Self-supervised transformer on synthetic hierarchical objects encodes object-membership abstractions in low-dimensional, linearly readable directions

- A small masked-reconstruction transformer encoder (3 blocks, 2 heads, 64-d embeddings) trained from scratch on a synthetic Hierarchical Object Dataset (board-of-tokens objects) develops intermediate "abstractions": object-membership information that linear probes read out near-perfectly at the attention-update stages, including probes that generalize from one set of fuzzy-object instances to held-out instances and from unmasked to masked tokens. [ferry-etal-2023-emergence-and-function-of-abstract-representations-in-self-supervised-transformers] - The authors describe these abstractions as "low-dimensional, linearly separable attractors"/low-dimensional manifolds not aligned with the network's units: interchange interventions in a probe-derived rotated basis often change the prediction by swapping a single feature, versus an average of about 18 units needed in the unit basis. [ferry-etal-2023-emergence-and-function-of-abstract-representations-in-self-supervised-transformers] - Whether composite-object abstractions are built from root-object abstractions (part-whole hierarchy) is supported by the authors as preliminary, qualitative case-study evidence; no natural-language or natural-image pretraining is involved. [ferry-etal-2023-emergence-and-function-of-abstract-representations-in-self-supervised-transformers]