The shift from cloud-based processing to on-device intelligence is reshaping how artificial intelligence systems are designed and deployed. Moving computation to the edge enables lower latency and improved privacy, but also imposes strict energy and resource constraints. As a result, many edge AI systems rely on static, pre-trained models that prioritize efficiency over adaptability and quickly become outdated in dynamic environments. Bridging this gap requires learning systems that combine low energy consumption with the ability to adapt continuously.
The human brain remains one of the most powerful and efficient learning systems we know. It learns continually from daily experience, adapts to new environments without overwriting prior knowledge, and operates on roughly 20 watts of power. These capabilities make the brain not only a benchmark for energy efficiency, but also a key source of inspiration for building adaptive artificial systems capable of continual learning.
Neuromorphic computing seeks to translate these biological principles into artificial systems. By emulating the brain’s event-driven, parallel, and low-power processing through spiking neural networks (SNNs), neuromorphic approaches offer a framework for both energy-efficient computation and bio-inspired continual learning algorithms that mitigate catastrophic forgetting.
In this talk, we will examine the computational foundations and distinctive features of spiking neural networks, highlighting how and why they can achieve greater energy efficiency than conventional neural networks. We will then introduce the concept of continual learning and discuss how bio-inspired mechanisms can be integrated into SNNs to enable adaptive learning without catastrophic forgetting. Finally, we will outline the key requirements and challenges for implementing such systems into neuromorphic hardware.
27/2/2026 11:00 - 12:00
ESAT Aula L
