What Do Compressed Deep Neural Networks Forget? (1911.05248v3)

Abstract: 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.

• The paper reveals that compression methods cause uneven accuracy loss, significantly impacting challenging examples known as Pruning Identified Exemplars (PIEs).
• It identifies that PIEs—often misannotated, low-quality, or complex images—suffer the most pronounced performance drops after pruning.
• The research demonstrates that compressed models become more sensitive to adversarial inputs, underscoring a critical trade-off between efficiency and fairness.

Evaluation of Model Compression Impacts on Neural Network Behavior

The paper "What Do Compressed Deep Neural Networks Forget?" critically examines the implications of pruning and quantization on deep neural networks (DNNs), especially how these compression methods affect specific classes and instances within a dataset. The authors challenge the common notion that top-line performance metrics, such as top-1 accuracy, provide a complete picture of a model's generalization capacity after compression.

Summary of Key Findings

The research reveals several critical insights:

1. Disparate Impact on Data Distribution: Compression affects model accuracy disproportionately across different classes and data points. The paper identifies a subset of data, termed Pruning Identified Exemplars (PIEs), which are particularly vulnerable to performance loss after pruning. These PIEs are statistically shown to diverge significantly from the performance patterns observed in non-compressed models.
2. Characteristics of PIEs: PIEs are frequently images that present classification challenges due to issues such as label misannotation, low quality, multiple objects, or the need for fine-grained classification. These points often reside on the fringes of the data distribution, illustrating that compression might inadvertently prioritize mainstream or frequently occurring data instances.
3. Sensitivity to Adversarial Inputs: Compressing models enhances their sensitivity to distribution shifts such as natural adversarial examples and algorithmically corrupted images. The research highlights that this tendency is more pronounced at higher compression levels.
4. Comparison of Compression Techniques: While all compression strategies examined exhibited some level of disparate impact, quantization was observed to be less disruptive compared to pruning. This finding suggests a potential preference for quantization in applications where minimizing bias or impact on specific data subsets is critical.

Implications and Speculation

The paper speaks to two broad implications for the field:

• Practical Deployment: Given that many deployed DNNs utilize compressed models for efficiency reasons, understanding which portions of data might be adversely affected can inform deployment strategies in sensitive applications, such as healthcare or autonomous systems. Here, a careful balance between model size and performance equity must be struck.
• Theoretical Insights Regarding Model Capacity: These findings contribute to ongoing discussions about the role of model capacity in DNNs. The analysis suggests that excess parameters may play an essential role in capturing nuanced and less frequent patterns in data, which are often critical for robust generalization to unseen examples.

Potential Future Directions

Future research might explore the following avenues:

• Design of Compression Algorithms: Developing new compression algorithms that explicitly account for and aim to reduce disparate impacts on dataset subgroups could be beneficial. Tools for identifying and mitigating impacts on PIEs could be integrated into existing frameworks.
• Cross-Domain Analysis: Extending this analysis to domains such as natural language processing or speech recognition may reveal how compression impacts generalize across different types of data and tasks.
• Fairness and Bias Mitigation: Understanding the intersection between compression and fairness offers fertile ground for additional research, especially given the increasing regulatory attention on AI bias and fairness.

Conclusion: The research pushes the boundaries of understanding the subtleties of neural network compression, offering a framework for assessing the true costs of such efficiency-improving techniques. It invites a re-evaluation of current practices, fostering discussion on how to maintain both performance and fairness in AI systems.