Transformer architecture
make bigger
Transformer architecture
make bigger
Decoder
words are under, not properly displayed, same for some others as well
softmax(QKᵀ / √dₖ)V
formula needs to be presented in correct latex form
Same words, different roles
ai speak
A developing inquiry. One idea at a time
ai speak
We are thus led to examine how the parts of a sentence contribute to its interpretation. How do we determine which words belong together? What establishes the role of one word in relation to another? And how does the surrounding context help us resolve a relationship that would otherwise remain uncertain?
say something about semantics
Definitions and proofs will establish properties of the constructions themselves; experiments and interactive examples will allow us to investigate their behavior and what they acquire through learning. Throughout, we shall return to the linguistic questions that motivated the development.
This deviated from the math-speak we want
What draws me to this subject is the gap between how naturally we read and how difficult it is to say what a machine must represent. I want to make that difficulty visible, and then work through it carefully enough that the mathematics answers questions we have already learned to ask.
note about semantics
Language → structure → mathematics
ai speak
Beginning with language, and letting the questions lead us toward the mathematics.
ai speak, remove
ame words. A different event.
ai speak
An inquiry, one idea at a time ↓
ai speak