Dissecting Pruned Neural Networks (1907.00262v1)

Abstract: Pruning is a standard technique for removing unnecessary structure from a neural network to reduce its storage footprint, computational demands, or energy consumption. Pruning can reduce the parameter-counts of many state-of-the-art neural networks by an order of magnitude without compromising accuracy, meaning these networks contain a vast amount of unnecessary structure. In this paper, we study the relationship between pruning and interpretability. Namely, we consider the effect of removing unnecessary structure on the number of hidden units that learn disentangled representations of human-recognizable concepts as identified by network dissection. We aim to evaluate how the interpretability of pruned neural networks changes as they are compressed. We find that pruning has no detrimental effect on this measure of interpretability until so few parameters remain that accuracy beings to drop. Resnet-50 models trained on ImageNet maintain the same number of interpretable concepts and units until more than 90% of parameters have been pruned.

• The paper demonstrates that pruning up to 90% of parameters preserves interpretability with minimal impact on accuracy.
• The analysis employs Network Dissection on ResNet-50 models, confirming stable recognition of human-recognizable concepts post-pruning.
• The study uses sparse pruning and lottery ticket fine-tuning, suggesting that efficient model deployment is achievable without sacrificing clarity.

Dissecting Pruned Neural Networks: An In-Depth Analysis

This paper, entitled "Dissecting Pruned Neural Networks," by Jonathan Frankle and David Bau, explores the implications of pruning on the interpretability of neural networks. Specifically, it examines whether the reduction of unnecessary network parameters, a common practice for efficiency gains, affects the model's ability to maintain interpretable units that capture human-recognizable concepts. This research primarily investigates Resnet-50 models trained on ImageNet, using Network Dissection techniques to quantify interpretability.

Pruning and Interpretability

Pruning in neural networks refers to the practice of removing parameters deemed unnecessary, thereby reducing the computational and storage overhead without sacrificing the model’s accuracy. The conventional understanding is that many networks contain a large number of redundant parameters post-training. The authors challenge this assumption by exploring whether these redundant parameters also contribute to the network’s interpretability or if they can be pruned without losing the ability to understand the model’s underlying concepts.

The paper postulates three hypotheses regarding the relationship between pruning and interpretability:

1. No Relationship: Pruning does not significantly alter the interpretability until it compromises accuracy.
2. Improved Interpretability: Pruning enhances interpretability by removing noise and highlighting more critical components.
3. Reduced Interpretability: Pruning diminishes interpretability by compressing representations and obscuring human-recognizable concepts.

Research Findings

The findings from multiple experiments revealed a strong correlation supporting Hypothesis A: pruning does not substantially affect interpretability until it impacts the model’s accuracy. Specifically, the paper showed that Resnet-50 retains its interpretable nature until more than 90% of the parameters are pruned. This finding indicates that the parameters removed through pruning are largely extraneous with respect to the network’s interpretability, as well as its accuracy.

The detailed analysis of interpretability involved assessing the number of interpretable units and the total number of concepts learned by pruned models. Network dissection revealed nearly identical numbers of interpretable concepts and units across various levels of pruning, up to a threshold determined by accuracy loss.

Methodological Insights

The authors utilized a sparse pruning technique, removing weights based on magnitude at the end of training. The pruned networks were then subjected to fine-tuning using a lottery ticket approach, which resets weights to early values in training. This particular method of fine-tuning allows the network to potentially learn new representations suited to its reduced size, which the paper found beneficial in maintaining both accuracy and interpretability.

An intriguing aspect of the research was its examination of the consistency of concept recognition in pruned versus original networks. A significant portion of units continued to recognize the same concepts post-pruning, evidencing stable interpretability across pruning iterations.

Implications and Future Directions

The practical implications of these findings suggest robust avenues for deploying pruned networks without losing interpretability, beneficial for edge computing or environments with limited resources. Beyond practical applications, the results pose theoretical questions regarding the nature of network redundancy and the potential for more efficient model training paradigms that are inherently interpretable.

Future research could extend these findings by exploring different pruning methods, such as structured pruning that removes entire filters, or alternative fine-tuning strategies. Moreover, the transferability of these results across different architectures beyond Resnet-50 could provide additional insights into the generalizability of pruning effects on interpretability.

In summary, this paper makes a substantial contribution to the understanding of neural network pruning, offering a nuanced perspective that separates the concerns of accuracy and interpretability—a critical step towards building efficient yet interpretable AI systems.