Towards Understanding Modern Alchemy - Transformers as a Computational Model

Explore the intricacies of in-context learning (ICL) in language models through a 37-minute lecture by Ekin Akyurek from MIT. Delve into the concept of in-context learning of formal languages (ICLL) as a model problem for understanding ICL. Examine how Transformers outperform recurrent and convolutional models in learning regular languages sampled from random finite automata. Discover the role of specialized "n-gram heads" in Transformers' superior performance and their potential to improve natural language modeling. Learn how incorporating these heads into neural models can enhance perplexity scores on datasets like SlimPajama. Gain insights into the computational capabilities of Transformers and their implications for advancing language model performance.