We propose NovoGrad, a first-order stochastic gradient method with layer-wise gradient normalization via second moment estimators and with decoupled weight decay for a better regularization. The method requires half as much memory as Adam/AdamW. We evaluated NovoGrad on a diverse set of problems, including image classification, speech recognition, neural machine translation and language modeling. On these problems, NovoGrad performed equal to or better than SGD and Adam/AdamW. Empirically we show that NovoGrad (1) is very robust during the initial training phase and does not require learning rate warm-up, (2) works well with the same learning rate policy for different problems, and (3) generally performs better than other optimizers for very large batch sizes.

Neural Module Networks, originally proposed for the task of visual question answering, are a class of neural network architectures that involve human-specified neural modules, each designed for a specific form of reasoning. In current formulations of such networks only the parameters of the neural modules and/or the order of their execution is learned. In this work, we further expand this approach and also learn the underlying internal structure of modules in terms of the ordering and combination of simple and elementary arithmetic operators. Our results show that one is indeed able to simultaneously learn both internal module structure and module sequencing without extra supervisory signals for module execution sequencing. With this approach, we report performance comparable to models using hand-designed modules.

Graphs arise naturally in many real-world applications including social networks, recommender systems, ontologies, biology, and computational finance. Traditionally, machine learning models for graphs have been mostly designed for static graphs. However, many applications involve evolving graphs. This introduces important challenges for learning and inference since nodes, attributes, and edges change over time. In this survey, we review the recent advances in representation learning for dynamic graphs, including dynamic knowledge graphs. We describe existing models from an encoder-decoder perspective, categorize these encoders and decoders based on the techniques they employ, and analyze the approaches in each category. We also review several prominent applications and widely used datasets, and highlight directions for future research.

We present a recurrent neural network memory that uses sparse coding to create a combinatoric encoding of sequential inputs. Using several examples, we show that the network can associate distant causes and effects in a discrete stochastic process, predict partially-observable higher-order sequences, and enable a DQN agent to navigate a maze by giving it memory. The network uses only biologically-plausible, local and immediate credit assignment. Memory requirements are typically one order of magnitude less than existing LSTM, GRU and autoregressive feed-forward sequence learning models. The most significant limitation of the memory is generalization to unseen input sequences. We explore this limitation by measuring next-word prediction perplexity on the Penn Treebank dataset.

With the growing importance of personalized recommendation, numerous recommendation models have been proposed recently. Among them, Matrix Factorization (MF) based models are the most widely used in the recommendation field due to their high performance. However, MF based models suffer from cold start problems where user-item interactions are sparse. To deal with this problem, content based recommendation models which use the auxiliary attributes of users and items have been proposed. Since these models use auxiliary attributes, they are effective in cold start settings. However, most of the proposed models are either unable to capture complex feature interactions or not properly designed to combine user-item feedback information with content information. In this paper, we propose Self-Attentive Integration Network (SAIN) which is a model that effectively combines user-item feedback information and auxiliary information for recommendation task. In SAIN, a self-attention mechanism is used in the feature-level interaction layer to effectively consider interactions between multiple features, while the information integration layer adaptively combines content and feedback information. The experimental results on two public datasets show that our model outperforms the state-of-the-art models by 2.13%

Object segmentation is a crucial problem that is usually solved by using supervised learning approaches over very large datasets composed of both images and corresponding object masks. Since the masks have to be provided at pixel level, building such a dataset for any new domain can be very costly. We present ReDO, a new model able to extract objects from images without any annotation in an unsupervised way. It relies on the idea that it should be possible to change the textures or colors of the objects without changing the overall distribution of the dataset. Following this assumption, our approach is based on an adversarial architecture where the generator is guided by an input sample: given an image, it extracts the object mask, then redraws a new object at the same location. The generator is controlled by a discriminator that ensures that the distribution of generated images is aligned to the original one. We experiment with this method on different datasets and demonstrate the good quality of extracted masks.

