NRGPT: An Energy-based Alternative for GPT
This work addresses the challenge of integrating energy-based paradigms into popular language models, offering an incremental alternative for researchers in machine learning.
The authors tackled the problem of unifying GPT architectures with energy-based modeling by proposing NRGPT, a minimal modification that conceptualizes inference as exploring an energy landscape, and demonstrated its performance on tasks like Shakespeare, ListOPS, and OpenWebText, with models showing resistance to overfitting.
Generative Pre-trained Transformer (GPT) architectures are the most popular design for language modeling. Energy-based modeling is a different paradigm that views inference as a dynamical process operating on an energy landscape. We propose a minimal modification of the GPT setting to unify it with the EBM framework. The inference step of our model, which we call eNeRgy-GPT (NRGPT), is conceptualized as an exploration of the tokens on the energy landscape. We prove, and verify empirically, that under certain circumstances this exploration becomes gradient descent, although they don't necessarily lead to the best performing models. We demonstrate that our model performs well for simple language (Shakespeare dataset), algebraic ListOPS tasks, and richer settings such as OpenWebText language modeling. We also observe that our models may be more resistant to overfitting, doing so only during very long training.