Mind · Inference
PredictiveBehaviorEmbedding#c46b
Train a network to predict future motion; its internal features separate behaviours it was never told about
Mechanism
To find individual behavioural states without labels, train a model on a self-supervised prediction task and read its hidden layers. For ants: represent each pose as joint angles in exponential-map form (63 dimensions), take 25 observed frames at 15 fps, compress the 35-frame window (observed plus predicted) with a discrete cosine transform, and train a graph convolutional network (37 nodes, 256 hidden features, 12 residual blocks with batch norm, tanh and dropout 0.5, learnable node-to-node attention, about 2.1 M parameters) to predict the future joint angles with an L1 loss. Hook the output of the final residual block, standardise, and project with PCA: the first three components (47% of variance) already separate walking, running and antennation. A model that must forecast motion has to compress it, as in ModelAsCompression, and the compression aligns with behavioural state. The aim is per-agent state labels that, with network structure, let one study AdaptiveNetworkCoevolution between individual states and collective topology.
Invariants
- No behavioural labels are used in training; labels are only used afterwards to inspect the embedding
- Embeddings are taken from a fixed layer after training, then standardised before projection
Failure modes
- Visual separation in a 3-D projection is not a classifier; the split must be validated on held-out individuals
- Trained on very few individuals, the features may encode identity rather than behaviour
- Missing joint angles from tracking failures, if imputed naively, create spurious clusters
Where this came from
- Thesis Kevin Do, PhD thesis on ant collective behaviour (NCSU, 2026), Ch. 2.3, Fig. 2.5; Ch. 5.7.3
Builds on
Identity
The name is the hash of the definition. Change a word and it becomes a different pattern.
- Sema ID
sema:PredictiveBehaviorEmbedding#mh:SHA-256:c46b7187c440856b539933c739aa8631852808e27fc784330102a57239d97a3e- Reference
PredictiveBehaviorEmbedding#c46b- Ring
- 2
- Tier
- 3