Name
Abstract | ||
|---|---|---|
Machine learning for extracting case features can provide great benefit over feature engineering for retrieval in poorly understood or hard to characterize domains. The effectiveness of machine learning with deep neural networks has prompted much interest in neural network approaches to feature learning in case-based reasoning, with several works showing the value of feature extraction from input data using convolutional neural networks. Those approaches are based on plausible assumptions about where in the networks to extract features for maximal usefulness. This paper presents an empirical evaluation of those underlying assumptions. We compare three extraction approaches, for an image classification task: the most common feature extraction method, extracting after the convolution layer; a recently proposed alternative, extracting after the densely-connected layers; and a new approach, extracting after the densely-connected layers using multiple networks. Our results show that the latter two approaches substantially increase case retrieval accuracy in example-sparse domains, to which case-based reasoning systems are commonly applied. |
Year | DOI | Venue |
|---|---|---|
2022 | 10.1007/978-3-031-14923-8_6 | International Conference on Case-Based Reasoning |
Keywords | DocType | Citations |
Case-based reasoning,Deep learning,Feature learning,Hybrid systems,Indexing,Integrated systems,Retrieval | Conference | 0 |
PageRank | References | Authors |
0.34 | 0 | 3 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| David B. Leake | 1 | 1369 | 121.60 |
| Wilkerson Zachary | 2 | 0 | 0.34 |
| D. Crandall | 3 | 2111 | 168.58 |