What does a pruned deep neural network "forget"?
Very excited to share our recent work w Aaron Courville, Yann Dauphin and @DreFrome
weightpruningdamage.github.io
Very excited to share our recent work w Aaron Courville, Yann Dauphin and @DreFrome
weightpruningdamage.github.io
At face value, deep neural network pruning appears to promise you can (almost) have it all — remove the majority of weights with minimal degradation to top-1 accuracy. In this work, we explore this trade-off by asking whether certain classes are disproportionately impacted.
We find that pruning is better described as "selective brain damage" -- performance on a tiny subset of classes and images is cannibalized in order to preserve overall performance. The interesting part is what makes certain images more likely to be forgotten...
The images impacted most by pruning are far more challenging for the original model. As model capacity is varied, the model appears to "forget" the examples it was already uncertain about. For real world datasets, this suggests caution should be used before pruning.
At the very least, we should consider additional measures to articulate this trade-off beyond just top-1 accuracy. :) We propose some tools to start thinking about this problem, for example PIE surfaces a subset of impacted images for a human expert to inspect...
Inspecting PIE images reveals a lot about the mapping the model learns. This may also explain why I have shared so many poorly structured ImageNet test set images over the last few months. :) Many PIE ImageNet images are multi-object or challenging for a human...
After this project, I am more convinced that a good data cleaning of ImageNet (and less rigid treatment of this a single object classification task) may compensate for much of the capacity we have thrown at the problem. But... that is a topic for another paper. :)
On real world data, the stakes are often higher than correctly classifying “guacamole”. For tasks like patient risk stratification, understanding the features pruning is cannibalizing is crucial. I hope our work highlight the need for better tooling to understand these tradeoffs.
