Early in life, humans first learn to walk, and then a few years later, they learn to ride bicycles, and finally, as young adults, they learn to drive cars. In learning how to do each new task, humans don’t forget previous ones. Artificial neural networks, on the other hand, struggle to learn continually and consequently suffer from catastrophic forgetting: the tendency to lose almost all information about a previously learned task when attempting to learn a new one.
In this post, I will describe the technicalities of why neural networks do not learn continually, briefly discuss how the brain is thought to succeed at learning task after task, and finally highlight some exciting work in the machine learning community that builds on fundamental principles of neural computation to alleviate catastrophic forgetting.
Catastrophic forgetting in neural networks
Why do neural networks forget previously learned tasks so easily after learning a new one? Each task that a neural network learns to perform has a single error function that the network is trying to minimize, irrespective of the error of any other task. To be more specific, a typical neural network has a set of weights and biases, which are its parameters. For the network to change its behavior or learn how to perform a task, it must change its parameters, and this is exactly what happens during training. The figure below shows the weight space of a neural network: each axis corresponds to the value of a single parameter (imagine there are only two parameters for illustration purposes, but in reality there could be millions of dimensions). At any point in time, a typical neural network is at a single place in weight space based on its parameters’ values. Then, learning is equivalent to moving throughout weight space to a place where the error is small on the task being learned.
Say we have two tasks to be learned in sequence: distinguishing between different species of birds, and identifying various handwritten Greek letters. As a randomly initialized neural network learns to classify different species of birds, it will move through weight space to a place where the error on this task is low. Then, it will attempt to learn Greek letters. Just as before, it will move to a place where the error will be low on this new task. The places in weight space where the network can reliably identify different handwritten Greek letters are likely to be separate from those where it previously achieved success classifying bird types, especially since weight space is typically millions of dimensions.
In the diagram below, the orange circle represents the neural network at a place in weight space, and the areas shaded in green and purple correspond to where the error on learning a wide array of birds and greek letters is sufficiently small, respectively….[ ]