AI Goes Back to School

Lifelong learning in machine learning refers to the ability of an AI system to continuously acquire new knowledge and adapt its understanding based on incoming data throughout its operational lifespan. This concept is inspired by the human brain's capacity to learn and improve over time. Unlike traditional learning methods in which a model is trained on a fixed dataset and remains static once deployed, lifelong learning allows AI systems to evolve and stay relevant in dynamic environments.

Lifelong learning can also enhance data efficiency. Traditional methods often require large and diverse datasets for effective training. In contrast, lifelong learning leverages past experiences, requiring less data for adapting to new situations, making it more resource-efficient.

Additionally, lifelong learning enables AI systems to specialize in specific tasks as they accumulate knowledge and experience. This specialization fosters domain expertise and allows the model to become more proficient in its target areas, leading to more accurate and context-aware decision-making.

However, to continue the comparison with the human brain's ability to learn over time, these lifelong learners are like a person stranded on a desert island, unable to benefit from the knowledge of others. In reality, people do not acquire all of their knowledge from their own experience, rather they learn from others. Leaps forward in human knowledge are achieved by those that are standing on the shoulders of giants, so to speak.

Why should that not also be the case for artificially intelligent systems? That is the question recently asked by a team led by researchers at the University of Southern California. They have developed a distributed framework in which individual learning agents can work at building their own knowledge in a particular domain, but they can also share that knowledge with other agents. In this way, machine learning systems can rapidly build up a diverse base of knowledge.

The team’s technique, called SKILL (Shared Knowledge Lifelong Learning), consists of a number of independent learning agents sequentially learning new skills in parallel. When an agent becomes proficient at its skill, it shares it over a decentralized communication network. Along with the skill-specific data, it also sends summary information that represents its learned tasks in a common, task-agnostic manner that other agents can understand.

The other agents then receive this information over the network and incorporate it with their own knowledge, thereby enabling them to solve new tasks without having done the hard work of learning on their own. As this process continues, each agent eventually learns from every other agent, and they all become proficient in performing every task.

The researchers set up an experiment to test SKILL in which 102 learning agents each learned to perform a different image classification task. By sharing their knowledge, all of the agents ultimately mastered every skill. Moreover, they did it very quickly — by sharing the burden, the time needed to learn the tasks was reduced by a factor of 101.5.

Looking ahead, the team is focused on validating that the system can scale up to thousands, or even millions of tasks. They believe that such a capability could transform many fields, like medicine. They envision agents learning about different illnesses, treatments, recent research, and more to contribute to a universal medical assistant that could assist physicians.

But before that happens, there is still some work yet to be done. The researchers are looking for opportunities to speed up SKILL, and they also intend to enable additional use cases, beyond image recognition.