Theories or fragments?

Lake et al. argue persuasively that modelling human-like\ud intelligence requires flexible, compositional representations in order to\ud embody world knowledge. But human knowledge is too sparse and selfcontradictory\ud to be embedded in “intuitive theories.” We argue, instead,\ud that knowledge is grounded in exemplar-based learning and\ud generalization, combined with high flexible generalization, a viewpoint\ud compatible both with non-parametric Bayesian modelling and with subsymbolic\ud methods such as neural networks