Sample efficiency has been one of the major challenges for deep reinforcement learning.

Deep learning achieves state-of-the-art results in many tasks in computer vision and natural language processing.

By transferring both features and gradients between different layers, shortcut connections explored by ResNets allow us to effectively train very deep neural networks up to hundreds of layers. However, the additional computation costs induced by those shortcuts are often overlooked.

Communication overhead is a major bottleneck hampering the scalability of distributed machine learning systems.

In this work, we provide a theoretical foundation for crafting transferable adversarial examples to entire hypothesis classes.

Moreover, we identify the drawbacks of these greedy algorithms and propose new algorithms that fix these issues.

A prominent technique for self-supervised representation learning has been to contrast semantically similar and dissimilar pairs of samples.

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. Summary: The paper presents an extension of gated auto encoders to time-series data. The main idea is to use a gated auto encoder to model the time series in an autoregressive manner; predicting x_{t+1} from x_t using a gated autoencoder whose mapping unit values are initialised using a pair of contiguous datapoints. The paper introduces two interesting refinements: predictive training, and higher order relational features. Predictive training is a training criterion suitable for time series data that is different from the criterion normally used for gated auto encoders. Predictive training tries to minimise the square error in predicting x_{t+1} given x_{t} and the value of the mapping units that optimally predict x_{t} given x_{t-1}.