iris VS returnn-experiments

Yea, after all these LLMs are predicting one sequence of tokens from another sequence of tokens and the tokens could be anything, it just "happens" that text has the most knowledge and the easiest to input, then there are image, sound, video, but tokens could also be learned from world experience in RL:

Transformers are Sample-Efficient World Models:

https://github.com/eloialonso/iris#transformers-are-sample-e...

Transformers are Sample Efficient World Models https://github.com/eloialonso/iris https://arxiv.org/abs/2209.00588v1

This is actually a pretty neat, self-contained implementation that can super easily extended beyond stereotypical natural language models, for example to create world models for video games [1] or to create robot models that can learn to imitate from large, chaotic human demonstration data [2] (disclaimer, I'm an author on the second one.) Basically, GPT (or minGPT) models are EXCELLENT sequence modelers, almost to the point where you can throw any sensible sequence data at it and hope to get interesting results, as long as you don't overfit.

Even though I have only been working on machine learning for around six years, it's crazy to see how the landscape has changed so fast so recently, including diffusion models and transformers. It's not too much to say that we might expect more major breakthroughs by the end of this decade, and end in a place we can't even imagine right now!

[1] https://github.com/eloialonso/iris

The code is all released already. You find it here: https://github.com/rwth-i6/returnn-experiments/tree/master/2...

This is TensorFlow-based. But I also have another PyTorch-based implementation already, also public (inside our other repo, i6_experiments). It's not so easy currently to set this up, but I'm working on a simpler pipeline in PyTorch.

We don't have the models online yet, but we can upload them later. But I'm not sure how useful they are outside of research, as they are specifically for those research tasks (Librispeech, Tedlium), and probably don't perform too well on other data.

This works for an architecture which has been well tuned and studied before, like LSTM or Transformer.

Once you do research on the model, testing out things, it often tends to become such kwarg monster in many frameworks.

Having everything (relevant) in one file (even in the config file itself with hyper params) allows you to copy the file for every experiment and modify it inplace. This avoids the kwargs mess. But then the config files are very complex, and can become messy in other ways (esp for research projects). Example: https://github.com/rwth-i6/returnn-experiments/blob/master/2...

Such approach makes it much more flexible and does not mess with the baseline code. As you say, it's more like an evolutionary DNA-like approach, where you then tend to do crossovers with other evolved good-performing configs, etc.