The stellar performance of large language models (LLMs) such as ChatGPT has shocked the world. The breakthrough was made by the invention of the Transformer architecture, which is surprisingly simple and scalable. It is still built of deep learning neural networks. The main addition is the so-called “attention” mechanism that contextualizes each word token. Moreover, its unprecedented parallelisms endow LLMs with massive scalability and, therefore, impressive accuracy after training over billions of parameters.
The simplicity that the Transformer architecture has demonstrated is, in fact, comparable to the Turing machine. The difference is that the Turing machine controls what the machine can do at each step. The Transformer, however, is like a magic black box, learning from massive input data through parameter optimizations. Researchers and scientists are still intensely interested in discovering its potential and any theoretical implications for studying the human mind.
In this article, we will first discuss the four main features of the Transformer architecture: word embedding, attention mechanism, single-word prediction, and generalization capabilities such as multi-modal extension and transferred learning. The intention is to focus on why the architecture is so effective instead of how to build it (for which readers can find many…