Segmenting images using classification models. — Unpacking the features learned by a deep convolutional neural network (CNN) is a dauting task. Going through each layer to either visualize filters or features scales poorly with network depth, and although some cool figures can be created with this process, the result can be quite abstract, even psychedelic. Techniques…

Last summer I had the pleasure of working with a talented undergraduate researcher named Romain Speciel on a project that looked at how to regularize model ensembles in a way that improves robustness to adversarial examples. Our main concern was reducing the transferability of adversarial examples between models as that…

A recent attempt to improve the robustness of convolutional neural networks (CNNs) on image classification tasks has revealed an interesting link between robustness and interpretability. Models trained using adversarial training, a training procedure that augments training data with adversarial examples, have input gradients that qualitatively appear to be using more…

For the past couple of years, researchers and companies have been trying to make deep learning more accessible to non-experts by providing access to pre-trained computer vision or machine translation models. Using a pre-trained model for another task is known as transfer learning, but it still requires sufficient expertise to…

Adversarial examples are a large obstacle for a variety of machine learning systems to overcome. Their existence shows the tendency of models to rely on unreliable features to maximize performance, which if perturbed, can cause misclassifications with potentially catastrophic consequences. The informal definition of an adversarial example is an input…