From home to work, artificial intelligence has made in roads into virtually every aspect of our lives. It has transformed how we relate to others, do our jobs, and interact with the devices we use every day. AI is also reshaping entire sectors of the economy. In some—health, mobility, energy, and the environment—AI brings hope of unprecedented benefits. However, there are fundamental questions around AI that will have to be answered if the technology is to be deployed in industrial scenarios and, more broadly, accepted by society.
Artificial intelligence, with its capacity to analyze huge volumes of data at lightning speed and its startlingly wide array of potential applications, is a technological powerhouse. Different AIs—the technology comes in many forms—can be combined in a multitude of ways to solve all sorts of problems. AIs give autonomous vehicles their environmental perception and trajectory calculation capabilities, summarize text and translate it into multiple languages, and even interact with people. They are used to model physical phenomena, optimize complex processes, and predict events. And artificial intelligence is a formidable tool not only for innovation, but also for scientific research, pushing back the limits of what is possible. CEA-List has built on its long-standing image analysis, signal processing, and language processing research, developing deep knowledge of the AI technologies that underpin these three disciplines. The institute is ideally positioned to bring its partner companies effective responses to their challenges in health, transportation, cybersecurity, the Factory of the Future, and environmental technologies.
Environmental perception, scene analysis, object detection, image classification, robotics, etc.
Machine translation, automated information extraction, content generation, text summarization, chatbots, etc.
Predictive analytics, anomaly detection, audio/voice analysis, cyber intelligence, interpretation of physical measurements (electrical, vibration, etc.)
Complex simulation, flow and process optimization, etc.
In partnership with automotive supplier Valeo, CEA-List developed a 3D, 360° perception system that enables a vehicle to detect pedestrians, other vehicles, and any other element in its environment in real time. The system uses data gathered from the vehicle’s on-board fish-eye cameras. CEA-List’s Deep Manta algorithm, a multi-task neural network, was modified and enhanced to meet the detection needs of this project.
Although AI has much to offer, consumers remain reticent and deployment in industrial scenarios has yet to gain any real traction. The unanswered questions preventing the widespread adoption of AI mainly center around deep learning using large neural networks —currently the most efficient and, therefore, the most attractive type of AI. Neural networks are complex, making it difficult to describe how they actually work or what they base their answers to a given problem on.
“Black box” AIs of this type are problematic in scenarios where explainability is needed. Imagine medical diagnostics or credit score calculations, which directly affect individuals, for example, or decision-support AIs that could trigger immediate action in high-stakes professional contexts. But the need for transparent, explainable AI isn’t the only deal breaker for users. There are also the issues of biased training data and data privacy.
Industrial AI use cases come with their own set of demands in terms of quality, robustness, safety, and security, and new methods will be needed to respond effectively. An AI is, by definition, a learning system. And a system that learns changes constantly. Therefore, any new methods must be dynamic if they are to ensure that an AI continues to behave as expected.
Greater trust in AI and in deep learning is a prerequisite to the widespread deployment of these technologies in industrial scenarios. CEA-List is addressing this need by developing new methods and techniques to support trusted AI, and by contributing to trusted AI initiatives in France (the Confiance.ai consortium; Formal verification and AI, with Paris-Saclay University) and internationally (AI Safety, Waise, Safe AI). The institute is also engaged in EU projects Tailor, with 54 partners, and Starlight, which it coordinates.
CEA-List recently worked on a promising project to enable the formal validation of neural networks used for image recognition. To learn more, read this article from 24/09/2020.
Until recently, AI has been deployed on cloud architectures primarily addressing B2C use cases. Industrial AI, with its much higher latency and data privacy requirements, is driving the emergence of a new paradigm —Edge AI. Moving computing off the cloud and closer to devices and sensors does present one major challenge, however. AI is intensive in terms of processing, memory, and power, making the technology difficult to deploy on highly constrained Edge devices without negatively impacting performance. CEA-List is bringing its experience with embedded systems to these new challenges.
The institute is tackling both hardware and algorithms with two neural network accelerators, PNeuro and DNeuro, developed specifically for embedded architectures, and with the N2D2 environment, which aids in the development of more compact AI models. Ultimately, the goal is to move learning off the cloud as well, for more efficient processing and enhanced data security.
We now know how to reduce the size of AI models by a factor of between 10 and 100. This will open up a whole new range of potential applications for AI. Imagine a satellite with its own image processing capabilities that sends only those images relevant to what it was instructed to search for back to Earth. It would be a totally new paradigm likely to cost much less, too.
Dolphin Design and CEA-List have set up a joint R&D lab on embedded systems to develop an Edge AI solution for IoT offering software flexibility, energy efficiency, and performance. The new solution will be built on CEA-List’s PNeuro hardware accelerator and Dolphin Design’s processing platform.
LEN (Lifelong Exploratory Navigation) software and the EDNA* (Exploratory Digraph Navigation Using A*) algorithm are frugal enough in terms of processing power to be embedded into mobile robots, enabling the robots to move around and explore a space autonomously. LEN and EDNA* are built on machine learning and meet the safety requirements of this kind of use case.
A headline-making 2019 study out of the University of Massachusetts Amherst revealed the dramatic environmental costs of creating an AI. According to the study, the carbon footprint of training a single AI (BERT, a leading natural language processing model) is equivalent to that of a round-trip trans-America flight.
Natural language processing AIs are known to be particularly complex. At the same time, the growth of AI computing budgets—10x per year—continues to outpace improvements in energy efficiency. Digital technology is estimated to account for nearly 4% of total global carbon emissions. While deep learning algorithms currently represent just a tiny fraction of total tech-related emissions, the time to address energy efficiency and society’s concerns about the mushrooming environmental costs of AI is now.
Tomorrow’s AI—and the hardware architectures and data it is built on—will have to be frugal across the entire lifecycle, from development to use. CEA-List’s research in this area echoes what the institute is doing with embedded AI. Here, CEA-List is developing low-data-budget training methods, more energy efficient application design environments, and optimized hardware architectures.
In certain situations, AI training data may not be available in sufficient quantities. In others, sharing data may not be desirable. CEA-List is developing new methods for these scenarios, training AIs on limited and simulated data using generative adversarial networks, distributed AI, and one-shot and few-shot learning, for example.
EU sovereignty on the fundamental hardware and software technologies that will underpin the development of innovative AI and data-driven services is an imperative. The EU’s economic and industrial independence depends on it. Europe will not be able to prepare effectively for its future or fend off cyber threats without it. In this challenge lies an opportunity to create an alternative vision of AI—one that aligns with the EU’s ethical and societal values.
Initiatives are cropping up across France and all of Europe to address sovereignty. These informal communities, research institutes, and other organizations are sharing research resources, creating new knowledge, educating junior scientists, and breaking down silos between science and business. Their overriding objectives are to secure the development and production of innovative solutions on a broad scale and return to a position of leadership on key technologies like trusted AI.
CEA-List is a stakeholder in these initiatives. Confiance.ai is a French national initiative that aims to bring AI to critical products and services. It is the technological pillar of a broader French government strategy on AI-based system security, reliability, and certification launched in January 2021. CEA-List is helping develop sovereign trusted AI solutions.