Abstract
<p>The striking constraints of some human cognitive processes stand in stark contrastto the near limitless capability of others. While we can acquire and flexibly use vast amountsof information, the amount we can process at any one time is often stiflingly limited: forexample the number of items we can hold in working memory or the number of tasks that canbe performed at once. Here, we integrate ideas from information theory, cognitive science,and neuroscience to offer a unified account of why processing is often so limited. We arguethat this reflects a fundamental tradeoff between generalization—how effectively existingrepresentations can be used in novel settings—and how many distinct representations canbe processed in parallel. Representations that best promote strong forms of generalization— a characteristically human cognitive strength — come at the expense of surprisinglystrict limits in the number of items that can be processed at once, an equally characteristichuman weakness. We refer to this as the “curse of generalization.” We formulate this first ininformation-theoretic terms, and then in process models, including a neural network modelof classic tasks used to demonstrate strict limits in human processing capacity. This tensionoffers a potential explanation for a range of phenomena — from performance on the tasks onwhich we focus, to representational learning and skill acquisition more broadly — as wellas the performance of modern machine learning architectures that exhibit generalizationcapabilities comparable to humans.</p>