The project has identified three requirements for lifelong learning in low-power standalone devices: new algorithms that can be updated in the field when novelty is encountered, architectural reconfigurability when learning new examples on the fly in real time, and new memory technologies to build those new reconfigurable architectures and execute those new lifetime learning algorithms. The project proposes the co-designing of lifelong learning algorithms, architectures, and multi-level 3D memory technologies into tomorrow’s accelerators.
Key takeaways:
- The next generation of edge devices, such as AI and IoT devices, should have lifelong learning built-in, allowing them to acquire new knowledge autonomously in real time, according to a collaborative project by experts from various institutions.
- The project members have confirmed the feasibility of lifetime learning in network edge devices by reviewing 22 prototypes under development worldwide, although none of these prototypes include all the necessary features for lifelong learning.
- Three key requirements for lifelong learning in low-power standalone devices are new algorithms that can be updated in the field, architectural reconfigurability for learning new examples in real time, and new memory technologies for building reconfigurable architectures and executing lifetime learning algorithms.
- The project proposes co-designing of lifelong learning algorithms, architectures, and multi-level 3D memory technologies into tomorrow's accelerators, with the aim of creating devices that consume less than 1 milliwatt of power while generating a billion operations per second.