A significant issue in training deep neural networks to solve supervised learning tasks is the need for large numbers of labelled datapoints. The goal of semi-supervised learning is to leverage ubiquitous unlabelled data, together with small quantities of labelled data, to achieve high task performance. Though substantial recent progress has been made in developing semi-supervised algorithms that are effective for comparatively small datasets, many of these techniques do not scale readily to the large (unlaballed) datasets characteristic of real-world applications. In this paper we introduce a novel approach to scalable semi-supervised learning, called Local Label Propagation (LLP). Extending ideas from recent work on unsupervised embedding learning, LLP first embeds datapoints, labelled and otherwise, in a common latent space using a deep neural network. It then propagates pseudolabels from known to unknown datapoints in a manner that depends on the local geometry of the embedding, taking into account both inter-point distance and local data density as a weighting on propagation likelihood. The parameters of the deep embedding are then trained to simultaneously maximize pseudolabel categorization performance as well as a metric of the clustering of datapoints within each psuedo-label group, iteratively alternating stages of network training and label propagation. We illustrate the utility of the LLP method on the ImageNet dataset, achieving results that outperform previous state-of-the-art scalable semi-supervised learning algorithms by large margins, consistently across a wide variety of training regimes. We also show that the feature representation learned with LLP transfers well to scene recognition in the Places 205 dataset.

In this paper we introduce a new neural architecture for sorting unordered sequences where the correct sequence order is not easily defined but must rather be inferred from training data. We refer to this architecture as OrderNet and describe how it was constructed to be naturally permutation equivariant while still allowing for rich interactions of elements of the input set. We evaluate the capabilities of our architecture by training it to approximate solutions for the Traveling Salesman Problem and find that it outperforms previously studied supervised techniques in its ability to generalize to longer sequences than it was trained with. We further demonstrate the capability by reconstructing the order of sentences with scrambled word order.

Due to their hunger for big data, modern deep learning models are trained in parallel, often in distributed environments, where communication of model updates is the bottleneck. Various update compression (e.g., quantization, sparsification, dithering) techniques have been proposed in recent years as a successful tool to alleviate this problem. In this work, we introduce a new, remarkably simple and theoretically and practically effective compression technique, which we call natural compression (NC). Our technique is applied individually to all entries of the to-be-compressed update vector and works by randomized rounding to the nearest (negative or positive) power of two. NC is "natural" since the nearest power of two of a real expressed as a float can be obtained without any computation, simply by ignoring the mantissa. We show that compared to no compression, NC increases the second moment of the compressed vector by the tiny factor 9/8 only, which means that the effect of NC on the convergence speed of popular training algorithms, such as distributed SGD, is negligible. However, the communications savings enabled by NC are substantial, leading to 3-4x improvement in overall theoretical running time. For applications requiring more aggressive compression, we generalize NC to natural dithering, which we prove is exponentially better than the immensely popular random dithering technique. Our compression operators can be used on their own or in combination with existing operators for a more aggressive combined effect. Finally, we show that N is particularly effective for the in-network aggregation (INA) framework for distributed training, where the update aggregation is done on a switch, which can only perform integer computations.

We propose a novel unsupervised domain adaptation framework based on domain-specific batch normalization in deep neural networks. We aim to adapt to both domains by specializing batch normalization layers in convolutional neural networks while allowing them to share all other model parameters, which is realized by a two-stage algorithm. In the first stage, we estimate pseudo-labels for the examples in the target domain using an external unsupervised domain adaptation algorithm---for example, MSTN or CPUA---integrating the proposed domain-specific batch normalization. The second stage learns the final models using a multi-task classification loss for the source and target domains. Note that the two domains have separate batch normalization layers in both stages. Our framework can be easily incorporated into the domain adaptation techniques based on deep neural networks with batch normalization layers. We also present that our approach can be extended to the problem with multiple source domains. The proposed algorithm is evaluated on multiple benchmark datasets and achieves the state-of-the-art accuracy in the standard setting and the multi-source domain adaption scenario.