AI Goes Back to School

Continuous learning in machine learning refers to the capability of an artificial intelligence (AI) system to incessantly obtain fresh information and adjust its comprehension according to new data throughout its life in operation. This idea is derived from the human brain's aptitude to learn and enhance over time. Unlike conventional learning techniques where a model is trained on an unchanging dataset and remains unvarying once implemented, lifelong learning enables AI systems to progress and retain relevance in ever-changing surroundings.

Continuing to learn throughout one's life can also improve the effectiveness of data. Conventional approaches often necessitate extensive and varied datasets to train effectively. Conversely, lifelong learning capitalizes on previous knowledge and necessitates less data to adapt to novel circumstances, thus exhibiting greater efficiency in resource utilization.

Furthermore, continuous learning empowers AI systems to specialize in particular duties while gathering knowledge and expertise. This specialization not only enhances domain proficiency but also enables the model to excel in its designated fields, resulting in more precise and context-sensitive decision-making.

The ability facilitates the exchange of information between individuals who engage in continuous learning throughout their lives (Image source: Y. Ge et al.)

Nevertheless, if we are to extend the analogy regarding the human brain's capacity for acquiring knowledge, these individuals who prioritize lifelong learning resemble an individual who finds themselves marooned on a remote island, devoid of the privilege to extract wisdom from others. In actuality, humans do not solely rely on personal experiences to amass knowledge; instead, they acquire it from fellow beings. Substantial strides in expanding human understanding are only made possible by those who build upon the wisdom and expertise of those who came before them, metaphorically speaking.

Is it not also reasonable for artificially intelligent systems to adopt this approach? This query has been posed by a group of scientists headed by individuals from the University of Southern California. They have crafted a decentralized structure wherein independent learning entities can strive to amass their own expertise in a specific field, while also being able to exchange and transfer this knowledge to fellow entities. By doing so, machine learning systems can swiftly accumulate a wide-ranging foundation of knowledge.

The team has developed a method known as SKILL (Shared Knowledge Lifelong Learning), which involves multiple learning agents independently acquiring new skills simultaneously. Once an agent becomes skilled in a particular area, it shares its knowledge across a decentralized communication network. In addition to the specific skill-related information, it also provides a summary of its learned tasks in a universal format that other agents can comprehend, regardless of the task.

The other individuals then obtain this data through the network and merge it with their own understanding, thus empowering them to tackle fresh assignments without going through the challenging process of learning independently. As this progression persists, each individual eventually gains knowledge from each and every other person, and they all become skilled in effectively accomplishing every task.

Working of a self-sufficient agent (Image: Y. Ge et al.)

The scientists organized a trial to assess EXPERTISE by having 102 educational entities acquire distinct abilities related to recognizing images. Through collaboration, all of the entities managed to successfully grasp every proficiency. Furthermore, they achieved this in an impressively short span of time—thanks to sharing the responsibilities, the learning process for the tasks was expedited by a remarkable 101.5 times.

In the future, the team is concentrating on verifying that the system has the ability to expand and handle a large number of tasks, potentially even reaching millions. They are confident that this capability has the potential to create significant changes in various industries, such as medicine. They have a vision of agents acquiring knowledge about various illnesses, treatments, recent studies, and additional information to support the development of a universal medical assistant that can aid physicians.

However, prior to that occurrence, there are still remaining tasks to be accomplished. The scientists are seeking chances to enhance the efficiency of SKILL, and they also plan to facilitate further applications apart from just identifying images.

I excel in research and development, innovation, and crafting revolutionary ideas that you didn't even realize you desired.