Gavin William Taylor
Publications
Divide and Contrast: Learning Robust Temporal Features without Augmentation
We present Di-COT, a framework for unsupervised time-series representation learning. By contrasting sub-blocks within instances, it avoids data augmentation, achieving state-of-the-art accuracy on downstream tasks with much lower training time.
Visualizing the loss landscape of neural nets
We introduce a novel approach for visualizing the loss landscape of large neural networks. We change network architectures and training parameters and then visualize the loss landscape, developing intuition into what makes one network better than another.
MetaPoison: Practical General-purpose Clean-label Data Poisoning
We introduce MetaPoison, a clean-label data poisoning framework. By approximating bilevel optimization via meta-learning, we craft imperceptible perturbations that force targeted misclassifications on deep models trained from scratch.
I have published in a number of different areas, including neural network optimization, data poisoning/adversarial examples, reinforcement learning, and representation learning for time series. My full set of publications can be found on my Google Scholar page, and selected publications spanning these topics are linked above.