Triplet Embedding Judgements- CFD Faces (cleaning)

This notebook is for documenting how datafiles straight from the Sapir server are cleaned for use in triplet embedding algorithms. Participants each completed a jsPsych-based study with 1000 judgements each. A validation set of 12 triplets is included for each participant.

---

After generating embeddings… (these use SALMON embedding code)

Using everyone’s embeddings…

p_1000

p_2000

Embeddings for individual participants…

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

## [1] "------------------------------------------------"

Planned analyses…

1. How similar are individual people’s embeddings to the group? To other individuals? To their demographic group?

• cosine similarity analysis
• Measure of “spread” for individual images (& image groups) in embedding space?

2. Improved embedding? Pick for model accuracy over training time? Different loss?

3. How could this relate to behavior about faces?