Category: Offsites

Posted by Edith Cohen and Uri Stemmer, Research Scientists, Google Research

Differential privacy (DP) is a rigorous mathematical definition of privacy. DP algorithms are randomized to protect user data by ensuring that the probability of any particular output is nearly unchanged when a data point is added or removed. Therefore, the output of a DP algorithm does not disclose the presence of any one data point. There has been significant progress in both foundational research and adoption of differential privacy with contributions such as the Privacy Sandbox and Google Open Source Library.

ML and data analytics algorithms can often be described as performing multiple basic computation steps on the same dataset. When each such step is differentially private, so is the output, but with multiple steps the overall privacy guarantee deteriorates, a phenomenon known as the cost of composition. Composition theorems bound the increase in privacy loss with the number k of computations: In the general case, the privacy loss increases with the square root of k. This means that we need much stricter privacy guarantees for each step in order to meet our overall privacy guarantee goal. But in that case, we lose utility. One way to improve the privacy vs. utility trade-off is to identify when the use cases admit a tighter privacy analysis than what follows from composition theorems.

Good candidates for such improvement are when each step is applied to a disjoint part (slice) of the dataset. When the slices are selected in a data-independent way, each point affects only one of the k outputs and the privacy guarantees do not deteriorate with k. However, there are applications in which we need to select the slices adaptively (that is, in a way that depends on the output of prior steps). In these cases, a change of a single data point may cascade — changing multiple slices and thus increasing composition cost.

In “Õptimal Differentially Private Learning of Thresholds and Quasi-Concave Optimization”, presented at STOC 2023, we describe a new paradigm that allows for slices to be selected adaptively and yet avoids composition cost. We show that DP algorithms for multiple fundamental aggregation and learning tasks can be expressed in this Reorder-Slice-Compute (RSC) paradigm, gaining significant improvements in utility.

The Reorder-Slice-Compute (RSC) paradigm

An algorithm A falls in the RSC paradigm if it can be expressed in the following general form (see visualization below). The input is a sensitive set D of data points. The algorithm then performs a sequence of k steps as follows:

1. Select an ordering over data points, a slice size m, and a DP algorithm M. The selection may depend on the output of A in prior steps (and hence is adaptive).
2. Slice out the (approximately) top m data points according to the order from the dataset D, apply M to the slice, and output the result.

A visualization of three Reorder-Slice-Compute (RSC) steps.

If we analyze the overall privacy loss of an RSC algorithm using DP composition theorems, the privacy guarantee suffers from the expected composition cost, i.e., it deteriorates with the square root of the number of steps k. To eliminate this composition cost, we provide a novel analysis that removes the dependence on k altogether: the overall privacy guarantee is close to that of a single step! The idea behind our tighter analysis is a novel technique that limits the potential cascade of affected steps when a single data point is modified (details in the paper).

Tighter privacy analysis means better utility. The effectiveness of DP algorithms is often stated in terms of the smallest input size (number of data points) that suffices in order to release a correct result that meets the privacy requirements. We describe several problems with algorithms that can be expressed in the RSC paradigm and for which our tighter analysis improved utility.

Private interval point

We start with the following basic aggregation task. The input is a dataset D of n points from an ordered domain X (think of the domain as the natural numbers between 1 and |X|). The goal is to return a point y in X that is in the interval of D, that is between the minimum and the maximum points in D.

The solution to the interval point problem is trivial without the privacy requirement: simply return any point in the dataset D. But this solution is not privacy-preserving as it discloses the presence of a particular datapoint in the input. We can also see that if there is only one point in the dataset, a privacy-preserving solution is not possible, as it must return that point. We can therefore ask the following fundamental question: What is the smallest input size N for which we can solve the private interval point problem?

It is known that N must increase with the domain size |X| and that this dependence is at least the iterated log function log* |X| [1, 2]. On the other hand, the best prior DP algorithm required the input size to be at least (log* |X|)^1.5. To close this gap, we designed an RSC algorithm that requires only an order of log* |X| points.

The iterated log function is extremely slow growing: It is the number of times we need to take a logarithm of a value before we reach a value that is equal to or smaller than 1. How did this function naturally come out in the analysis? Each step of the RSC algorithm remapped the domain to a logarithm of its prior size. Therefore there were log* |X| steps in total. The tighter RSC analysis eliminated a square root of the number of steps from the required input size.

