Science Advances

Supplementary Materials

This PDF file includes:

  • Section S1. Alternative architectures and loss functions
  • Section S2. Impact of pretrained weights and dropout rate
  • Section S3. Functionally interpreting ML/MF and f3 using the generative model
  • Section S4. Psychophysics methods and model-free analysis
  • Fig. S1. A more detailed diagram of the modeling framework.
  • Fig. S2. Evaluation of VGG-Raw, VGG+, and EIG networks based on the FIV image set (extending Fig. 3).
  • Fig. S3. Scatter plots of data and model similarity matrices and analysis of earlier network layers (extending Fig. 3).
  • Fig. S4. Evaluation of alternative models using the FIV-S image set.
  • Fig. S5. Evaluation of the VAE models using the FIV-S image set.
  • Fig. S6. Trade-off arising from the choice of training targets and the use of pretrained weights.
  • Fig. S7. Variants of the EIG network architecture each trained from scratch without pretraining.
  • Fig. S8. Comparison of intermediate stages of the generative model to f3.
  • Fig. S9. Decoding analysis.
  • Fig. S10. Learning curve analysis.
  • Fig. S11. Lighting direction judgment experiment.
  • Fig. S12. Snapshot of a trial from the depth judgment experiment.
  • Fig. S13. Decoding lighting elevation and profile depth from the VGG network.
  • Table S1. ID network architecture their architectures, loss functions, and training procedures.
  • Table S2. VAE decoder architecture.
  • Table S3. VAE-QN pose architecture dimensional vector.
  • References (6165)

Download PDF

Files in this Data Supplement: