Explainable Trustworthy brain-like AI for Data Intensive Applications

Project Information

EXTRA-BRAIN

Grant agreement ID: 101135809

DOI

10.3030/101135809

EC signature date 22 November 2023

Start date 1 January 2024

End date 28 February 2027

Funded under

• Digital, Industry and Space

Total cost

€ 4 994 312,50

EU contribution

€ 4 994 312,50

Coordinated by

KUNGLIGA TEKNISKA HOEGSKOLAN
Sweden

Periodic Reporting for period 1 - EXTRA-BRAIN (Explainable Trustworthy brain-like AI for Data Intensive Applications)

Reporting period: 2024-01-01 to 2025-06-30

Summary of the context and overall objectives of the project

A fast spreading adoption of AI in wide-ranging industrial sectors reflects its potential to enhance productivity and economic efficiency. The main technological engine behind AI driven innovation has been a large family of deep neural networks (DNNs), more recently empowered by generative approaches. While generative AI opens up new opportunities for innovation far beyond the industrial domain, AI applications stimulate ever-increasing demands in computing and data among others, resulting in excessive need for resources with unreasonable carbon footprint for industrial deployments. This is particularly challenging for applications with real-time constraints yet requiring functionality, robust adaptivity and operational flexibility, especially in resource limited environments. Current DNNs do not offer a robust and sustainable solution for systems that require low-power consumption, dynamic scalability, adaptation to changing conditions, sustainable scalability, data efficient learning and deployment flexibility in the edge-cloud continuum (ECC). In the attempt to find a manageable compromise between these objectives and requirements, today’s DNNs are also vulnerable to various threats that undermine user trust and reliability. It is therefore strategically important to develop and deploy new core AI technologies that can be more sustainably scaled and utilised in resource constrained application domains, thereby driving the innovation potential of European industries. EXTRA-BRAIN intends to respond to this strategic demand by promoting a brain-like computing paradigm as a technological backbone of the next-generation AI systems. Our mission is to develop brain-like neural networks (BLNNs) building on principles of neural information processing in the human brain and exploiting low-power neuromorphic implementations with the aim of providing AI solutions that are resource efficient, still operationally robust, functionally flexible, trustworthy, and with potential for deployments in the ECC. To further align with the ambition for a sustainable and trustworthy AI the project also integrates application dependent data processing pipelines and explores how different explainability techniques promote trust in EXTRA-BRAIN’s solutions. Finally, to demonstrate the innovative potential, human centricity and real-world applicability of the proposed brain-like AI approach we deploy and systematically evaluate the EXTRA-BRAIN’s AI methods in a range of use case (UC) scenarios spanning digital finance, telecommunication and search-and-rescue (SAR) robotics.

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

The main technical and scientific achievements of EXTRA-BRAIN so far revolve around key advancements of the core AI technology and the EXTRA-BRAIN’s AI ecosystem. To start with, BLNNs building on the KTH Bayesian Confidence Propagating Neural Networks (BCPNNs) have been developed and evaluated to demonstrate their functional capabilities, robustness and learning efficiency. Since their scalability due to the architectural modularity and locality of learning render them suitable for low-power neuromorphic implementations, we are systematically evaluating the performance and energy footprint of different BCPNN deployments on FPGA hardware. Our co-design activities help optimise the resulting neuromorphic artefacts by aligning algorithmic development with hardware implementation. The co-design process has also paved the way for building prototype BLNNs suited for robotic vision tasks in SAR robotics UC and edge user allocation (EUA) in the telecommunication domain. In the digital finance UC we have made effort on integrating risk estimation tools and sentiment analysis for personalized investment decisions with focus on reducing data requirements and adaptability using more traditional deep learning. In the context of the EXTRA-BRAIN’s UCs we have systematically collated user requirements, reviewed key dimensions for application specific trustworthy AI operations including explainability aspects and started developing data pipelines tailored to concrete UC requirements.
In parallel we have designed the architecture of the EXTRA-BRAIN AI system and devised a co-creation approach guiding deployment in concrete applications. To facilitate flexible allocation of computing resources across a distributed infrastructure, we have made key steps towards building an orchestration layer supporting ECC deployments. Next we will intensify our efforts on integrating the existing architectural components and execute our co-creation strategy for securing trustworthy AI in UCs.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

Most of the results reported so far represent the scientific progress in core brain-like AI technology beyond the state of the art, as reflected in the project’s scientific publications. Moreover, we have devised novel brain-like computing approaches to concrete problems in the EXTRA-BRAIN’s UC domains: a novel BLNN method for EUA problem with a neuromorphic implementation for industrial validation (pending patent application), and another BLNN prototype for perception capabilities of a four-legged robot in SAR scenarios. The use of neuromorphic computing for “intelligent” perception has strong innovative value in autonomous SAR robotics. Also, integrating risk estimation tools and sentiment analysis for personalized investment decisions with federated learning capabilities offers attractive prospects for innovation in digital finance. Other ongoing developments also show potential for generating impact like EXTRA-BRAIN’s AI platform with flexible deployments in ECC capitalising on adaptive orchestration layers and resource efficient neuromorphic implementation. We see potential in generated volumes of image and video datasets in SAR robotics, where demands for publicly available data and benchmarks are rapidly expanding. We intend to make our synthetic data available and propose a new computer vision benchmark in this fast growing field.
In summary, currently the most tangible results beyond the state of the art are of scientific nature and represent a significant research contribution in brain-like AI, edge AI, neuromorphic computing and in application domains linked with specific UCs. We still advocate for more research to enhance the operational capabilities, robustness of our AI methods, paving the way for demonstrations in industrial scenarios and higher TRL. We recognise the importance of patenting, especially with future commercialisation prospects. There are opportunities for the EXTRA-BRAIN’s AI technology to become relevant in a wide range of industrial sectors, particularly in areas where robust resource efficient computing with limited access to bulks of annotated data, online task learning capabilities, continuous adaptation and flexible autonomy are in demand (towards sustainable AI systems). Further market analysis will help us identify more specific business opportunities.

erf-2024-extra-brain.jpg