My pipeline

[Microcosmos22]

Plan is as follows:

1. Make predictions using pretrained Google Perch.
2. Calculate Priors Table, which contain (marginal) distributions given location and time
   P(class | site, hour)
3. Combine them either using:

• fuse_scores_with_tables: linear combination using parameter lambda
• Add simple stacker that learns:

final = w1 × base + w2 × prior + bias

And learns: w1, w2 per class.

5. Add PCA and Probes
   PCA reduces the feature vector to its most relevant components:
   2048 → maybe 64 or 128 dimensions
   Probes grab the feature vectors + other characteristics and predicts the scores itself, rather than letting Perch compute the scores from the feature/embeddings only
   probe sees:

PCA embedding        ← acoustic features
raw score            ← Perch output
prior score          ← site/hour prior
base score           ← fused base
prev/next score      ← temporal context
mean/max score       ← smoothing context

7. Fine-tune, unlikely