Even though the interval point task seems very basic, it captures the essence of the difficulty of private solutions for common aggregation tasks. We next describe two of these tasks and express the required input size to these tasks in terms of N.

Private approximate median

One of these common aggregation tasks is approximate median: The input is a dataset D of n points from an ordered domain X. The goal is to return a point y that is between the ⅓ and ⅔ quantiles of D. That is, at least a third of the points in D are smaller or equal to y and at least a third of the points are larger or equal to y. Note that returning an exact median is not possible with differential privacy, since it discloses the presence of a datapoint. Hence we consider the relaxed requirement of an approximate median (shown below).

We can compute an approximate median by finding an interval point: We slice out the N smallest points and the N largest points and then compute an interval point of the remaining points. The latter must be an approximate median. This works when the dataset size is at least 3N.

An example of a data D over domain X, the set of interval points, and the set of approximate medians.

Private learning of axis-aligned rectangles

For the next task, the input is a set of n labeled data points, where each point x = (x₁,….,x_d) is a d-dimensional vector over a domain X. Displayed below, the goal is to learn values a_i , b_i for the axes i=1,…,d that define a d-dimensional rectangle, so that for each example x

• If x is positively labeled (shown as red plus signs below) then it lies within the rectangle, that is, for all axes i, x_i is in the interval [a_i ,b_i], and
• If x is negatively labeled (shown as blue minus signs below) then it lies outside the rectangle, that is, for at least one axis i, x_i is outside the interval [a_i ,b_i].

A set of 2-dimensional labeled points and a respective rectangle.

Any DP solution for this problem must be approximate in that the learned rectangle must be allowed to mislabel some data points, with some positively labeled points outside the rectangle or negatively labeled points inside it. This is because an exact solution could be very sensitive to the presence of a particular data point and would not be private. The goal is a DP solution that keeps this necessary number of mislabeled points small.

We first consider the one-dimensional case (d = 1). We are looking for an interval [a,b] that covers all positive points and none of the negative points. We show that we can do this with at most 2N mislabeled points. We focus on the positively labeled points. In the first RSC step we slice out the N smallest points and compute a private interval point as a. We then slice out the N largest points and compute a private interval point as b. The solution [a,b] correctly labels all negatively labeled points and mislabels at most 2N of the positively labeled points. Thus, at most ~2N points are mislabeled in total.

Illustration for d = 1, we slice out N left positive points and compute an interval point a, slice out N right positive points and compute an interval point b.

With d > 1, we iterate over the axes i = 1,….,d and apply the above for the i^th coordinates of input points to obtain the values a_i , b_i. In each iteration, we perform two RSC steps and slice out 2N positively labeled points. In total, we slice out 2dN points and all remaining points were correctly labeled. That is, all negatively-labeled points are outside the final d-dimensional rectangle and all positively-labeled points, except perhaps ~2dN, lie inside the rectangle. Note that this algorithm uses the full flexibility of RSC in that the points are ordered differently by each axis. Since we perform d steps, the RSC analysis shaves off a factor of square root of d from the number of mislabeled points.

Training ML models with adaptive selection of training examples

The training efficiency or performance of ML models can sometimes be improved by selecting training examples in a way that depends on the current state of the model, e.g., self-paced curriculum learning or active learning.

The most common method for private training of ML models is DP-SGD, where noise is added to the gradient update from each minibatch of training examples. Privacy analysis with DP-SGD typically assumes that training examples are randomly partitioned into minibatches. But if we impose a data-dependent selection order on training examples, and further modify the selection criteria k times during training, then analysis through DP composition results in deterioration of the privacy guarantees of a magnitude equal to the square root of k.

Fortunately, example selection with DP-SGD can be naturally expressed in the RSC paradigm: each selection criteria reorders the training examples and each minibatch is a slice (for which we compute a noisy gradient). With RSC analysis, there is no privacy deterioration with k, which brings DP-SGD training with example selection into the practical domain.

Conclusion

The RSC paradigm was introduced in order to tackle an open problem that is primarily of theoretical significance, but turns out to be a versatile tool with the potential to enhance data efficiency in production environments.

Acknowledgments

The work described here was done jointly with Xin Lyu, Jelani Nelson, and Tamas Sarlos.

