Thesis Overview Invariance and Same-Equivariance Measures for Convolutional Neural Networks

Abstract

The main contributions of this thesis include:

• A comparative analysis of Neural Network based models for sign language handshape classification.
• An analysis of strategies to achieve equivariance to rotations in neural networks for:
  • Comparing the performance of strategies based on data augmentation and specially designed networks and layers.
  • Determining strategies to retrain networks so that they acquire equivariance to rotations.
• A set of measures to empirically analyze the equivariance of Neural Networks, as well as any other model based on latent representations, and the corresponding:
  • Validation of the measures to establish if they are indeed measuring the purported quantity. Analysis of the different variants of the proposed measures. Analysis of the properties of the measures, in terms of their variability to transformations, models and weight initialization.
  • Analysis of the impact of several hyperparameters of the models on the structure of their equivariance, including Max Pooling layers, Batch Normalization, and kernel size. Analysis of the structure of the equivariance in several well known CNN models such as ResNet, All Convolutional and VGG. Analysis of the impact on the equivariance of using specialized models to obtain equivariance such as Transformational Invariance Pooling.
  • Analysis of the class dependency of equivariance. Analysis of the effect of varying the complexity and diversity of the transformations on the measures.