Neural networks
Some of the initial enthusiasm for connectionist modeling waned as researchers discovered that neural-network solutions to learning problems were difficult to understand and thus difficult to relate to human solutions. These difficulties resulted, in large part, from a lack of techniques for analyzing the knowledge representations learned by neural networks. We explored a number of techniques for identifying knowledge representations, including graphing approximated functions and various techniques for reducing the dimensionality of network contributions. Network contributions are the products of sending unit activations and output-side connection weights. Doing a Principal Components Analysis of network contributions provides useful insights into how neural networks learn to solve problems. These network representations can then be compared to human knowledge representations at various stages of learning.
Unlike most neural networks, people rarely learn from scratch. Instead, people are likely to retrieve and modify their existing knowledge to deal with new problems. This tendency to rely on existing knowledge in part explains why people sometimes learn complex tasks so rapidly and why their learning is often biased in particular ways. Neural networks are useful devices for exploring the complex relations between knowledge and learning. To study such issues, we developed a new algorithm, called knowledge-based cascade-correlation that is able to recruit previously learned sub-networks in the service of new learning. This use of existing knowledge typically speeds learning and sometimes makes learning possible.
Unlike most neural networks, people rarely learn from scratch. Instead, people are likely to retrieve and modify their existing knowledge to deal with new problems. This tendency to rely on existing knowledge in part explains why people sometimes learn complex tasks so rapidly and why their learning is often biased in particular ways. Neural networks are useful devices for exploring the complex relations between knowledge and learning. To study such issues, we developed a new algorithm, called knowledge-based cascade-correlation that is able to recruit previously learned sub-networks in the service of new learning. This use of existing knowledge typically speeds learning and sometimes makes learning possible.
