Machine Learning Stories

 This paper introduces a new way to sequential learning of tasks without forgetting previously learned tasks. In this regime,      D ata/Tasks are presented to the model sequentially.     The model needs to adapt to each task  while remembering the knowledge from the previous tasks.     There should not be a mechanism to  store and replay the data from the prior tasks as the overhead would be large when there are too many tasks. What is the crux of this paper? The authors try to solve catastrophic forgetting in machine learning models by analogizing it to how human brains avoid it. Knowledge is encoded in the brain by reducing the plasticity of synapses vital to previously learned tasks. From the neuro biological models of synaptic consolidation - the authors devise EWC or elastic weight consolidation. This algorithm slows down learning of certain weights based on how important they were to the previous tasks. How is it achieved? Supp...

Identifying the most influential parameters in a representation learned for one task plays an important role in continual learning. This paper addresses two learning directions:  1) completely eliminating catastrophic forgetting using weight pruning(dropping neurons/weights?), where a hard selection mask freezes the most certain parameters (BLLL-PRN)   2) reducing catastrophic forgetting by adaptively modifying the learning rates based on how uncertain the parameter values are (BLLL-REG)( This method has the added advantage that there's no need have the information of task boundaries ) Since we have considered a Bayesian approach, we have to look at Bayesian Neural networks. The posterior over the weights of the neural network is learned in a BNN. One of the popular ways to perform inference in a BNN is using Variational Inference. This paper exactly follows BayesByBackProp algorithm by [Blundell, 2015]. Each weight parameter of the network is represente...

This paper is about an approach that is used for solving task-based sequential learning problems in Neural Networks. In TBSL, datasets arrive as tasks sequentially, and the model trains them one by one as they arrive. The data once trained is lost and cannot be obtained to re-train the model again. If the new data fed to the model is very different from the one it was trained on before, it will completely adapt to serve the new data and forget what it learned before. This issue is termed as catastrophic forgetting. All the experiments that tested the method discussed in the paper involves classification tasks. What is the crux of the paper? This paper is about doing continual learning within a single neural network itself. The authors have identified that in order for the model to remember the older tasks, the important parameters for a particular task should not be overwritten during training of future tasks. The important parameters are generally referred to as the weights...