Concepts
“Concepts are the glue that holds our mental world together” (Murphy, 2002, p1).
This session will explore recent attempts to model the structure of human concepts. This will involve engaging with a historical debate and distinction between more “statistical” and more “rule-based” models and their implications but also thinking about how our concepts relate to how we categorize things but also how we think and generate new ideas.
Primary Readings
Everyone should read these and be prepared to discuss.
| Murphy, G. L. (2002) | Chapter 2, The big book of concepts. MIT Press (On Learn) |
| Goodman, N. D., Tenenbaum, J. B., Feldman, J., & Griffiths, T. L. (2008) | A rational analysis of rule‐based concept learning. Cognitive Science, 32(1), 108-154. |
Secondary Readings
The presenter should read and incorporate these.
| Lake, B. M., Salakhutdinov, R., & Tenenbaum, J. B. (2015). | Human-level concept learning through probabilistic program induction. Science, 350(6266), 1332-1338.
The article highlights two aspects of human conceptual knowledge that have eluded machine systems: the ability to learn new concepts from just one or a few examples, and the ability to learn rich and flexible representations. The authors propose the Bayesian Program Learning (BPL) framework, which can learn a wide range of visual concepts from just a single example and generalize in ways that are similar to human behavior. BPL represents concepts as stochastic programs and utilizes compositionality, causality, and learning to learn as key ideas. The authors present a computational model that captures these human learning abilities specifically for simple visual concepts, using handwritten characters from different alphabets as examples. The model represents concepts as simple programs that best explain observed examples, using a Bayesian criterion. In a one-shot classification task, the model achieves human-level performance and outperforms recent deep learning approaches. The model’s creative generalization abilities are tested through several “visual Turing tests,” which show that in many cases, the model’s behavior is indistinguishable from human behavior. This is the first really modern modelling paper on the reading list. It showcases the power of hierarchical Bayesian program induction as a theory of conceptual structure. In being based in symbolic rules/programs, it is a successor to Goodman et al (2008), but the domain is much more naturalistic and the computational cost of fitting this model is far higher. Note also the differences in writing style compared to the classic papers, e.g. compressed, concise, visual, snappy but key details in an appendix. |
| Shepard, R. N., Hovland, C. I. and Jenkins, H. M. (1961) | Learning and Memorization of Classifications. Psychological Monographs: General and Applied, 75(13), 1.The study includes a series of experiments that investigate the learning and memorization of different types of classifications. The results show that the structure of the classification and the complexity of the rules involved affect the difficulty of the task. Overall, the study provides insights into the factors that determine the difficulty of learning and remembering classifications. It highlights the importance of considering the structure of the classification, the complexity of the rules, and the use of different stimuli in understanding the learning and memorization process. The experiments in this paper are long and complicated but what makes it famous is the way Shepard visualizes concepts/classification rules of different complexity (i.e. Figures 2 and 3). These visualizations and the logic behind them are widely used in categorization research and teaching today. |
Questions under discussion
- How do concepts relate to categories?
- What about to previous topics? e.g. Are concepts representations?
