Measuring the interpretability of unsupervised representations via quantized reversed probing
Self-supervised visual representation learning has recently attracted significant research interest. While a common way to evaluate self-supervised representations is through transfer to various downstream tasks, we instead investigate the problem of measuring their interpretability, i.e. understand...
| Main Authors: | , , |
|---|---|
| Format: | Conference item |
| Language: | English |
| Published: |
OpenReview
2021
|
| _version_ | 1826307688704770048 |
|---|---|
| author | Laina, I Asano, YM Vedaldi, A |
| author_facet | Laina, I Asano, YM Vedaldi, A |
| author_sort | Laina, I |
| collection | OXFORD |
| description | Self-supervised visual representation learning has recently attracted significant research interest. While a common way to evaluate self-supervised representations is through transfer to various downstream tasks, we instead investigate the problem of measuring their interpretability, i.e. understanding the semantics encoded in raw representations. We formulate the latter as estimating the mutual information between the representation and a space of manually labelled concepts. To quantify this we introduce a decoding bottleneck: information must be captured by simple predictors, mapping concepts to clusters in representation space. This approach, which we call reverse linear probing, provides a single number sensitive to the semanticity of the representation. This measure is also able to detect when the representation contains combinations of concepts (e.g., "red apple'') instead of just individual attributes ("red'' and "apple'' independently). Finally, we propose to use supervised classifiers to automatically label large datasets in order to enrich the space of concepts used for probing. We use our method to evaluate a large number of self-supervised representations, ranking them by interpretability, highlight the differences that emerge compared to the standard evaluation with linear probes and discuss several qualitative insights. Code at: https://github.com/iro-cp/ssl-qrp. |
| first_indexed | 2024-03-07T07:06:52Z |
| format | Conference item |
| id | oxford-uuid:dfdcf107-438a-4779-bd0c-e228db70bc61 |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-03-07T07:06:52Z |
| publishDate | 2021 |
| publisher | OpenReview |
| record_format | dspace |
| spelling | oxford-uuid:dfdcf107-438a-4779-bd0c-e228db70bc612022-05-13T11:09:11ZMeasuring the interpretability of unsupervised representations via quantized reversed probingConference itemhttp://purl.org/coar/resource_type/c_5794uuid:dfdcf107-438a-4779-bd0c-e228db70bc61EnglishSymplectic ElementsOpenReview2021Laina, IAsano, YMVedaldi, ASelf-supervised visual representation learning has recently attracted significant research interest. While a common way to evaluate self-supervised representations is through transfer to various downstream tasks, we instead investigate the problem of measuring their interpretability, i.e. understanding the semantics encoded in raw representations. We formulate the latter as estimating the mutual information between the representation and a space of manually labelled concepts. To quantify this we introduce a decoding bottleneck: information must be captured by simple predictors, mapping concepts to clusters in representation space. This approach, which we call reverse linear probing, provides a single number sensitive to the semanticity of the representation. This measure is also able to detect when the representation contains combinations of concepts (e.g., "red apple'') instead of just individual attributes ("red'' and "apple'' independently). Finally, we propose to use supervised classifiers to automatically label large datasets in order to enrich the space of concepts used for probing. We use our method to evaluate a large number of self-supervised representations, ranking them by interpretability, highlight the differences that emerge compared to the standard evaluation with linear probes and discuss several qualitative insights. Code at: https://github.com/iro-cp/ssl-qrp. |
| spellingShingle | Laina, I Asano, YM Vedaldi, A Measuring the interpretability of unsupervised representations via quantized reversed probing |
| title | Measuring the interpretability of unsupervised representations via quantized reversed probing |
| title_full | Measuring the interpretability of unsupervised representations via quantized reversed probing |
| title_fullStr | Measuring the interpretability of unsupervised representations via quantized reversed probing |
| title_full_unstemmed | Measuring the interpretability of unsupervised representations via quantized reversed probing |
| title_short | Measuring the interpretability of unsupervised representations via quantized reversed probing |
| title_sort | measuring the interpretability of unsupervised representations via quantized reversed probing |
| work_keys_str_mv | AT lainai measuringtheinterpretabilityofunsupervisedrepresentationsviaquantizedreversedprobing AT asanoym measuringtheinterpretabilityofunsupervisedrepresentationsviaquantizedreversedprobing AT vedaldia measuringtheinterpretabilityofunsupervisedrepresentationsviaquantizedreversedprobing |