English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Deep convolutional neural network (CNN) training via iterative optimizationhas had incredible success in finding optimal parameters. However, modern CNNarchitectures often contain millions of parameters. Thus, any given model for asingle architecture resides in a massive parameter space. Models with similarloss could have drastically different characteristics such as adversarialrobustness, generalizability, and quantization robustness. For deep learning onthe edge, quantization robustness is often crucial. Finding a model that isquantization-robust can sometimes require significant efforts. Recent worksusing Graph Hypernetworks (GHN) have shown remarkable performance predictinghigh-performant parameters of varying CNN architectures. Inspired by thesesuccesses, we wonder if the graph representations of GHN-2 can be leveraged topredict quantization-robust parameters as well, which we call GHN-Q. We conductthe first-ever study exploring the use of graph hypernetworks for predictingparameters of unseen quantized CNN architectures. We focus on a reduced CNNsearch space and find that GHN-Q can in fact predict quantization-robustparameters for various 8-bit quantized CNNs. Decent quantized accuracies areobserved even with 4-bit quantization despite GHN-Q not being trained on it.Quantized finetuning of GHN-Q at lower bitwidths may bring further improvementsand is currently being explored.

GHN-Q: Parameter Prediction for Unseen Quantized Convolutional Architectures via Graph Hypernetworks