There is a growing interest in the learning-to-learn paradigm, also known as\nmeta-learning, where models infer on new tasks using a few training examples.\nRecently, meta-learning based methods have been widely used in few-shot\nclassification, regression, reinforcement learning, and domain adaptation. The\nmodel-agnostic meta-learning (MAML) algorithm is a well-known algorithm that\nobtains model parameter initialization at meta-training phase. In the meta-test\nphase, this initialization is rapidly adapted to new tasks by using gradient\ndescent. However, meta-learning models are prone to overfitting since there are\ninsufficient training tasks resulting in over-parameterized models with poor\ngeneralization performance for unseen tasks. In this paper, we propose a\nBayesian neural network based MAML algorithm, which we refer to as the B-SMALL\nalgorithm. The proposed framework incorporates a sparse variational loss term\nalongside the loss function of MAML, which uses a sparsifying approximated KL\ndivergence as a regularizer. We demonstrate the performance of B-MAML using\nclassification and regression tasks, and highlight that training a sparsifying\nBNN using MAML indeed improves the parameter footprint of the model while\nperforming at par or even outperforming the MAML approach. We also illustrate\napplicability of our approach in distributed sensor networks, where sparsity\nand meta-learning can be beneficial.\n