Value-centric abstraction represents relationships through continuous similarity, perception, and intuition rather than exact structural rules. It is the kind of abstraction used when recognizing patterns that may resemble one another without sharing an explicit programmatic form. section 19.4.3 section 19.5
Deep learning is especially effective at this form of abstraction, but it does not by itself provide program-centric abstraction: discrete reasoning, planning, and exact structural manipulation. Human-like intelligence requires both approaches, which occupy a spectrum rather than two completely separate categories. section 19.4.3 section 20.3
Sorting versus recognizing digits
Sorting a vector is a poor fit for value-centric learning: a model needs extensive data, can still make mistakes, and may require retraining for longer vectors. A short sorting program generalizes to vectors of any length. Conversely, classifying handwritten MNIST digits is a poor fit for pure discrete programming because perception requires laborious hand-coded detectors for visual properties. section 19.4.3 section 19.4.3
Value-centric abstraction contributes to AI systems by:
- Recognizing patterns through learned geometric representations. section 19.5
- Providing informal intuition that can guide searches through possible programs. section 20.3
- Combining with algorithmic modules in systems intended to generalize across new tasks. section 19.4.7
