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

Many image processing tasks involve image-to-image mapping, which can beaddressed well by fully convolutional networks (FCN) without any heavypreprocessing. Although empirically designing and training FCNs can achievesatisfactory results, reasons for the improvement in performance are slightlyambiguous. Our study is to make progress in understanding their generalizationabilities through visualizing the optimization landscapes. The visualization ofobjective functions is obtained by choosing a solution and projecting itsvicinity onto a 3D space. We compare three FCN-based networks (two existingmodels and a new proposed in this paper for comparison) on multiple datasets.It has been observed in practice that the connections from the pre-pooledfeature maps to the post-upsampled can achieve better results. We investigatethe cause and provide experiments to shows that the skip-layer connections inFCN can promote flat optimization landscape, which is well known to generalizebetter. Additionally, we explore the relationship between the modelsgeneralization ability and loss surface under different batch sizes. Resultsshow that large-batch training makes the model converge to sharp minimizerswith chaotic vicinities while small-batch method leads the model to flatminimizers with smooth and nearly convex regions. Our work may contribute toinsights and analysis for designing and training FCNs.

Visualized Insights into the Optimization Landscape of Fully Convolutional Networks