Humans, and generally all intelligent beings, have the innate ability to learn and adapt continually to the changing environment where they are immersed. People continually learn new skills or behaviors, while not forgetting past knowledge or experience. This ability is essential for any intelligence (real or artificial) that aims to cope with the astonishing complexity of the real world.
Continual Learning (CL) aims to provide artificial intelligence (AI) models this ability, making possible the creation of a real lifelong artificial intelligence, that, as humans, can learn and accumulate knowledge for the entire duration of its “life”.
In recent years, Machine Learning (ML) has shown impressive results in many different fields, making possible the realization of previously unthinkable technologies now used in our everyday objects. Machine learning is based on the concept of “training” a model on a huge amount of data so that the model can learn directly from the data how to resolve the given task.
In a nutshell, ML workflow can be summarized as follows:
- Acquire a great amount of data from the real world.
- Split the acquired data into training and test data.
- Train the ML model on the training data.
- Test the ML model performance using the test data.
- If the performance is satisfactory, deploy the model to the real world.
But, what if the data changes over time? What if the real-world data is different from the data acquired in the first step? What if we want to adapt our ML to new data? What if we want to train continually our ML model?
The first thought may be to collect some other data from the real world and re-apply the workflow. Unfortunately, this leads to the catastrophic forgetting problem, where the model completely forgets how to deal with past data and forgets past knowledge while new knowledge is learned.
The main goal of continual learning is to address the catastrophic forgetting problem, making it possible to train ML models incrementally without forgetting old knowledge and behaviours.
Another problem with the classical ML workflow is the deployment of the model. In that workflow, there are two separate phases in the life of an ML model: training and deployment. In the training phase, the model is “offline”, so it cannot be used in the real world. On the contrary, in the deployment phase, the model is active but cannot learn anything new from the data it processes.
Continual learning also addresses this problem, since training a model continually means that the last version of the model (the one trained with the last seen data) is always deployed and usable, without the need for a time and resource-consuming re-training.
Continual learning can be the key to solving many problems of ML systems that have been puzzled scientists for years, and one of the keys to a real general artificial intelligence. In this sense, continual learning is heavily influenced, and can heavily influence, fields of research as neuroscience, brain studies, reinforcement learning, agent-based artificial intelligence, machine learning, statistical learning, and many others.
- Members of MI@Biolab are among the founders of Continual AI
- Strong contribution to the design and development of Avalanche (continual learning Framework)
- Introduced Core50 (a dataset and benchmark) for continual object recognition
- Designed AR1: a state of the art CL algorithm
- Proposed Latent Replay for real time continual learning
- Introduced Generative Negative Replay
- Continual Learning in Real-Life Applications (Graffieti G., Borghi G., Maltoni D.) IEEE Robotics and Automation Letters [PDF]
Join us in the quest for continual artificial intelligence!