Neural Information Processing Systems http://nips.cc/

A.1 Training dataset pre-processing We used 40000publicly available videos from YouTube which were available in a spatial resolution of at least 1920 1080 pixels. In an attempt not to skew the distribution of content too far from what may inform biological representation learning, we excluded most artificial content such as screenshots and videos of computer games. To reduce video compression artifacts and prevent systematic downsampling artifacts, each segment was then spatially downsampled to a randomized height between 128 and 160. Each segment was then separated into 15 pairs of neighboring frames, and a randomly placed, but spatially colocated patch of 64 64 pixels was cropped out of each frame pair. The order of the frame pairs was then randomized in a running buffer, and all RGB pixel values were normalized to the range between 0 and 1 before being fed into the model.

Tractable models of human perception have proved to be challenging to build. Hand-designed models such as MS-SSIM remain popular predictors of human image quality judgements due to their simplicity and speed. Recent modern deep learning approaches can perform better, but they rely on supervised data which can be costly to gather: large sets of class labels such as ImageNet, image quality ratings, or both. We combine recent advances in information-theoretic objective functions with a computational architecture informed by the physiology of the human visual system and unsupervised training on pairs of video frames, yielding our Perceptual Information Metric (PIM)1. We show that PIM is competitive with supervised metrics on the recent and challenging BAPPS image quality assessment dataset and outperforms them in predicting the ranking of image compression methods in CLIC 2020. We also perform qualitative experiments using the ImageNet-C dataset, and establish that PIM is robust with respect to architectural details.

Graph neural networks (GNNs) have been successful in learning representations from graphs. Many popular GNNs follow the pattern of aggregate-transform: they aggregate the neighbors' attributes and then transform the results of aggregation with a learnable function. Analyses of these GNNs explain which pairs of non-identical graphs have different representations. However, we still lack an understanding of how similar these representations will be. We adopt kernel distance and propose transform-sum-cat as an alternative to aggregate-transform to reflect the continuous similarity between the node neighborhoods in the neighborhood aggregation. The idea leads to a simple and efficient graph similarity, which we name Weisfeiler-Leman similarity (WLS). In contrast to existing graph kernels, WLS is easy to implement with common deep learning frameworks. In graph classification experiments, transform-sum-cat significantly outperforms other neighborhood aggregation methods from popular GNN models. We also develop a simple and fast GNN model based on transform-sum-cat, which obtains, in comparison with widely used GNN models, (1) a higher accuracy in node classification, (2) a lower absolute error in graph regression, and (3) greater stability in adversarial training of graph generation.

We sincerely appreciate the reviewers for their careful reading, constructive questions and suggestions. We would very1 much like further exchanges to improve our work, but the following is our best effort within the current limits.2 First, we address questions appeared at least twice. We write P1, P2 for paragraph reference, and Rx for reviewers.3 We discuss two main motivations here: lack of graph loss, and empirical failure4 of distinguishing power.

Autonomous agents situated in real-world environments must be able to master large repertoires of skills. While a single short skill can be learned quickly, it would be impractical to learn every task independently. Instead, the agent should share knowledge across behaviors such that each task can be learned efficiently, and such that the resulting model can generalize to new tasks, especially ones that are compositions or subsets of tasks seen previously. A policy conditioned on a goal or demonstration has the potential to share knowledge between tasks if it sees enough diversity of inputs. However, these methods may not generalize to a more complex task at test time. We introduce compositional plan vectors (CPVs) to enable a policy to perform compositions of tasks without additional supervision. CPVs represent trajectories as the sum of the subtasks within them. We show that CPVs can be learned within a one-shot imitation learning framework without any additional supervision or information about task hierarchy, and enable a demonstration-conditioned policy to generalize to tasks that sequence twice as many skills as the tasks seen during training. Analogously to embeddings such as word2vec in NLP, CPVs can also support simple arithmetic operations - for example, we can add the CPVs for two different tasks to command an agent to compose both tasks, without any additional training.