For empiricism, imagination is a way out of a tight spot. If all ideas come from experience, where do winged horses or the New Jerusalem come from? Hume answers that imagination cannot exceed the material supplied by the senses, but has an unlimited power of mixing, compounding, separating and dividing ideas (Buckner, 2024, pp. 190, 196). Buckner (2024) illustrates the productivity of thought in his classes with a pink elephant in a tutu hanging from the ceiling while singing a sea shanty: no one has seen it, everyone understands it (2024, p. 196). Fodor praised Hume for seeing the problem and faulted him for not saying how the faculty works: a "something-I-know-not-what" (2024, p. 199).
Buckner's thesis is that deep learning can redeem that promissory note (2024, p. 191). Discriminative networks go from exemplars to abstraction; generative ones (GANs, VAEs, generative transformers) travel the other way, turning abstractions into new exemplars (2024, p. 205). They can thus synthesize instances of a category, fuse categories, and simulate future or counterfactual experience for decision-making or for training offline (2024, p. 237). DeepMind's I2A agent generates imagined "rollouts" of its actions before choosing, and so solved Sokoban puzzles (2024, pp. 232–233) (see world model).
The nativist objection is that these networks merely interpolate within their data, whereas human imagination extrapolates. Buckner replies that the distinction depends on how the "inside" of the data is defined: measured over pixels, nearly every decision a network makes on a new case already counts as extrapolation (2024, pp. 221–222). He also cautions that the vocabulary of "imagination" in AI may be mere loose inspiration (2024, p. 192, n. 1).
When an image generator produces a winged horse, does it compose ideas as Hume said, or interpolate in a latent space? And is there a difference, if composition itself could be a form of interpolation in the right space?