Posted by Shantanu Shahane, Software Engineer, and Matthias Ihme, Research Scientist, Athena Team

Turbulence is ubiquitous in environmental and engineering fluid flows, and is encountered routinely in everyday life. A better understanding of these turbulent processes could provide valuable insights across a variety of research areas — improving the prediction of cloud formation by atmospheric transport and the spreading of wildfires by turbulent energy exchange, understanding sedimentation of deposits in rivers, and improving the efficiency of combustion in aircraft engines to reduce emissions, to name a few. However, despite its importance, our current understanding and our ability to reliably predict such flows remains limited. This is mainly attributed to the highly chaotic nature and the enormous spatial and temporal scales these fluid flows occupy, ranging from energetic, large-scale movements on the order of several meters on the high-end, where energy is injected into the fluid flow, all the way down to micrometers (μm) on the low-end, where the turbulence is dissipated into heat by viscous friction.

A powerful tool to understand these turbulent flows is the direct numerical simulation (DNS), which provides a detailed representation of the unsteady three-dimensional flow-field without making any approximations or simplifications. More specifically, this approach utilizes a discrete grid with small enough grid spacing to capture the underlying continuous equations that govern the dynamics of the system (in this case, variable-density Navier-Stokes equations, which govern all fluid flow dynamics). When the grid spacing is small enough, the discrete grid points are enough to represent the true (continuous) equations without the loss of accuracy. While this is attractive, such simulations require tremendous computational resources in order to capture the correct fluid-flow behaviors across such a wide range of spatial scales.

The actual span in spatial resolution to which direct numerical calculations must be applied depends on the task and is determined by the Reynolds number, which compares inertial to viscous forces. Typically, the Reynolds number can range between 10² up to 10⁷ (even larger for atmospheric or interstellar problems). In 3D, the grid size for the resolution required scales roughly with the Reynolds number to the power of 4.5! Because of this strong scaling dependency, simulating such flows is generally limited to flow regimes with moderate Reynolds numbers, and typically requires access to high-performance computing systems with millions of CPU/GPU cores.

In “A TensorFlow simulation framework for scientific computing of fluid flows on tensor processing units”, we introduce a new simulation framework that enables the computation of fluid flows with TPUs. By leveraging latest advances on TensorFlow software and TPU-hardware architecture, this software tool allows detailed large-scale simulations of turbulent flows at unprecedented scale, pushing the boundaries of scientific discovery and turbulence analysis. We demonstrate that this framework scales efficiently to accommodate the scale of the problem or, alternatively, improved run times, which is remarkable since most large-scale distributed computation frameworks exhibit reduced efficiency with scaling. The software is available as an open-source project on GitHub.

Large-scale scientific computation with accelerators

The software solves variable-density Navier-Stokes equations on TPU architectures using the TensorFlow framework. The single-instruction, multiple-data (SIMD) approach is adopted for parallelization of the TPU solver implementation. The finite difference operators on a colocated structured mesh are cast as filters of the convolution function of TensorFlow, leveraging TPU’s matrix multiply unit (MXU). The framework takes advantage of the low-latency high-bandwidth inter-chips interconnect (ICI) between the TPU accelerators. In addition, by leveraging the single-precision floating-point computations and highly optimized executable through the accelerated linear algebra (XLA) compiler, it’s possible to perform large-scale simulations with excellent scaling on TPU hardware architectures.

This research effort demonstrates that the graph-based TensorFlow in combination with new types of ML special purpose hardware, can be used as a programming paradigm to solve partial differential equations representing multiphysics flows. The latter is achieved by augmenting the Navier-Stokes equations with physical models to account for chemical reactions, heat-transfer, and density changes to enable, for example, simulations of cloud formation and wildfires.

It’s worth noting that this framework is the first open-source computational fluid dynamics (CFD) framework for high-performance, large-scale simulations to fully leverage the cloud accelerators that have become common (and become a commodity) with the advancement of machine learning (ML) in recent years. While our work focuses on using TPU accelerators, the code can be easily adjusted for other accelerators, such as GPU clusters.

This framework demonstrates a way to greatly reduce the cost and turn-around time associated with running large-scale scientific CFD simulations and enables even greater iteration speed in fields, such as climate and weather research. Since the framework is implemented using TensorFlow, an ML language, it also enables the ready integration with ML methods and allows the exploration of ML approaches on CFD problems. With the general accessibility of TPU and GPU hardware, this approach lowers the barrier for researchers to contribute to our understanding of large-scale turbulent systems.

