CARLS: Cross-platform Asynchronous Representation Learning System

In this work, we propose CARLS, a novel framework for augmenting the capacity of existing deep learning frameworks by enabling multiple components -- model trainers, knowledge makers and knowledge banks -- to concertedly work together in an asynchronous fashion across hardware platforms. The proposed CARLS is particularly suitable for learning paradigms where model training benefits from additional knowledge inferred or discovered during training, such as node embeddings for graph neural networks or reliable pseudo labels from model predictions. We also describe three learning paradigms -- semi-supervised learning, curriculum learning and multimodal learning -- as examples that can be scaled up efficiently by CARLS. One version of CARLS has been open-sourced and available for download at: https://github.com/tensorflow/neural-structured-learning/tree/master/research/carls

DynamicEmbedding: Extending TensorFlow for Colossal-Scale Applications

One of the limitations of deep learning models with sparse features today stems from the predefined nature of their input, which requires a dictionary be defined prior to the training. With this paper we propose both a theory and a working system design which remove this limitation, and show that the resulting models are able to perform better and efficiently run at a much larger scale. Specifically, we achieve this by decoupling a model's content from its form to tackle architecture evolution and memory growth separately. To efficiently handle model growth, we propose a new neuron model, called DynamicCell, drawing inspiration from from the free energy principle [15] to introduce the concept of reaction to discharge non-digestive energy, which also subsumes gradient descent based approaches as its special cases. We implement DynamicCell by introducing a new server into TensorFlow to take over most of the work involving model growth. Consequently, it enables any existing deep learning models to efficiently handle arbitrary number of distinct sparse features (e.g., search queries), and grow incessantly without redefining the model. Most notably, one of our models, which has been reliably running in production for over a year, is capable of suggesting high quality keywords for advertisers of Google Smart Campaigns and achieved significant accuracy gains based on a challenging metric -- evidence that data-driven, self-evolving systems can potentially exceed the performance of traditional rule-based approaches.

Context Aware Machine Learning

We propose a principle for exploring context in machine learning models. Starting with a simple assumption that each observation may or may not depend on its context, a conditional probability distribution is decomposed into two parts: context-free and context-sensitive. Then by employing the log-linear word production model for relating random variables to their embedding space representation and making use of the convexity of natural exponential function, we show that the embedding of an observation can also be decomposed into a weighted sum of two vectors, representing its context-free and context-sensitive parts, respectively. This simple treatment of context provides a unified view of many existing deep learning models, leading to revisions of these models able to achieve significant performance boost. Specifically, our upgraded version of a recent sentence embedding model not only outperforms the original one by a large margin, but also leads to a new, principled approach for compositing the embeddings of bag-of-words features, as well as a new architecture for modeling attention in deep neural networks. More surprisingly, our new principle provides a novel understanding of the gates and equations defined by the long short term memory model, which also leads to a new model that is able to converge significantly faster and achieve much lower prediction errors. Furthermore, our principle also inspires a new type of generic neural network layer that better resembles real biological neurons than the traditional linear mapping plus nonlinear activation based architecture. Its multi-layer extension provides a new principle for deep neural networks which subsumes residual network (ResNet) as its special case, and its extension to convolutional neutral network model accounts for irrelevant input (e.g., background in an image) in addition to filtering.