- Abstract: Representation learning is one of the fundamental problems of machine learning. On its own, this problem can be cast as an unsupervised dimensionality reduction problem. However, representation learning is often also used as an implicit step in supervised learning (SL) or reinforcement learning (RL) problems. In this paper, we study the possible "interference" supervision, commonly provided through a loss function in SL or a reward function in RL, might have on learning representations, through the lens of learning from limited data and continual learning. Particularly, in connectionist networks, we often face the problem of catastrophic interference whereby changes in the data distribution cause networks to fail to remember previously learned information and learning representations can be done without labeled data. A primary running hypothesis is that representations learned using unsupervised learning are more robust to changes in the data distribution as compared to the intermediate representations learned when using supervision because supervision interferes with otherwise "unconstrained" representation learning objectives. To empirically test hypotheses, we perform experiments using a standard dataset for continual learning, permuted MNIST. Additionally, through a heuristic quantifying the amount of change in the data distribution, we verify that the results are statistically significant.