Framework validation and homogeneous isotropic turbulence

Beyond demonstrating the performance and the scaling capabilities, it is also critical to validate the correctness of this framework to ensure that when it is used for CFD problems, we get reasonable results. For this purpose, researchers typically use idealized benchmark problems during CFD solver development, many of which we adopted in our work (more details in the paper).

One such benchmark for turbulence analysis is homogeneous isotropic turbulence (HIT), which is a canonical and well studied flow in which the statistical properties, such as kinetic energy, are invariant under translations and rotations of the coordinate axes. By pushing the resolution to the limits of the current state of the art, we were able to perform direct numerical simulations with more than eight billion degrees of freedom — equivalent to a three-dimensional mesh with 2,048 grid points along each of the three directions. We used 512 TPU-v4 cores, distributing the computation of the grid points along the x, y, and z axes to a distribution of [2,2,128] cores, respectively, optimized for the performance on TPU. The wall clock time per timestep was around 425 milliseconds and the flow was simulated for a total of 400,000 timesteps. 50 TB data, which includes the velocity and density fields, is stored for 400 timesteps (every 1,000th step). To our knowledge, this is one of the largest turbulent flow simulations of its kind conducted to date.

Due to the complex, chaotic nature of the turbulent flow field, which extends across several magnitudes of resolution, simulating the system in high resolution is necessary. Because we employ a fine-resolution grid with eight billion points, we are able to accurately resolve the field.

Contours of x-component of velocity along the z midplane. The high resolution of the simulation is critical to accurately represent the turbulent field.

The turbulent kinetic energy and dissipation rates are two statistical quantities commonly used to analyze a turbulent flow. The temporal decay of these properties in a turbulent field without additional energy injection is due to viscous dissipation and the decay asymptotes follow the expected analytical power law. This is in agreement with the theoretical asymptotes and observations reported in the literature and thus, validates our framework.

Solid line: Temporal evolution of turbulent kinetic energy (k). Dashed line: Analytical power laws for decaying homogeneous isotropic turbulence (n=1.3) (Ⲧl: eddy turnover time).

Solid line: Temporal evolution of dissipation rate (ε). Dashed line: Analytical power laws for decaying homogeneous isotropic turbulence (n=1.3).

The energy spectrum of a turbulent flow represents the energy content across wavenumber, where the wavenumber k is proportional to the inverse wavelength λ (i.e., k ∝ 1/λ). Generally, the spectrum can be qualitatively divided into three ranges: source range, inertial range and viscous dissipative range (from left to right on the wavenumber axis, below). The lowest wavenumbers in the source range correspond to the largest turbulent eddies, which have the most energy content. These large eddies transfer energy to turbulence in the intermediate wavenumbers (inertial range), which is statistically isotropic (i.e., essentially uniform in all directions). The smallest eddies, corresponding to the largest wavenumbers, are dissipated into thermal energy by the viscosity of the fluid. By virtue of the fine grid having 2,048 points in each of the three spatial directions, we are able to resolve the flow field up to the length scale at which viscous dissipation takes place. This direct numerical simulation approach is the most accurate as it does not require any closure model to approximate the energy cascade below the grid size.

Spectrum of turbulent kinetic energy at different time instances. The spectrum is normalized by the instantaneous integral length (l) and the turbulent kinetic energy (k).

A new era for turbulent flows research

More recently, we extended this framework to predict wildfires and atmospheric flows, which is relevant for climate-risk assessment. Apart from enabling high-fidelity simulations of complex turbulent flows, this simulation framework also provides capabilities for scientific machine learning (SciML) — for example, downsampling from a fine to a coarse grid (model reduction) or building models that run at lower resolution while still capturing the correct dynamic behaviors. It could also provide avenues for further scientific discovery, such as building ML-based models to better parameterize microphysics of turbulent flows, including physical relationships between temperature, pressure, vapor fraction, etc., and could improve upon various control tasks, e.g., to reduce the energy consumption of buildings or find more efficient propeller shapes. While attractive, a main bottleneck in SciML has been the availability of data for training. To explore this, we have been working with groups at Stanford and Kaggle to make the data from our high-resolution HIT simulation available through a community-hosted web-platform, BLASTNet, to provide broad access to high-fidelity data to the research community via a network-of-datasets approach. We hope that the availability of these emerging high-fidelity simulation tools in conjunction with community-driven datasets will lead to significant advances in various areas of fluid mechanics.

