Can Continual Learning be the key to Machine Intelligence?

Learning has been essential to our existence. And, continuous learning helps an individual avoid stagnation in any profession and ensures that one keeps on moving forward towards reaching his goal and potential. The same also goes for machine models that are backed by AI's machine learning algorithm. Enter continual learning. Continual learning, also called lifelong learning or online machine learning, is a fundamental idea in machine learning in which models continuously learn and evolve based on the input of increasing amounts of data while retaining previously learned knowledge. In practice, this refers to supporting a model's ability to autonomously learn and adapt in production as new data comes in. Just like us, this concept too, is based on mimicking humans' ability to learn incrementally by acquiring, fine-tuning, and transferring knowledge and skills throughout their lifespan.

What is Machine Learning?

Machine Learning came into existence in 1946 when Polish scientist Stanislaw Ulam was looking for solutions in an attempt to figure out the probability of winning a game of solitaire. Today it is defined as an application of artificial intelligence where a computer/machine learns from past experiences (input data) and makes future predictions. This allows the machine learning models to make assumptions, test them, and learn autonomously, without being explicitly programmed. If implemented correctly, ML shall unlock the ability to empower organizations to modernize the way they function.

Need for Continual Learning

In today's context, most of the machine learning models of AI are trained offline with input feeds. However, with the changing setting of industrial demands, it is high time we work on experimenting with an algorithm that relies on continuous learning. This is also crucial since data is no longer static and structured; instead, it is changing and scattered. For instance, ten years ago, the most searched keyword on Google was Swine Flu, but this year it is COVID-19. Next, every four years Olympics, Presidential Debate feature in trends too, thus implying data can have patterns too. Therefore, Google will need to retrain the model and display updates about these topics under 'trending.' This applies to every search engine, trending hashtags of social media, and recommender systems used by e-commerce sites and online stream platforms.  This is the simplest application of continual learning scenarios where the data distributions stay the same, but the data keeps coming.

Challenges

While this technology looks promising, there are several longstanding challenges exist in applying them. One of the many hurdles is catastrophic forgetting (or catastrophic interference phenomenon). This happens during the continual acquisition of incrementally available information from non-stationary data distributions, leading to the interference of the new infor­mation with what the model has already learned. This occurrence can lead to an abrupt decrease in performance while the new data is being integrated or, even worse, an overwrite of the model's previous knowledge with the new data. Although some researchers have proposed retraining the model every time new data are available, this process can be computationally expensive and inhibit real-time inferences.

Another challenge in continual learning is deploying new models to the same environment without negatively affecting users' experience and maintaining high accuracy. This is because deploying models for continual learning is a bit different from classic model deployment.