Source
David S. Watson (2023), “On the Philosophy of Unsupervised Learning”, LIT-187, Philosophy & Technology 36:28, read in full in NOTE-172.
What was actually shown
One structure, three times. Each task gets an epistemic thesis and an ontological one.
- Clustering (§2). EC: we learn to identify natural kinds via clustering or something like it. Accepted, on functionalist grounds. OC: natural kinds are what clustering ought to find under ideal conditions. Rejected: there is “no a priori reason to believe that all natural kinds are effectively computable”, so identifiability via clustering is “not a necessary condition for natural kindhood” (p. 7). Sufficiency is left to a “broadly pragmatic” sketch: kinds are grouped for a context and purpose, constrained by inductive usefulness and participation in laws, and kinds and algorithms “may even be co-constitutive” (p. 8).
- Abstraction (§3). EA accepted. OA rejected: “Abstraction algorithms may be sufficient to identify some essential properties, but essential properties are not necessarily identifiable via abstraction algorithms” (p. 12).
- Generative models (§4). EG accepted. OG rejected: it needs all possibilities to be effectively computable, and restricting possibility to the computable is “ad hoc and circular” (pp. 16–17).
Objectivity (§5). Hyperparameters (k, cut height, distance, architecture), data provenance and opacity condition every result, and clustering never tests k = 1, so it can produce false positives without a correction such as the gap statistic. Against vacuity: practitioners use a few simple distances, and pipelines corroborate clusters, for example cancer subtypes checked with log-rank survival tests (pp. 18–19). The proposed remedy is Mayo’s severity, made partly concrete as resampling stability, generalization of clusters to held-out data, out-of-sample likelihood and train-on-synthetic scores (pp. 20–21). No severity is computed for any example.
What could have come out otherwise. A kind could have been shown to be uncomputable, which would have turned the rejection into a result. The paper gives no such case.
What this does not say
- That some kind is not computable. The ontological theses are rejected because a kind might be uncomputable. No example is offered.
- That clustering finds kinds. The epistemic theses are hedged (“or something very much like them”) so heavily that it is unclear what would falsify them.
- That kinds are pragmatic. That view is asserted: “I will go to no great lengths defending this” (p. 8).
- That unsupervised evidence is privileged. The gene-expression example says an unsupervised separation is “arguably better evidence” of a treatment effect. That is hedged, and Søgaard (LIT-162) denies any such privilege.
- That unsupervised learning is the most fundamental branch of AI. The abstract says so; the body supports it with one sentence (p. 3).
- Anything about modern pretraining. Footnote 1 excludes self-supervised learning and foundation models.
Connections
- QUESTION-028. The question asks when robust, generalizable, intervention-relevant structure warrants treating a communicative category as a real kind. This account constrains the answer: an algorithm’s finding a category is evidence about it, and on this view cannot be what makes it a kind. It also concurs with A191’s “None of these independently establishes a metaphysical ontology”. It does not say what would establish one beyond the pragmatic sketch.
- Natural kinds. THEORY-273 holds that inductive usefulness is necessary but not sufficient for a natural kind. Watson’s sketch of sufficiency leans on inductive usefulness and participation in laws, so the two meet there.
- Søgaard (LIT-162) replies to the line of argument that clustering is epistemically privileged. NOTE-172 pairs them as claim and rebuttal.
- Persistence as a criterion. QUESTION-113 asks whether stability under refinement picks the same categories as evidence does. Watson’s stability tests are one of the severity tools he names, and he does not say stability suffices.