Related Experiment Videos
Connectionism and the learning of probabilistic concepts
1MRC Applied Psychology Unit, Cambridge, U.K.
Summary
This study investigated if connectionist networks can simulate medical diagnosis learning. Experiments found that a specific network model accurately predicted human bias in probability judgments, supporting its use in simulating medical concept learning.
Area of Science:
- Cognitive Science
- Artificial Intelligence
- Psychology
Background:
- Gluck and Bower proposed connectionist networks could simulate medical concept learning, with symptoms as input and diagnosis as output.
- Previous studies showed human bias in probability judgments for rare vs. common diseases, which Gluck and Bower linked to their network model.
Purpose of the Study:
- To examine the validity of Gluck and Bower's claim regarding connectionist network simulation of medical concept learning.
- To investigate whether a connectionist network using the Rescorla-Wagner learning rule could predict human bias in probability estimates.
Main Methods:
- Three experiments were conducted, involving subjects estimating disease probabilities given symptoms.
- The first experiment replicated Gluck and Bower's findings.
- The second and third experiments modified the design to address assumptions about probability estimation and network architecture.
Main Results:
- The initial experiment replicated human bias, but the connectionist network, as applied by Gluck and Bower, did not predict this bias.
- A revised experimental design in the third experiment showed that human biases were predicted by the connectionist network when it lacked hidden units.
Conclusions:
- Evidence was found supporting the connectionist network model's ability to simulate human probability judgments in medical diagnosis tasks.
- The findings suggest that connectionist models, particularly those without hidden units, can account for observed biases in medical concept learning, challenging purely normative accounts.