Artificial intelligence (AI) researchers at Ohio State University are making strides in creating AI systems that closely mimic human learning. At a machine learning conference in Honolulu, the team discussed “continual learning”, a process that enables computers to continuously acquire new skills while retaining previous knowledge – similar to how humans build upon past experiences to learn new things.
One of the challenges with artificial neural networks is “catastrophic forgetting”, where as they learn new tasks, they tend to lose information from their earlier training. This can be problematic as society increasingly relies on AI systems, such as in self-driving cars, where forgetting critical lessons could compromise safety.
Ness Shroff, a professor of computer science and engineering at Ohio State University, who led the study, emphasized the importance of AI systems retaining past knowledge as they learn new things. To overcome this issue, the researchers found that AI networks better retain information when trained on diverse and dissimilar tasks, rather than tasks with shared features. By exposing algorithms to varied tasks early on, their capacity to absorb new information is expanded.
Shroff expressed optimism that their work represents a significant step towards developing AI capable of lifelong, human-like learning. This advancement could accelerate the scaling up of algorithms and their adaptability to changing environments.
The International Conference on Machine Learning, which featured the Ohio State University research, also showcased other groundbreaking work. Researchers from MIT presented a technique to disrupt the creation of deepfake images by injecting small disruptive bits of code into source images.
Tech giant Google also played a prominent role at the conference, with over 80 scientific papers, including advancements in the AlphaFold 3-D protein modeler, fusion science, and new models like PaLM-E for robotics and Phenaki for generating video from text.
Shakir Mohamed, director for science, technology, and society at Google DeepMind, delivered a keynote speech highlighting the potential of machine learning to address major societal challenges, from healthcare to climate change. He emphasized the importance of diverse perspectives in developing AI that benefits everyone.
Researchers are making significant strides in creating AI systems that learn more like humans, with continual learning and retention of past knowledge. This could have profound implications for AI’s scalability and adaptability, potentially revolutionizing various industries and benefiting society as a whole.
