What Do Compressed Deep Neural Networks Forget

By Sara Hooker, Aaron C. Courville, Gregory Clark, Yann Dauphin, Andrea Frome. 2019

221 citations recorded by the source corpus.

Deep neural network pruning and quantization techniques have demonstrated it is possible to achieve high levels of compression with surprisingly little degradation to test set accuracy. However, this measure of performance conceals significant differences in how different classes and images are impacted by model compression techniques. We find that models with radically different numbers of weights have comparable top-line performance metrics but diverge considerably in behavior on a narrow subset of the dataset. This small subset of data points, which we term Pruning Identified Exemplars (PIEs) are systematically more impacted by the introduction of sparsity. Compression disproportionately impacts model performance on the underrepresented long-tail of the data distribution. PIEs over-index on atypical or noisy images that are far more challenging for both humans and algorithms to classify. Our work provides intuition into the role of capacity in deep neural networks and the trade-offs incurred by compression. An understanding of this disparate impact is critical given the widespread deployment of compressed models in the wild.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Learning with Recoverable Forgetting (2022, neutral): “…(20% parameters of the original network); the other is for knowledge filtering that pruning would be better described as ‘selective knowledge damage’ [16], where only the activated filters are kept such that we only deposit the sample-specific knowledge of D r rather than the whole knowledge.”
Characterising Bias in Compressed Models (2020, background): “In additional to the measure of divergence proposed by proposed by Hooker et al. (2019b) which we term Modal CIE , we consider an additional measure of divergence Taxicab CIE .”
Moving beyond “algorithmic bias is a data problem” (2021, background): “11,12 These are not the only design choices that matter—even more subtle choices like learning rate and length of training can also disproportionately impact error rates on the long-tail of the data-set.”
Metadata Archaeology: Unearthing Data Subsets by Leveraging Training Dynamics (2022, methodology): “Prior work has developed a rough taxonomy of data properties, or metadata which different examples might exhibit, including but not limited to: noisy [68, 71, 62, 63], atypical [25, 10, 21, 60], challenging [24, 3, 8, 49, 2], prototypical or core subset selection [49, 55, 56, 27] and out-of-distribu”
Intriguing Properties of Compression on Multilingual Models (2022, background): “This ﬁnding disagrees with some prior works (Du et al., 2021a; Hooker et al., 2019; Sehwag et al., 2019) which ﬁnd sparsity erodes different measures of robustness.”
How Well Do Sparse ImageNet Models Transfer? (2021, background): “Yet, we wish to highlight recent work [29, 30, 44] which examines the robustness of pruned models to input perturbations, as well as the impact of pruning on accuracy of specific segments of the data.”
Randomness In Neural Network Training: Characterizing The Impact of Tooling (2021, background): “Trade-off with fairness objectives Recent work [Hooker, 2021, Yona et al., 2021, D’Amour et al., 2020, Hooker et al., 2019] has identiﬁed that models with similar top-line metrics can evidence unacceptable performance on subsets of the distribution.”
FAIR-Ensemble: When Fairness Naturally Emerges From Deep Ensembling (2023, methodology): “…focus on datasets collection and curation (Barocas et al., 2019; Zhao et al., 2017; Shankar et al., 2017), with limited work to-date understanding the role of model design or optimization choices on amplifying or curbing bias (Ogueji et al., 2022; Hooker et al., 2019; Balestriero et al., 2022).”
Deep Learning Through the Lens of Example Difficulty (2021, background): “There are a number of definitions of example difficulty in the literature (E.g. see Carlini et al. (2019); Hooker et al. (2019); Lalor et al. (2018); Agarwal and Hooker (2020)).”
The Low-Resource Double Bind: An Empirical Study of Pruning for Low-Resource Machine Translation (2021, background): “Recent work restricted to the computer vision domain has found that sparse models with comparable top-line performance metrics diverge considerably in behavior on the long-tail of the distribution and are sensitive to distribution shifts (Hooker et al., 2020a; Liebenwein et al., 2021).”