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/

Zendy Plus

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

Zendy Tools

Zendy Open

Model compression has emerged as an important area of research for deployingdeep learning models on Internet-of-Things (IoT). However, for extremelymemory-constrained scenarios, even the compressed models cannot fit within thememory of a single device and, as a result, must be distributed across multipledevices. This leads to a distributed inference paradigm in which memory andcommunication costs represent a major bottleneck. Yet, existing modelcompression techniques are not communication-aware. Therefore, we proposeNetwork of Neural Networks (NoNN), a new distributed IoT learning paradigm thatcompresses a large pretrained 'teacher' deep network into several disjoint andhighly-compressed 'student' modules, without loss of accuracy. Moreover, wepropose a network science-based knowledge partitioning algorithm for theteacher model, and then train individual students on the resulting disjointpartitions. Extensive experimentation on five image classification datasets,for user-defined memory/performance budgets, show that NoNN achieves higheraccuracy than several baselines and similar accuracy as the teacher model,while using minimal communication among students. Finally, as a case study, wedeploy the proposed model for CIFAR-10 dataset on edge devices and demonstratesignificant improvements in memory footprint (up to 24x), performance (up to12x), and energy per node (up to 14x) compared to the large teacher model. Wefurther show that for distributed inference on multiple edge devices, ourproposed NoNN model results in up to 33x reduction in total latency w.r.t. astate-of-the-art model compression baseline.

Memory- and Communication-Aware Model Compression for Distributed Deep Learning Inference on IoT