While the method looks extremely similar to diffusion, the difference is that we frame the entire process as a single Markovian chain and don’t denoise, thus saving on compute)
Interestingly, I’ve already tried some preliminary experiments. Turns out, learning
f(x)=x
is much faster for this arch than memorizing a simple sequence! So the hypothesis of recurrence being a strong prior against overfitting and promoting generalization has some support empirically.
(Very) Initial runs on addition seem to generalize decently on a couple of digits outside training distribution. Further experimentation would be needed to see whether I can push that to its limits and generalize for more than a handful of digits.
My hypothesis is that recurrence introduces a prior for generalization and combating memorization; However, seeing results and other works, It seems more a property of employing differential equations. Still, empirical results are a bit lacking here so that’s what I’d be focusing on!