Rewriting the Rules of a Classifier
Observations of various deep neural network architectures indicate that deep networks may be spontaneously learning representations of concepts with semantic meaning, and encoding a relational structure or rule between these concepts. We refer to these encoded relationships between concepts in the n...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2022
|
| Online Access: | https://hdl.handle.net/1721.1/139151 |
| _version_ | 1826216388884168704 |
|---|---|
| author | Elango, Mahalaxmi |
| author2 | Torralba, Antonio |
| author_facet | Torralba, Antonio Elango, Mahalaxmi |
| author_sort | Elango, Mahalaxmi |
| collection | MIT |
| description | Observations of various deep neural network architectures indicate that deep networks may be spontaneously learning representations of concepts with semantic meaning, and encoding a relational structure or rule between these concepts. We refer to these encoded relationships between concepts in the network as rules. In classifiers, we rewrite an existing rule in the network as desired, referred to as the rewriting technique.
We demonstrate that using our rewriting technique and simple human knowledge about how to classify the world around us, we can generalize existing classes to unseen variants, identify spurious correlations present in the dataset, mitigate the effects of spurious correlations, and introduce new classes. We find that our technique reduces the need for: computing resources, because we only re-train a single layer’s weights; new training images, because our rewriting technique can rewrite using concepts already encoded in the network; and domain knowledge, because what we choose to edit to improve classification is derived from logical rules a human would construct to classify images. |
| first_indexed | 2024-09-23T16:46:52Z |
| format | Thesis |
| id | mit-1721.1/139151 |
| institution | Massachusetts Institute of Technology |
| last_indexed | 2024-09-23T16:46:52Z |
| publishDate | 2022 |
| publisher | Massachusetts Institute of Technology |
| record_format | dspace |
| spelling | mit-1721.1/1391512022-01-15T03:02:54Z Rewriting the Rules of a Classifier Elango, Mahalaxmi Torralba, Antonio Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science Observations of various deep neural network architectures indicate that deep networks may be spontaneously learning representations of concepts with semantic meaning, and encoding a relational structure or rule between these concepts. We refer to these encoded relationships between concepts in the network as rules. In classifiers, we rewrite an existing rule in the network as desired, referred to as the rewriting technique. We demonstrate that using our rewriting technique and simple human knowledge about how to classify the world around us, we can generalize existing classes to unseen variants, identify spurious correlations present in the dataset, mitigate the effects of spurious correlations, and introduce new classes. We find that our technique reduces the need for: computing resources, because we only re-train a single layer’s weights; new training images, because our rewriting technique can rewrite using concepts already encoded in the network; and domain knowledge, because what we choose to edit to improve classification is derived from logical rules a human would construct to classify images. M.Eng. 2022-01-14T14:52:58Z 2022-01-14T14:52:58Z 2021-06 2021-06-17T20:13:11.250Z Thesis https://hdl.handle.net/1721.1/139151 In Copyright - Educational Use Permitted Copyright MIT http://rightsstatements.org/page/InC-EDU/1.0/ application/pdf Massachusetts Institute of Technology |
| spellingShingle | Elango, Mahalaxmi Rewriting the Rules of a Classifier |
| title | Rewriting the Rules of a Classifier |
| title_full | Rewriting the Rules of a Classifier |
| title_fullStr | Rewriting the Rules of a Classifier |
| title_full_unstemmed | Rewriting the Rules of a Classifier |
| title_short | Rewriting the Rules of a Classifier |
| title_sort | rewriting the rules of a classifier |
| url | https://hdl.handle.net/1721.1/139151 |
| work_keys_str_mv | AT elangomahalaxmi rewritingtherulesofaclassifier |