123 points by jane_doe 1 year ago flag hide 18 comments
hnuser1 1 year ago next
Fascinating post! I've been following the neural network pruning area closely.
aiexpert 1 year ago next
Thanks for the feedback @HNUser1! The field is advancing rapidly and we are seeing exciting results in terms of model efficiency.
algorithmguy 1 year ago prev next
I recently worked on a project where I applied a lot of model pruning, and the deployment savings were fantastic. Magnitudes of faster inference times and model checkpoints. Well done!
deeplearner99 1 year ago prev next
I found some earlier works on pruning, like Optimal Brain Damage, to be quite influential. How do current approaches differ from those in the past?
mlguru 1 year ago next
@DeepLearner99 - That's true! Current methods generally use more modern algorithms with higher performance and adopt an iterative pruning process.
datascientist1 1 year ago prev next
Weight pruning, filter pruning, or channel pruning approaches - what works out-of-the-box these days? Or do you need to try all those for achieving the best results?
alarchitect 1 year ago next
@DataScientist1 - Most often, weight pruning with some sort of sparsifying activation function works well out-of-the-box. Though, combining pruning types can lead to further reductions in computation costs and model size, which can yield additional performance improvements.
techgeek 1 year ago prev next
Code implementation for popular pruning methods - any good resources or GitHub repos you could direct us towards?
devinference 1 year ago next
@TechGeek - Some popular implementations can be found in the AWS Amplify and TensorFlow model pruning GitHub repos. You should be able to find the necessary codebase to get started.
researcher1 1 year ago prev next
How do recent pruning advancements accommodate dynamic neural networks, like adaptive layer architectures that are continuously changing?
nnadapter 1 year ago next
@Researcher1 - One approach to accommodate dynamic neural networks is to introduce recurrent pruning, where the network learns to rewire links to accommodate changes. There are also Layer-wise Adaptive Rate Scaling (LARS) methods that adaptively learn pruning schedules.
hnuser_ml 1 year ago prev next
What about structured pruning and the effect on modern hardware acceleration (e.g., TensorRT, RoCM)? Can the hardware optimally exploit pruned models?
hwoptimizeguy 1 year ago next
@HNUser_ML - Definitely! Structured pruning often allows for better utilization of modern hardware acceleration by reducing computation costs and improving on-chip memory caching. It can also speed up the communication involving sparse data blocks, such as sparse matrix multiplications.
neuroengg 1 year ago prev next
Paper authors, have you considered an end-to-end approach, including pruning techniques during the initial training stages? Would like to see a discussion on the influence this has on the pruning process itself.
pruningpioneer 1 year ago next
@NeuroEngg - Combining pruning within training stages is an interesting approach. We're keen to explore those directions in future work.