Abstract
The Computational Gauntlet of Human-Like Learning
Pat Langley1,2
1
    Institute for the Study of Learning and Expertise, 2164 Staunton Court, Palo Alto, CA 94306 USA
2
    Center for Design Research, Mechanical Engineering, Stanford University, Stanford, CA 94305 USA



Despite impressive advances, machine learning has abandoned many of its early profound
insights. The discipline has become overly reliant on statistical analysis of massive data sets and
strayed far from its conceptual roots. A promising alternative, common at the field’s founding
four decades ago, is to develop mechanisms that learn like humans. We can use findings from
cognitive psychology to devise a computational gauntlet that systems must traverse, giving new
criteria for evaluation. Here are six core features of human learning that provide constraints:
 • Learning involves the acquisition of modular cognitive structures. This does not specify
   the structures’ details; only that expertise consists of discrete mental elements.
 • Learned cognitive structures can be composed during performance. Relevant elements
   of expertise are accessed and then combined as needed to produce behavior.
 • Expertise is acquired in a piecemeal manner, with one element added at a time. Humans
   learn one cognitive structure, then another, continuing until they achieve broad coverage.
 • Learning is an incremental activity that processes one experience at a time. This is linked
   to on-line approaches that interleave learning tightly with performance.
 • Learning is guided by knowledge that aids interpretation of new experiences. Because
   acquisition is incremental, it occurs in the context of existing mental structures.
 • Cognitive structures are acquired and refined rapidly, each from small numbers of train-
   ing cases. This is enabled by piecemeal, incremental, and knowledge-guided processing.
The literature does contain systems, some older and others more recent, that exhibit these
characteristics and thus run the computational gauntlet successfully. These include:
 • Cobweb (Fisher, 1987), which constructs probabilistic taxonomies from unsupervised cases;
 • Prodigy (Minton, 1990), which acquires control rules from planning traces to guide search;
 • SAGE (McClure, 2015), which invokes structural analogy to learn relational concepts; and
 • MIL (Muggleton, 2018), which uses meta-interpretive abduction to acquire visual categories.
These artifacts offer positive role models for the research community, but we need more efforts
that take seriously the power of human learning and that aim to replicate its main features.
Note: This is an abstract of a published paper: Langley, P. (2022). The computational gauntlet of
human-like learning. Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence
(pp. 12268–12273). Vancouver, BC: AAAI Press. https://doi.org/10.1609/aaai.v36i11.21489
Third International Workshop on Human-Like Computing, September 28–30, 2022, Windsor Great Park, UK
Envelope-Open langley@stanford.edu (P. Langley)
GLOBE http://www.isle.org/~langley/ (P. Langley)
Orcid 0000-0001-5260-7048 (P. Langley)
                                       © 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
    CEUR
    Workshop
    Proceedings
                  http://ceur-ws.org
                  ISSN 1613-0073
                                       CEUR Workshop Proceedings (CEUR-WS.org)