Catastrophic oblivion, Achilles heel of artificial intelligence

Artificial intelligences are often described as black boxes, because specialists don’t really know what’s going on there. Neural networks are trained on large amounts of data and have to perform certain tasks, but the resulting algorithms are too complex to analyze. However, all AIs face the same problem: catastrophic forgetting.

Researchers at Ohio State University have been studying this phenomenon in an attempt to find solutions. As the AI ​​learns new tasks, it forgets previously learned information. Catastrophic forgetting is especially problematic for continuous learning and can become dangerous, for example self-driving carself-driving car forget the basic safety rules.

Learn distinct tasks early

Researchers have found that AI memory bears similarities to human memory. Humans may have difficulty remembering the details of two similar scenarios and more easily remember two very separate events. Similarly, artificial neural networks retain information better when the subsequent tasks are very different.

Specifically, the researchers found that by learning a set of tasks early, the AI ​​can more easily learn new information and similar tasks later. The researchers hope that by analyzing the similarities between the machines and the brainbrain the human will will create AI that can learn and adapt like us. They will present their research findings at the International Conference on Machine Learning (ICML) next week.