email_034-attachment-send-file-code-css

facebookresearch / decodable_information_bottleneck

We characterize representations that are ``optimal'' -- in terms of test loss -- for a given functional family (e.g. 2 layer MLP) by proposing notions of sufficiency (being able to predict the labels) and minimality (not being able to distinguish between examples with the same labels) with respect to that functional family.

Summary


		Main Code: 8,669 LOC (230 files) = PY (76%) + YAML (23%) Secondary code: Test: 0 LOC (0); Generated: 0 LOC (0); Build & Deploy: 0 LOC (0); Other: 114 LOC (5);
		Duplication: 12%
		File Size: 0% long (>1000 LOC), 40% short (<= 200 LOC)
		Unit Size: 0% long (>100 LOC), 62% short (<= 10 LOC)
		Conditional Complexity: 1% complex (McCabe index > 50), 63% simple (McCabe index <= 5)
		Logical Component Decomposition: primary (17 components)
		Goals: Keep the system simple and easy to change (4)
	Features of interest: TODOs 2 files