Analyses¶

We treat sampled neural-network weights as data. Each topic asks one question; each experiment page states its own question, answer, protocol and results.

Topic	Question
Posterior geometry	What does the sampled posterior look like in weight space?
Sampler trajectories	What do sampler paths look like, and why?
Samplers	Which samplers do better, and at what cost?
Weight representations	How compactly and portably can single draws be encoded?
Weight generation	Can we generate posterior populations?

Campaigns¶

Dataset	Architecture	Sampler	Chains × draws
Airfoil	FCN 16 × 16 × 16 × 1 · relu	MCLMC	32 × 200
Bikesharing	FCN 16 × 16 × 16 × 1 · relu	MCLMC	32 × 200
Energy	FCN 16 × 16 × 16 × 1 · relu	MCLMC	32 × 200
CIFAR-10	ResNet7	SMILE	8 × 200

Additional campaign configurations

Dataset	Architecture	Sampler	Chains × draws
Airfoil	FCN 16 × 16 × 16 × 1 · relu	MCLMC	256 × 50
Energy	FCN 16 × 16 × 16 × 1 · relu	MCLMC	256 × 50
Airfoil	FCN 16 × 16 × 16 × 1 · relu	MCLMC	8 × 1000
Airfoil	FCN 8 × 8 × 8 × 1 · relu	MCLMC	16 × 100
Airfoil	FCN 32 × 32 × 32 × 1 · relu	MCLMC	16 × 100
Airfoil	FCN 64 × 64 × 64 × 1 · relu	MCLMC	16 × 100
Airfoil	FCN 16 × 16 × 16 × 1 · relu	MCLMC	8 × 100000
Bikesharing	FCN 16 × 16 × 16 × 1 · relu	MCLMC	8 × 100000
Energy	FCN 16 × 16 × 16 × 1 · relu	MCLMC	8 × 100000
Airfoil	FCN 16 × 16 × 16 × 1 · relu	MCLMC	8 × 1000
Airfoil	FCN 16 × 16 × 16 × 1 · relu	NUTS	8 × 1000
Airfoil	FCN 16 × 16 × 16 × 1 · relu	RANDOM_WALK	8 × 1000
Bikesharing	FCN 16 × 16 × 16 × 1 · relu	MCLMC	8 × 1000
Bikesharing	FCN 16 × 16 × 16 × 1 · relu	NUTS	8 × 1000
Bikesharing	FCN 16 × 16 × 16 × 1 · relu	RANDOM_WALK	8 × 1000
Energy	FCN 16 × 16 × 16 × 1 · relu	MCLMC	8 × 1000
Energy	FCN 16 × 16 × 16 × 1 · relu	NUTS	8 × 1000
Energy	FCN 16 × 16 × 16 × 1 · relu	RANDOM_WALK	8 × 1000