The paper's minimum description length (Appendix C), in bits per input at distortion δ = distortion_fraction · (mean per-dimension variance of x):
More...
#include <block_sparse_featurizer.hpp>
The paper's minimum description length (Appendix C), in bits per input at distortion δ = distortion_fraction · (mean per-dimension variance of x):
- support:
log2 C(G, ℓ) for G features (blocks) and ℓ active per input on average;
- code:
Σ_g p_g Σ_j ½ log2(1 + σ²_gj / δ), σ²_gj the eigenvalues of block g's code covariance where it is active and p_g its firing rate;
- residual:
Σ_a ½ log2(1 + λ_a / δ) over the residual covariance's eigenvalues;
- dictionary:
½ · G·b(d - b) / N · log2 N (the free parameters of G points on the Grassmannian Gr(b, d)). A shorter description at the same distortion is the paper's evidence that blocks fit the activations better than single directions.
◆ code
| double pulsatrix::DescriptionLength::code = 0 |
◆ delta
| double pulsatrix::DescriptionLength::delta = 0 |
◆ dictionary
| double pulsatrix::DescriptionLength::dictionary = 0 |
◆ residual
| double pulsatrix::DescriptionLength::residual = 0 |
◆ support
| double pulsatrix::DescriptionLength::support = 0 |
◆ total
| double pulsatrix::DescriptionLength::total = 0 |
The documentation for this struct was generated from the following file: