SPD paper report
Run information accompanying the Stochastic Parameter decomposition paper
Created on June 11|Last edited on June 15
Comment
In this report, we collect all the runs used in the paper and show some additional information (more information can be found by clicking on the runs below). All code is produced using https://github.com/goodfire-ai/spd.
Figure guide:
- causal_importance_upper_leaky: Also shown in the paper, these figures show the importance values of each subcomponent in response to one-hot inputs (xi = 0.75). As mentioned in Appendix A.2, we use an upper leaky hard sigmoid because the causal importance values cannot be allowed to take negative values, else the training would not be incentivized to sparsify them. These are the values we show here.
- stochastic_recon, stochastic_recon_layerwise and importance_minimality: As described in the paper.
- mean_component_activation_counts: Over an evaluation dataset of 100 batches, calculates the frequency with which each causal importance value is above 0 (i.e. that the activation of this component is causally important). Note, in resid_mlp 3-layer, there are many causal importance values slightly larger than 0 and thus the corresponding components have little influence on the masked forward pass. Future versions of the code set a higher threshold for these plots (e.g. 0.1).
TMS 5-2 results
Run: nrandmasks1_randrecon1.00e+00_layerwiserandrecon1.00e+00_p2.00e+00_import1.00e-04_C200_sd0_lr1.00e-03_bs4096_ft40_hid10hid-layers0
1
TMS 5-2 + Identity results
TMS 40-10
TMS 40-10 + Identity Results
Resid_MLP 1-layer
Run: nrandmasks1_randrecon1.00e+00_layerwiserandrecon1.00e+00_p2.00e+00_import1.00e-04_C200_sd0_lr1.00e-03_bs4096_ft40_hid10hid-layers0
1
Resid_MLP 2-layer
Resid_MLP 3-layer
Add a comment