Benjamin Lee

Aurko Roy, Rohan Anil, Guangda Lai et al. · deepmind

Transformer models have recently emerged as one of the foundational models in natural language processing, and as a byproduct, there is significant recent interest and investment in scaling these models. However, the training and inference costs of these large Transformer language models are prohibitive, thus necessitating more research in identifying more efficient variants. In this work, we propose a simple yet effective modification to the Transformer architecture inspired by the literature in statistical language modeling, by augmenting the model with n-grams that are constructed from a discrete latent representation of the text sequence. We evaluate our model, the N-Grammer on language modeling on the C4 data-set as well as text classification on the SuperGLUE data-set, and find that it outperforms several strong baselines such as the Transformer and the Primer. We open-source our model for reproducibility purposes in Jax.

Haiyang Huang, Newsha Ardalani, Anna Sun et al.

Mixture-of-Experts (MoE) models have gained popularity in achieving state-of-the-art performance in a wide range of tasks in computer vision and natural language processing. They effectively expand the model capacity while incurring a minimal increase in computation cost during training. However, deploying such models for inference is difficult due to their large size and complex communication pattern. In this work, we provide a characterization of two MoE workloads, namely Language Modeling (LM) and Machine Translation (MT) and identify their sources of inefficiencies at deployment. We propose three optimization techniques to mitigate sources of inefficiencies, namely (1) Dynamic gating, (2) Expert Buffering, and (3) Expert load balancing. We show that dynamic gating improves maximum throughput by 6.21-11.23$\times$ for LM, 5.75-10.98$\times$ for MT Encoder and 2.58-5.71$\times$ for MT Decoder. It also reduces memory usage by up to 1.36$\times$ for LM and up to 1.1$\times$ for MT. We further propose Expert Buffering, a new caching mechanism that only keeps hot, active experts in GPU memory while buffering the rest in CPU memory. This reduces static memory allocation by up to 1.47$\times$. We finally propose a load balancing methodology that provides additional scalability to the workload.

Yucheng Lu, Tobias Rau, Benjamin Lee et al.

We present a design space for animated transitions of the appearance of 3D spatial datasets in a hybrid Augmented Reality (AR)-desktop context. Such hybrid interfaces combine both traditional and immersive displays to facilitate the exploration of 2D and 3D data representations in the environment in which they are best displayed. One key aspect is to introduce transitional animations that change between the different dimensionalities to illustrate the connection between the different representations and to reduce the potential cognitive load on the user. The specific transitions to be used depend on the type of data, the needs of the application domain, and other factors. We summarize these as a transition design space to simplify the decision-making process and provide inspiration for future designs. First, we discuss 3D visualizations from a spatial perspective: a spatial encoding pipeline, where 3D data sampled from the physical world goes through various transformations, being mapped to visual representations, and then being integrated into a hybrid AR-desktop environment. The transition design then focuses on interpolating between two spatial encoding pipelines to provide a smooth experience. To illustrate the use of our design space, we apply it to three case studies that focus on applications in astronomy, radiology, and chemistry; we then discuss lessons learned from these applications.

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Benjamin Lee, Dave Brown, Bongshin Lee et al.

A fundamental part of data visualization is transforming data to map abstract information onto visual attributes. While this abstraction is a powerful basis for data visualization, the connection between the representation and the original underlying data (i.e., what the quantities and measurements actually correspond with in reality) can be lost. On the other hand, virtual reality (VR) is being increasingly used to represent real and abstract models as natural experiences to users. In this work, we explore the potential of using VR to help restore the basic understanding of units and measures that are often abstracted away in data visualization in an approach we call data visceralization. By building VR prototypes as design probes, we identify key themes and factors for data visceralization. We do this first through a critical reflection by the authors, then by involving external participants. We find that data visceralization is an engaging way of understanding the qualitative aspects of physical measures and their real-life form, which complements analytical and quantitative understanding commonly gained from data visualization. However, data visceralization is most effective when there is a one-to-one mapping between data and representation, with transformations such as scaling affecting this understanding. We conclude with a discussion of future directions for data visceralization.

Benjamin Lee, Xiaoyun Hu, Maxime Cordeil et al.

Immersive technologies offer new opportunities to support collaborative visual data analysis by providing each collaborator a personal, high-resolution view of a flexible shared visualisation space through a head mounted display. However, most prior studies of collaborative immersive analytics have focused on how groups interact with surface interfaces such as tabletops and wall displays. This paper reports on a study in which teams of three co-located participants are given flexible visualisation authoring tools to allow a great deal of control in how they structure their shared workspace. They do so using a prototype system we call FIESTA: the Free-roaming Immersive Environment to Support Team-based Analysis. Unlike traditional visualisation tools, FIESTA allows users to freely position authoring interfaces and visualisation artefacts anywhere in the virtual environment, either on virtual surfaces or suspended within the interaction space. Our participants solved visual analytics tasks on a multivariate data set, doing so individually and collaboratively by creating a large number of 2D and 3D visualisations. Their behaviours suggest that the usage of surfaces is coupled with the type of visualisation used, often using walls to organise 2D visualisations, but positioning 3D visualisations in the space around them. Outside of tightly-coupled collaboration, participants followed social protocols and did not interact with visualisations that did not belong to them even if outside of its owner's personal workspace.

Jonathan Shen, Patrick Nguyen, Yonghui Wu et al.

Lingvo is a Tensorflow framework offering a complete solution for collaborative deep learning research, with a particular focus towards sequence-to-sequence models. Lingvo models are composed of modular building blocks that are flexible and easily extensible, and experiment configurations are centralized and highly customizable. Distributed training and quantized inference are supported directly within the framework, and it contains existing implementations of a large number of utilities, helper functions, and the newest research ideas. Lingvo has been used in collaboration by dozens of researchers in more than 20 papers over the last two years. This document outlines the underlying design of Lingvo and serves as an introduction to the various pieces of the framework, while also offering examples of advanced features that showcase the capabilities of the framework.