Meta Learning Model: Model that learns about Models
- In this project we attempt to uncover patterns within the black box of neural networks by training a meta-classifier.
- A solid understanding of these inner workings could lead to vast improvements in performance, and increase our confidence that a given model is behaving as we expect.
- Our approach consists of first training many binary classifiers on the MNIST dataset, then using this collection of models as a dataset for a meta-model. This meta-model takes weights of an MNIST classifier as input, and classifies them according to which number they were trained to identify.
- After much experimentation, we were unable to achieve satisfactory results from the meta-model, but came away with many insights on how to approach the problem in the future.
