For a neural networks, the goal is to find isomorphism's with real data. This may be performed by creating embedding of objects and then weighted attention based on the similarities of the embedding's

There are several kinds of top down attention

_the_increase_of_stimulus_salience_by_enhancing_de.jpeg)

Low sharpness means high generalization

Generative pertaining: Focus on the data first. Use a small semantically aligned model to filter out specific kinds of examples (ex pornography, hate speech) Also use feature encoder to identify and downweight data that is unclear (blurry images)

There needs to be a way to interchange intelligence on every major scale of engineering:

• distributed systems that hunt for / inform each other of their API’s
• nano-packages that a system can use
• weights and neural networks for transfer