Acknowledgements

We would like to thank Qing Wang, Yi-Fan Chen, and John Anderson for consulting and advice, Tyler Russell and Carla Bromberg for program management.

Posted by Si-An Chen, Student Researcher, Cloud AI Team, and Chun-Liang Li, Research Scientist, Cloud AI Team

Time series forecasting is critical to various real-world applications, from demand forecasting to pandemic spread prediction. In multivariate time series forecasting (forecasting multiple variants at the same time), one can split existing methods into two categories: univariate models and multivariate models. Univariate models focus on inter-series interactions or temporal patterns that encompass trends and seasonal patterns on a time series with a single variable. Examples of such trends and seasonal patterns might be the way mortgage rates increase due to inflation, and how traffic peaks during rush hour. In addition to inter-series patterns, multivariate models process intra-series features, known as cross-variate information, which is especially useful when one series is an advanced indicator of another series. For example, a rise in body weight may cause an increase in blood pressure, and increasing the price of a product may lead to a decrease in sales. Multivariate models have recently become popular solutions for multivariate forecasting as practitioners believe their capability of handling cross-variate information may lead to better performance.

In recent years, deep learning Transformer-based architectures have become a popular choice for multivariate forecasting models due to their superior performance on sequence tasks. However, advanced multivariate models perform surprisingly worse than simple univariate linear models on commonly-used long-term forecasting benchmarks, such as Electricity Transformer Temperature (ETT), Electricity, Traffic, and Weather. These results raise two questions:

• Does cross-variate information benefit time series forecasting?
• When cross-variate information is not beneficial, can multivariate models still perform as well as univariate models?

In “TSMixer: An All-MLP Architecture for Time Series Forecasting”, we analyze the advantages of univariate linear models and reveal their effectiveness. Insights from this analysis lead us to develop Time-Series Mixer (TSMixer), an advanced multivariate model that leverages linear model characteristics and performs well on long-term forecasting benchmarks. To the best of our knowledge, TSMixer is the first multivariate model that performs as well as state-of-the-art univariate models on long-term forecasting benchmarks, where we show that cross-variate information is less beneficial. To demonstrate the importance of cross-variate information, we evaluate a more challenging real-world application, M5. Finally, empirical results show that TSMixer outperforms state-of-the-art models, such as PatchTST, Fedformer, Autoformer, DeepAR and TFT.

TSMixer architecture

A key difference between linear models and Transformers is how they capture temporal patterns. On one hand, linear models apply fixed and time-step-dependent weights to capture static temporal patterns, and are unable to process cross-variate information. On the other hand, Transformers use attention mechanisms that apply dynamic and data-dependent weights at each time step, capturing dynamic temporal patterns and enabling them to process cross-variate information.

In our analysis, we show that under common assumptions of temporal patterns, linear models have naïve solutions to perfectly recover the time series or place bounds on the error, which means they are great solutions for learning static temporal patterns of univariate time series more effectively. In contrast, it is non-trivial to find similar solutions for attention mechanisms, as the weights applied to each time step are dynamic. Consequently, we develop a new architecture by replacing Transformer attention layers with linear layers. The resulting TSMixer model, which is similar to the computer vision MLP-Mixer method, alternates between applications of the multi-layer perceptron in different directions, which we call time-mixing and feature-mixing, respectively. The TSMixer architecture efficiently captures both temporal patterns and cross-variate information, as shown in the figure below. The residual designs ensure that TSMixer retains the capacity of temporal linear models while still being able to exploit cross-variate information.

Transformer block and TSMixer block architectures. TSMixer replaces the multi-head attention layer with time-mixing, a linear model applied on the time dimension.

Comparison between data-dependent (attention mechanisms) and time-step-dependent (linear models). This is an example of forecasting the next time step by learning the weights of the previous three time steps.

