Enter the next generation of machine learning…
In 1979, Donald B. Maudsley described meta learning as ‘the process by which learners become aware … and increasingly in control of habits of perception, inquiry, learning, and growth’. Maudsley was speaking in terms of social psychology, but the term has since been applied to computer science. In this field, meta learning refers to a subdivision of machine learning in which automatic learning algorithms are applied on meta data – data about data – rather than particular datasets.
Traditionally, software models are trained on specific data that helps them achieve a certain task. In contrast, the meta learning approach attempts to make artificially intelligent systems more flexible through learning to learn. In reinforcement learning, for instance, the model learns using an action policy instead of a specific dataset. ‘Few shots meta learning’ is another meta learning approach in which deep neural networks mimic the way babies learn to identify objects with minimal exposure. ‘Optimiser meta learning’ refers to the application of one neural network’s optimisations to the hyper parameters of another. In each case, the aim is to move away from specialised learning to a more general ability to complete tasks.
This is where Maudsley’s definition becomes relevant. If an automatic learning model can ‘meta learn’, then it becomes more humanlike in its acquisition of knowledge and therefore also the tasks it can complete. Ultimately this should deliver greater value, reach a wider range of applications and encourage the democratisation of machine learning. However, it also represents a tentative step towards the hotly debated artificial general intelligence (AGI) and the potential eclipse of human ability.
For more information about the development of machine learning, sign up for our free newsletter.