What happens when an AI model needs to keep learning instead of being trained once and deployed forever?
In this session, I explore the problem of continual learning through my journey of building and experimenting with NeuroAdapt, an adaptive AI research prototype. A model may perform well when learning one task, but when it is trained on new tasks, its performance on previously learned knowledge can deteriorate — a phenomenon known as catastrophic forgetting.
I’ll break down why this happens, introduce the stability-plasticity dilemma, and explore approaches that can help AI systems learn new information while preserving what they already know. Rather than presenting continual learning as just another ML concept, I’ll share the experiments, challenges, failures, and lessons from actually trying to build and evaluate an adaptive system.
The session will also explore how ideas from neuroscience can inspire the design of AI systems that are more capable of adapting over time, and what developers should consider when building models for environments where the data, tasks, and requirements continuously change.