Report

Teaser, summary, work performed and final results

Periodic Reporting for period 1 - ALOHA (software framework for runtime-Adaptive and secure deep Learning On Heterogeneous Architectures)

Teaser

The trend of the market is showing that more than 21 billion of Internet-of-Things (IoT) and embedded devices will be in use in the 2020. Although this growth of devices poses a great business opportunity, some challenges are hindering its wide exploitation. Industry is...

Summary

The trend of the market is showing that more than 21 billion of Internet-of-Things (IoT) and embedded devices will be in use in the 2020. Although this growth of devices poses a great business opportunity, some challenges are hindering its wide exploitation.
Industry is pushing towards Smart Things capable of taking decisions on behalf of the whole system. It is widely accepted that Artificial Intelligence (AI) is the key driver to solve this problem.
Deep Learning, an extremely promising instrument in the machine learning and artificial intelligence landscape. DL algorithms allow achieving very high performance in numerous applications involving recognition, identification and/or classification tasks, however, their adoption, and in general of AI technologies, is hindered by the lack of availability of low-cost and energy-efficient solutions.
Novel algorithm configurations, exploited in different domains, continuously improve the precision of DL systems. However, such advancement comes at the price of significant requirements in terms of processing power. Moreover, while the training phase is typically executed on high-performance computing facilities, recent trends of modern computing landscape push towards an ever-increasing deployment of DL inference on embedded devices. Using such an approach, according to the edge computing paradigm, DL systems may overcome limitations of cloud-based computing, when it comes to latency, bandwidth requirements, security, privacy, and availability. Nevertheless, when DL is moved at the edge, severe performance requirements must coexist with tight constraints in terms of power and energy consumption.
The main goal of ALOHA is to facilitate implementation of DL algorithms on heterogeneous low energy computing platforms providing automation for optimal algorithm selection, resource allocation and deployment.
In the ALOHA project the tool flow is associated to three use cases, where the partners involved as end user provides the requirements of their area of competence. The tool flow, on the other hand, targets three different types of platform, although it is meant to be adaptable to new platforms in the future.

Work performed

The project has reached the half of its duration, and the results obtained in this period are summarized here.
The first use case implementation focuses on the set of inputs needed by the ALOHA Tool flow to perform the expected optimizations. This includes datasets, reference algorithms, target hardware platform definition, constraints on accuracy, performance, security and power that each target deep learning task must satisfy, and the tools developed for the purpose of generating the needed data.
The packages developed in the project have been released as open source packages on GITHUB at the following address:
https://gitlab.com/aloha.eu/

Final results

The impact goal of the project, as stated in the DoA, is to “Reinforce and broaden Europe\'s strong position in low-energy computing by reducing the effort needed to include digital technology inside any type of product or service, including outside the traditional “high-tech” sectors.”

To address this target ALOHA intends “To study and provide methodologies and computer aided design support for effective implementation of Deep Learning algorithms on embedded systems, considering their prospective implementation on low-power computing platforms, helping the developer in all aspects, from application definition to deployment”
Evaluating the impact of ALOHA is directly connected with the actual assessment of the quality of its results on real design cases. This activity, in the first half of the project, has only involved a qualitative estimation of the usability and functional correctness of the tools and utilities that have just been released in an exploitable version. This first assessment confirms the possibility to use the toolflow to obtain the expected effects on the capability of small business and less skilled programmers to use digital technology based on deep learning and low-power heterogeneous architectures.
A more quantitative evaluation will take place from M18, involving measuring of results against the reference performance indicators defined for the exploitation.
Different stakeholders that can be impacted by the project results have been considered:
• The market, in terms of customers that can benefit from the use of the ALOHA toolchain and the related platforms presented (but not limited to them)
• The engineers: considering the persons that in the companies and in the research institutions will use the tools
• The DL architecture: the various DL architectures that are used and developed by the community of developers
• The DL frameworks: the multiple open source frameworks for the development of networks can be highly impacted by the ALOHA toolchain integrating it for the deployment and optimization of the high-level networks on the embedded target platforms.
• The embedded devices, whose adoption and usability can benefit from the toolchain
The actual status of the key results is described extensively in the deliverable D6.4, that is the actual report on the exploitation activities. Several components of the ALOHA toolchain have been already completed (e.g. KR4: CNN‐to‐DataFlow model conversion tool, KR12: Other key exploitation items – NEURAghe reference platform), and others are releases in a first usable version, to be enriched within the integration process (e.g. KR2: DNN approximation tool for parsimonious inference, KR5a: Design Space Exploration engine, KR5b: Training engine, KR8: Post‐Training tool for Parsimonious Inference, KR9: Sesame tool for application partitioning & mapping exploration, KR10: Architecture Optimization Workbench, KR11: Automated adaptive middleware generation/customization).
The integration of the ALOHA toolflow (KR1) is actually capable of evaluating and exploring different algorithm configurations and different ports on three reference architectures. Testing has been performed so far on generic and use-case related datasets.
From the impact point of view, it is important that the adoption of the toolflow by potential stakeholders is properly fostered and encouraged. Some Key Exploitable Results, obtained from the work performed in the different work packages and collected in D6.1, have been released as open-source in the ALOHA Gitlab repository. Adequate dissemination and publicity activities have been carried out (workshops with prospective users, as in Linz 2019) and planned (workshop with prospective users to be held in Cagliari in 2020). We have also spread the word about the project within the computer science community, participating actively to events organized by HIPEAC and to other more industry-related happenings.

Website & more info

More info: https://www.aloha-h2020.eu/.