What is it about?

Convolutional Neural Networks (CNNs) achieved state-of-the-art accuracy in many applications. Unfortunately their complexity prevent an effective deployment in portable devices. This work introduces a practical solution to design CNNs able to adapt to the context and scale their computational effort depending on the resources available. The proposed solution exploits the hierarchical organization of concepts that we, as humans, use during our everyday activities.

Featured Image

Why is it important?

The proposed strategy is well suited for embedded AI technologies. It enables the portability of Convolutional Neural Networks on resource constrained IoT end-nodes. This is achieved by transforming static CNNs into adaptive, hence flexible, models that adapt their energy consumption. The results demonstrate remarkable improvements.

Perspectives

We hope this article will inspire new practical uses of Convolutional Neural Network. The availability of a lightweight mechanism that make CNNs able to play with different levels of complexity is something that might improve the smartness of IoT embedded systems.

Andrea Calimera
Politecnico di Torino

Read the Original

This page is a summary of: Scalable-effort ConvNets for multilevel classification, January 2018, ACM (Association for Computing Machinery),
DOI: 10.1145/3240765.3240845.
You can read the full text:

Read

Resources

Contributors

The following have contributed to this page