Reading a map of learned representations AI
The image map reduces each final 512-value pooled representation to two coordinates. The activity map reduces each 128-value final hidden state from selected epoch 15. Each point represents an input example, not an individual channel, hidden unit, or learned weight.
UMAP builds a low-dimensional arrangement from neighborhood relationships. Here cosine distance compares vector directions; 15 neighbors and minimum distance 0.15 control the display. Labels color the points but do not enter the fit. See the UMAP parameter guide.
The classifier uses the original representation, not the two coordinates. Nearby points are candidates for similar responses, but apparent gaps, cluster sizes and axis directions are not direct measurements of category boundaries. The pet report’s nearest-neighbor cards use the original vector space to help check the projection.
Both layouts remain fixed while inspecting predictions across training. Their points do not trace trajectories between checkpoints. Nor are the image and activity maps aligned with each other.
The Research Atlas representation note describes a different use: a separate PCA/Ward grouping pipeline alongside a UMAP display. This note concerns classifier representations and their connection to saved activations.