Transformers Beyond ChatGPT
A chatbot is the least surprising place to find a transformer. Words already arrive in sequence. The stranger cases begin when the “sentence” is a medical scan, several camera feeds, a protein representation, or a stream of security events.
That is a useful way to think about transformers beyond ChatGPT. The architecture does not require language. It requires a representation made of tokens and a task where relationships among those tokens matter.
Once you see that, eight applications that look unrelated start to rhyme. Image patches can borrow context from distant anatomy. Robot actions can be generated beside visual observations and instructions. Satellite patches can be compared across spectral bands and time. User interactions can become an ordered history rather than a bag of clicks.
Eight types of data converge on a central lattice: medical scans, robot observations, satellite tiles, driving scenes, protein representations, security events, item histories, and sensor...
Copyright of this story solely belongs to hackernoon.com. To see the full text click HERE
