Architecture Intelligence
A perspective. A set of open questions.
Architecture here includes network structure, learning algorithms, and the organization of computation. It is a working perspective for reading and practice, not an established unified explanation.
Architecture & computation
Which computations does a network structure make easier, and what does it constrain?
Look for the assumptions introduced by structural choices, and ask which controls could distinguish competing explanations.
Learning rules & capabilities
What capabilities emerge when the same structure is trained with different learning algorithms?
Consider structure, objectives, and update rules together. Track which conditions a paper changes and which it holds fixed.
Representations & mechanisms
How far is an observed manifold from an explanation of behavior?
Use representational geometry as an entry point, then ask how representations form, how they are used, and what interventions could test.
Neuroscience & AI
Which neuroscience questions can help us design better AI experiments?
Look for mechanisms and questions that can be compared, while recording where biological and artificial systems resist a direct analogy.
These questions will change with the reading and experiments.
Visit the notebook →