To abstract is to move from the particular to the general: from many triangles, each with its own size, color, and rotation, to the idea of triangle. Locke held that this general idea had to be "neither oblique, nor rectangle, neither equilateral, equicrural, nor scalenon; but all and none of these at once," and Berkeley mocked such a contradictory object (Buckner, 2024, p. 101). Buckner (2024) distinguishes four variants in the empiricist tradition: abstraction as subtraction (leaving out the details that vary), as composition (building abstract ideas out of simpler ones), as representation (reasoning with a particular exemplar that stands in for the whole class, in the manner of Berkeley and Hume), and as invariance (what stays constant under transformations) (2024, pp. 103–108).
His own proposal, transformational abstraction, starts from the fact that variation among exemplars is not chaotic noise: position, scale, or rotation are "nuisance variables" with exploitable symmetries. A system that learns the right transformations can bring any triangle into a form comparable with any other, so that categories come apart in a feature space (2024, pp. 112–114). Convolutional networks would do this with an "abstraction sandwich": convolution composes and pooling subtracts; transformers would achieve subtraction through attention (2024, p. 127).
Critics do not deny that this works, only that it suffices. Chollet (2021) describes deep learning's generalization as interpolation on a learned manifold: local generalization, which falls short of the extreme generalization that requires abstraction, reasoning, and symbolic models (2021, sec. 5.1.2). M. Mitchell (2019) calls the formation of humanlike conceptual abstractions "an almost completely unsolved problem" and illustrates it with Bongard problems (2019, ch. 15). Buckner grants that this is the heart of the dispute: do networks only interpolate, or do they also extrapolate (2024, p. 221)?
For philosophy, the question is whether abstraction requires explicit symbols and concepts or can be a subpersonal transformation of experience, as moderate empiricism suggests.