Memories can be elusive for both humans and machines, and understanding why artificial agents experience gaps in their cognitive processes is crucial for the advancement of artificial intelligence (AI). Electrical engineers at The Ohio State University have delved into the impact of “continual learning” on overall performance in AI systems. Continual learning involves training computers to continuously learn a sequence of tasks, utilizing past knowledge to improve learning new ones.

However, one major challenge scientists face is overcoming the machine learning equivalent of memory loss, known as “catastrophic forgetting.” As AI agents are trained on successive tasks, they tend to lose information from previous tasks, posing risks as AI becomes more integrated into society. Ness Shroff, an Ohio Eminent Scholar and professor of computer science and engineering at Ohio State, emphasized the importance of preventing these AI systems, such as automated driving applications or robotic systems, from forgetting crucial lessons.

The research team, including postdoctoral researchers Sen Lin and Peizhong Ju, as well as professors Yingbin Liang and Shroff, discovered intriguing parallels between human memory and AI memory. Just as humans may struggle to recall contrasting facts about similar situations but easily remember inherently different scenarios, artificial neural networks perform better when facing diverse tasks sequentially rather than tasks with similar features.

The team’s findings, to be presented at the International Conference on Machine Learning in Honolulu, Hawaii, reveal that dynamic, lifelong learning in autonomous systems can significantly enhance machine learning algorithms. This adaptability allows scientists to scale up AI algorithms more rapidly and equip them to handle evolving environments and unexpected situations. Ultimately, the goal is to have these AI systems mimic the learning capabilities of humans.

To optimize algorithm memory, Shroff suggests teaching dissimilar tasks early in the continual learning process. This approach expands the network’s capacity for new information, improving its ability to learn more similar tasks later on.

The research holds immense significance as it enhances our understanding of AI and its relationship to the human brain. By uncovering the similarities between machines and human learning processes, the team is paving the way for a new era of intelligent machines that learn and adapt like their human counterparts.

By Impact Lab