The new approach for artificial intelligence is with enabling AI to learn new tasks more successfully while forgetting fewer of the areas that have previously been learned when completing previous tasks. In other words – the ideal pupil.
To achieve this, researchers at North Carolina State University researchers, with support from the U.S. Army Research Laboratory, have come up with an alternate framework designed for deep neural networks, which enables artificial intelligence systems recall more of what has taken place before.
Like humans, AI systems tend to replace older learnt things with more recently learnt things, a process called ‘catastrophic forgetting‘. This is a factor of the design of most deep neural networks. Other limitations are with AI systems forgetting some of the things they knew about old tasks, while not learning to do new ones as well.
The new approach decouples network structure learning and model parameter learning, which enables improved recall while enabling an AI platform to continue to learn new things. This is called architecture optimization and it moves away from the standard way by which information is captured and recorded. Instead, a network is presented with more choices within its system when it receives something new. The AI can decide to: skip a layer; use a layer in the same way that previous tasks used it; attach a lightweight adapter to a layer, which modifies it slightly; or create an entirely new layer to store the new activity.
According to one of the researchers, Dr. Mary Anne Fields: “We expect the Army’s intelligent systems to continually acquire new skills as they conduct missions on battlefields around the world without forgetting skills that have already been trained.”
Drawing on an example, she says: “For instance, while conducting an urban operation, a wheeled robot may learn new navigation parameters for dense urban cities, but it still needs to operate efficiently in a previously encountered environment like a forest.”
The research was presented to the 36th International Conference on Machine Learning, held during June 9-15 in Long Beach, California. There has also been a white paper published, titled “Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting.”
