Using Word Embeddings to Explore the Learned Representations of Convolutional Neural Networks

Dhanush Dharmaretnam, Chris Foster, Alona Fyshe

Abstract: As deep neural net architectures minimize loss, they build up information in a hierarchy of learned representations that ultimately serve their final goal. Different architectures tackle this problem in slightly different ways, but all models aim to create representational spaces that accumulate information through the depth of the network. Here we build on previous work that indicated that two very different model classes trained on two very different tasks actually build knowledge representations that have similar underlying representations. Namely, we compare word embeddings from SkipGram (trained to predict co-occurring words) to several CNN architectures (trained for image classification) in order to understand how this accumulation of knowledge behaves in CNNs. We improve upon previous work by including 5 times more ImageNet classes in our experiments, and further expand the scope of the analyses to include a network trained on CIFAR-100. We characterize network behavior in pretrained models, and also during training, misclassification, and adversarial attack. Our work illustrates the power of using one model to explore another, gives new insights for CNN models, and provides a framework for others to perform similar analyses when developing new architectures.

Keywords: Distributional Semantics, word embeddings, cnns, interpretability

TL;DR: A simple technique using word embeddings provides multiple insights into the function and performance of CNNs, both during and after training, and for misclassified and adversarial examples.