DeepMind Researchers Develop an AI that Does not Overwrite Past Knowledge

Researchers at DeepMind announced that they have developed an AI called “ .

In a blog post entitled “Enabling Continual Learning in Neural Networks,” DeepMind researchers wrote that one of the ways humans learns new things is through synaptic consolidation – a process in which “connections between neurons are less likely to be overwritten if they have been important in previously learnt tasks.”

DeepMind researchers said they took inspiration in the synaptic consolidation of the human brain in developing EWC to address the problem of “catastrophic forgetting” – a fundamental limitation of neural networks whereby new adaptations overwrite the previous knowledge that the neural network had previously learned.

“A neural network consists of several connections in much the same way as a brain. After learning a task, we compute how important each connection is to that task. When we learn a new task, each connection is protected from modification by an amount proportional to its importance to the old tasks,” DeepMind researchers said.

To test if EWC can learn new things by remembering past knowledge, the researchers exposed EWC to Atari games sequentially. The researchers said that EWC did not forget as easily and was able to learn to play several games, one after the other. DeepMind researchers added that EWC’s development “represents a step towards programs that can learn in a more flexible and efficient way.”