Physics-based Small-scale Hybrid artificial intelligence for low-power edge computing robotic applications

True innovations in artificial intelligence come from exploring alternatives for power-hungry large neural networks such as e.g. Large Language Models (LLMs).  As one of the pioneers of AI,  Yann Lecunn, advised to students: “You should work on next-gen AI systems that lift the limitations of LLMs.” (https://x.com/VivaTech/status/1793310822657032391).

In line with this advice, one of the activities at imec targets low-power in-the-edge AI computations with locally collected data. This includes, for example, the processing and interpretation of sensor information from wearable devices for e.g. health monitoring and biometric authentication. Since the sensor information is collected and processed locally without communication with any network, the system provides a natural secure hub respecting privacy.  A further application that we envision lies in the field of robotics, with highly personalized information driving a prosthesis or an exoskeleton.

To achieve this goal, small-scale hardware-integrated machine learning systems with emphasis on low power performance need to be developed. Efficient implementations can be obtained by exploiting inherent physical phenomena in electronic devices to perform computations, resulting in alternative computing paradigms. Furthermore, this approach requires combining and adapting techniques from different AI algorithms into a hybrid system.  We target one-shot learning or low-cost training from limited training material as well as real-time learning and analysis.

For this PhD, we expect the student to focus on the following aspects:

(i) The student understands and learns about various physical mechanisms - often undesired reliability issues - in existing or novel electronic device concepts and looks for opportunities to exploit these mechanisms for alternative computing purposes. The student will work in a group that has many decades of experience on detailed physical modeling of reliability phenomena.

(ii) The student has or is willing to acquire a good conceptual insight into the principles of several AI algorithms.

(iii) The student shows a creative approach towards optimizing different aspects of AI algorithms into a hybrid approach for processing data with the available physical processes. This PhD is a complete bottom-up approach where the student starts from the device properties and builds an AI implementation on top of this.  At this moment, imec has a concept that combines Phase Space Reconstruction with Reservoir Computing, but new ideas or extensions are encouraged.

(iv) The student helps in the hardware design of a small-scale system on chip. This will be done in collaboration with other students. The student will test and explore the working of this chip and identify and solve possible operational bottlenecks.

(v) The student applies the new approach to robotic test cases (in collaboration with robotics scientists) in order to demonstrate the feasibility, opportunities as well as limitations of a small-scale AI on chip.

In summary, we are looking for a creative student with an open mind that would like to work on small-scale hardware- and physics-anchored, alternative artificial intelligence solutions aiming at overcoming present-day fundamental issues of neural networks.