Scientists Making Progress Towards AI That Learns Like Humans
Scientists are making strides in developing artificial intelligence (AI) that can learn continuously like humans, gradually building upon its previous knowledge.
According to researchers from the University of Ohio, continuous learning involves training a computer to consistently learn a series of tasks, using its past knowledge to improve its grasp of new tasks. This was presented at the International Conference on Machine Learning (ICML).
For years, scientists have grappled with the problem of 'catastrophic forgetting', where artificial neural networks lose information gained while acquiring new knowledge. This poses risks as AI becomes more essential in complex tasks.
The University of Ohio claims to have tackled this challenge by exposing AI to diverse tasks in succession, rather than tasks with similar features. This mirrors human learning, where we might struggle to recall contrasting facts about similar situations but easily remember inherently different scenarios.
While teaching dynamic, lifelong learning to autonomous systems remains a significant challenge, possessing such capabilities could speed up the advancement of machine learning algorithms and help them adapt to changing environments and unexpected situations.
The ultimate goal is to have these systems mimic human learning capabilities.
ICML is a major event in the machine learning industry. This year's event featured keynote presentations from John Schulman, co-founder of OpenAI, and Shakir Mohammed, head of research at Google DeepMind, the AI subsidiary in London.
Despite advocating for machine learning and AI adoption, ICML banned the use of AI language tools like ChatGPT to write academic papers at the event. Organizers argued that AI-generated papers lack novelty and are derivatives of existing work.
Watch: Does AI know what it’s doing
