Frugal AI for robotic
From the very beginning with Karel Čapek's play, the first robots were envisioned as androids, mechatronic alter ego of humans. While their form and their movements allow for an anthropomorphic appearance, the robot's perception of its environment, its analysis, and subsequent adaptation of the robot's behavior are major challenges to ensure autonomy and a faithful mimicry of human behavior.
Artificial intelligence offers a generic framework for processing and exploiting signals from sensors: from denoising to recognizing a complex stimulus such as a phrase, a gesture, or a situation. Nevertheless, jointly implementing AI and sensors poses major challenges in terms of latency, memory and computational resources, under the constrain of limited energy budget.
The CEA (Commissariat à l'énergie atomique et aux énergies alternatives) is developing a set of solutions to address these challenges with original approaches that leverage all aspects, from algorithms to microelectronic technologies. This is particularly true for vision, which, among the five human senses, generates the largest volume of data and is used in the most diverse ways: from detection to a semantic analysis of a scene.
We will illustrate the CEA's transversal approach with a few examples of achievements. For instance, to enable wake-up functions, the CEA has developed always-on imagers capable of recognizing objects with power consumption of a few microwatts. To address the computational cost due to the first layers of convolutional networks, the NeuroCorgi circuit relies on a fixed architecture for the initial layers, allowing for 30fps recognition at a few mW. As for gesture processing, i.e. data that requires dealing with sequences of images, the CEA’s solution involves a combination of in-memory computing, 3D IC integration, and highly quantized neural networks. Finally, to deal with the extremes of catastrophic forgetting and catastrophic remembering, we introduce Metaplasticity from Synaptic Uncertainty (MESU). This Bayesian framework features a compact implementation with Ferroelectric Memory Field-Effect Transistors (FeMFETs).
Finally, we will discuss the perspectives opened, on the one hand by the possibility of implementing incremental learning under limited memory resources and on the other hand, by heterogeneous chiplets.
