We propose a new algorithm for temporal difference learning that learns faster, and is more stable.

We propose an algorithm for scalable recurrent learning, and evaluate it on prediction problems on the Arcade Learning Environment (ALE).

We propose an algorithm to approximates the gradient in for recurrent state learning, and meta-learning in O(n) operations and memory.

We propose a method for learning models that do not rely on spurious correlations. Our work builds on IRM (M Arjovsky, 2019) except unlike IRM, it can be implemented online to (1) detect spurious features for a set of given features and (2) learn non-spurious features from sensory data.

We propose OML, an objective for learning representations by using catastrophic interference as a training signal. Resultant representations are naturally sparse, accelerate future learning and are robust to forgetting under online updates in continual learning.

We propose a simple drop-in procedure for approximating the Bayesian credible regions of patient-specific survival functions that can be applied to many ISD models.

We isolate the truly effective existing ideas for incremental classifier learning from those that only work under certain conditions. Moreover, we propose a dynamic threshold moving algorithm that can successfully remove bias from an incrementally learned classifier when learning by knowledge distillation.

We propose a computationally efficient document segmentation algorithm that recursively uses convolutional neural networks to precisely localize a document in a natural image in real-time.

Recent Talks