ML-Enabled IoT Devices and Embedded AI

Research Partners

NespressoNestl Nespresso S.A.
RichemontRichemont
ClearSpaceClearSpace SA


Description

Machine learning is becoming increasingly ubiquitous. In the last years we have seen an unprecedented increase in the recognition accuracy of neural networks, which can, in some cases, even beat human beings in pattern recognition tasks. However, the huge processing requirements of neural networks have traditionally confined them to cloud servers. Such is the case, for example, of the speech recognition capabilities of some well known digital assistants in smart phones.

The cloud-centric approach faces serious network latency and bandwidth limitations, aside from requiring permanent coverage. Additionally, it also creates pressure on processing resources at the cloud side —where data from all the devices must be processed— and may even raise privacy concerns.

By moving part of the recognition and classification capabilities to the (embedded) devices, some of these issues can be solved.

At ESL we are designing innovative architectures and accelerators to efficiently implement neural networks into resource constrained embedded devices and IoT sensors. Some of the research avenues under exploration include hardware solutions such as multiprocessing and custom accelerators, or software modifications such as SIMD exploitation and quantization of the weights and parameters of the elements that compose the neural network.

These efforts will help to reduce network bandwidth usage and response latency, which may in some cases be quite relevant, and to create truly autonomous devices capable of operating even in cases of network unavailability. Additionally, the pressure for increased processing power at the data center side will be relieved, easing some of their performance and energy consumption challenges through a more distributed hierarchical recognition and classification model.

 

Related news:

 

Related Publications

E2CNN: Ensembles of Convolutional Neural Networks to Improve Robustness Against Memory Errors in Edge-Computing Devices
Ponzina, Flavio; Peon Quiros, Miguel; Burg, Andreas Peter; Atienza Alonso, David
2021IEEE - Transactions on ComputersPublication funded by Compusapien (Next-gen computing systems inspired by the human brain)Publication funded by WiPLASH H2020 (New on-chip wireless communication plane)Publication funded by SNF ML-edge (ML-edge: Enabling Machine-Learning-Based Health Monitoring in Edge Sensors via Architectural Customization)