Multi-Headed Attention is likely the most important architectural paradigm in machine learning. This summary goes over all critical mathematical operations within multi-headed self attention, allowing you to understand it’s inner workings at a fundamental level. If you’d like to learn more about the intuition behind this topic, check out the IAEE article.
Multi-headed self attention (MHSA) is used in a variety of contexts, each of which might format the input differently. In a natural language processing context one would likely use a word to vector embedding, paired with positional encoding, to calculate a vector that represents each word. Generally, regardless of the type of data, multi-headed self attention expects of sequence of vectors, where each vector represents something.