Synthetic Benchmark Creation for Few shot Image Classification and Its performance on meta-learning Algorithms
Sambhavi Tiwari, P. Aditya Kumar, Manas Gogoi, Shekhar Verma · 2022 IEEE 6th Conference on Information and Communication Technology (CICT) · 2022
Meta-learning is the paradigm of learning how to learn, which eventually enables adapting to new tasks quickly with few data samples, owing to the training over a large number of tasks. Optimization-based approaches are of critical interest in all the different meta-learning algorithms as it generalizes to almost all kinds of tasks, from regression and classification to reinforcement learning. The idea of synthetic data sets used for performance evaluation is to create benchmark data sets that are diverse enough to encourage meta-learning. However, such data sets have been previously designed for analyzing and evaluating popular models like ANIL and MAML but only for regression tasks. This work extends the same to create synthetic data sets for the N-way classification problem. Analysis of our synthetic data set shows that using data that require a higher degree of representation change in the hidden layer of models is detrimental to the performance of ANIL. At the same time, MAML performs comparatively well under similar circumstances.