animal2vec workshop · MPIABDay 3 · 11:00–11:45
Explore the structure.
Test the interpretation.
An unsupervised-learning toolbox without unnecessary mathematics.
Julian C. Schäfer-Zimmermann
Max Planck Institute of Animal Behavior
Department for the Ecology of Animal Societies
Communication and Collective Movement (CoCoMo) Group
animal2vec workshop · MPIABDay 3 · 11:00–11:45

Start with the question—not the plotting function

Your question First tool Its immediate output
What resembles this example? Nearest neighbours A ranked set of observations
What directions contain variation? PCA Linear coordinates and explained variance
What is locally neighbouring? UMAP A neighbourhood-oriented layout
Can examples be grouped? Clustering Candidate assignments
Which examples are unusual? Outlier scoring A ranking for inspection
animal2vec workshop · MPIABDay 3 · 11:00–11:45

Nearest neighbours: retrieve first, interpret second

Interactive example requires JavaScript.

In your real dataset: inspect the audio, remove self-matches, and avoid near-duplicate segments.

animal2vec workshop · MPIABDay 3 · 11:00–11:45

PCA chooses directions that retain variation

Interactive example requires JavaScript.

animal2vec workshop · MPIABDay 3 · 11:00–11:45

A PCA map is a summary, not the complete space

What to read

Which differences appear along each axis?
How much variance do these axes retain?

What not to conclude

Large variation is not automatically important biology.
Two dimensions need not preserve all useful distinctions.

scikit-learn PCA centers columns; it does not automatically standardize their scales.
animal2vec workshop · MPIABDay 3 · 11:00–11:45

UMAP begins from relationships between neighbours

Interactive example requires JavaScript.

The graph below is a schematic. No UMAP fit is computed in this slide.

animal2vec workshop · MPIABDay 3 · 11:00–11:45

A UMAP island is not a discovered call category

Useful

Explore local neighbourhoods.
Inspect examples.
Compare metadata and parameter choices.

Not established by the map

Biological category count.
Meaning of empty gaps.
Original-space density or global distance.

Check stability and the original feature space before treating a visual separation as evidence.
animal2vec workshop · MPIABDay 3 · 11:00–11:45

k-means: ask for K groups, then get K groups

Interactive example requires JavaScript.

Cluster assignments are fitted in the 12-D feature space. The PCA coordinates only display them.

animal2vec workshop · MPIABDay 3 · 11:00–11:45

Density-based grouping can leave points unassigned

Interactive example requires JavaScript.

Grey observations are marked noise. “Noise” here is an algorithmic status—not a biological annotation.

animal2vec workshop · MPIABDay 3 · 11:00–11:45

A candidate group needs an audit

Inspect Why
Typical examples and boundary cases A centroid alone can hide variation
Individual / session / device / site Technical or sampling factors may explain the group
Different parameters and resamples A fragile partition may not support a stable interpretation
Independent annotated observations Test the proposed biological interpretation

What observation would make you reject your current interpretation of a cluster?

animal2vec workshop · MPIABDay 3 · 11:00–11:45

Outliers are candidates for inspection

Interactive example requires JavaScript.

Rare vocalization? Corrupted file? Unusual background? Annotation mistake? Listen before naming it.

animal2vec workshop · MPIABDay 3 · 11:00–11:45

Turn exploration into an annotation decision

Retrieve / group / rankGenerate candidates
→
Listen and annotateTest the interpretation
→
Update the datasetAdd examples and document changes

Seek coverage

Inspect diverse and underrepresented regions.

Keep a comparison sample

Do not review only what the current model considers unusual.

animal2vec workshop · MPIABDay 3 · 11:00–11:45

One visible structure can support several stories

Interactive example requires JavaScript.

State a hypothesis with no labels. Then reveal site, device, and call group in that order.

animal2vec workshop · MPIABDay 3 · 11:00–11:45

Change a decision; check whether the finding survives

Decision to perturb Question to ask
Model / layer choice Does the proposed structure depend on one representation?
Aggregation / normalization / metric Does the observation definition change the conclusion?
Projection / clustering settings / seed Is the visible or assigned structure stable?
Held-out sessions / individuals / sites Does the interpretation survive a new context?
Keep exploratory findings separate from confirmatory evidence.
animal2vec workshop · MPIABDay 3 · 11:00–11:45

Next: the mystery-dataset exercise

Before metadata reveal

Describe the structure.
Inspect examples.
Write a testable hypothesis.

After metadata reveal

Compare alternative explanations.
Revise the hypothesis.
Identify the next validation step.

Your deliverable is an argued interpretation—not just a plot.
animal2vec workshop · MPIABDay 3 · 11:00–11:45

Use the map to ask better questions

A neighbour is a retrieval result.
A cluster is a candidate grouping.
An outlier is a review priority.

A biological conclusion requires evidence beyond the geometry.

Which analysis will you try first—and what would convince you that it is useful?