Evaluation on long-term forecasting benchmarks

We evaluate TSMixer using seven popular long-term forecasting datasets (ETTm1, ETTm2, ETTh1, ETTh2, Electricity, Traffic, and Weather), where recent research has shown that univariate linear models outperform advanced multivariate models with large margins. We compare TSMixer with state-of-the-art multivariate models (TFT, FEDformer, Autoformer, Informer), and univariate models, including linear models and PatchTST. The figure below shows the average improvement of mean squared error (MSE) by TSMixer compared with others. The average is calculated across datasets and multiple forecasting horizons. We demonstrate that TSMixer significantly outperforms other multivariate models and performs on par with state-of-the-art univariate models. These results show that multivariate models are capable of performing as well as univariate models.

The average MSE improvement of TSMixer compared with other baselines. The red bars show multivariate methods and the blue bars show univariate methods. TSMixer achieves significant improvement over other multivariate models and achieves comparable results to univariate models.

Ablation study

We performed an ablation study to compare TSMixer with TMix-Only, a TSMixer variant that consists of time mixing layers only. The results show that TMix-Only performs almost the same as TSMixer, which means the additional feature mixing layers do not improve the performance and confirms that cross-variate information is less beneficial on popular benchmarks. The results validate the superior univariate model performance shown in previous research. However, existing long-term forecasting benchmarks are not well representative of the need for cross-variate information in some real-world applications where time series may be intermittent or sparse, hence temporal patterns may not be sufficient for forecasting. Therefore, it may be inappropriate to evaluate multivariate forecasting models solely on these benchmarks.

Evaluation on M5: Effectiveness of cross-variate information

To further demonstrate the benefit of multivariate models, we evaluate TSMixer on the challenging M5 benchmark, a large-scale retail dataset containing crucial cross-variate interactions. M5 contains the information of 30,490 products collected over 5 years. Each product description includes time series data, like daily sales, sell price, promotional event information, and static (non-time-series) features, such as store location and product category. The goal is to forecast the daily sales of each product for the next 28 days, evaluated using the weighted root mean square scaled error (WRMSSE) from the M5 competition. The complicated nature of retail makes it more challenging to forecast solely using univariate models that focus on temporal patterns, so multivariate models with cross-variate information and even auxiliary features are more essential.

First, we compare TSMixer to other methods only considering the historical data, such as daily sales and historical sell prices. The results show that multivariate models outperforms univariate models significantly, indicating the usefulness of cross-variate information. And among all compared methods, TSMixer effectively leverages the cross-variate information and achieves the best performance.

Additionally, to leverage more information, such as static features (e.g., store location, product category) and future time series (e.g., a promotional event scheduled in coming days) provided in M5, we propose a principle design to extend TSMixer. The extended TSMixer aligns different types of features into the same length, and then applies multiple mixing layers to the concatenated features to make predictions. The extended TSMixer architecture outperforms models popular in industrial applications, including DeepAR and TFT, showcasing its strong potential for real-world impact.

The architecture of the extended TSMixer. In the first stage (align stage), it aligns the different types of features into the same length before concatenating them. In the second stage (mixing stage) it applies multiple mixing layers conditioned with static features.

The WRMSSE on M5. The first three methods (blue) are univariate models. The middle three methods (orange) are multivariate models that consider only historical features. The last three methods (red) are multivariate models that consider historical, future, and static features.

Conclusion

We present TSMixer, an advanced multivariate model that leverages linear model characteristics and performs as well as state-of-the-art univariate models on long-term forecasting benchmarks. TSMixer creates new possibilities for the development of time series forecasting architectures by providing insights into the importance of cross-variate and auxiliary information in real-world scenarios. The empirical results highlight the need to consider more realistic benchmarks for multivariate forecasting models in future research. We hope that this work will inspire further exploration in the field of time series forecasting, and lead to the development of more powerful and effective models that can be applied to real-world applications.

Acknowledgements

This research was conducted by Si-An Chen, Chun-Liang Li, Nate Yoder, Sercan O. Arik, and Tomas Pfister.

Posted by Stephan Rasp, Research Scientist, and Carla Bromberg, Program Lead, Google Research

In 1950, weather forecasting started its digital revolution when researchers used the first programmable, general-purpose computer ENIAC to solve mathematical equations describing how weather evolves. In the more than 70 years since, continuous advancements in computing power and improvements to the model formulations have led to steady gains in weather forecast skill: a 7-day forecast today is about as accurate as a 5-day forecast in 2000 and a 3-day forecast in 1980. While improving forecast accuracy at the pace of approximately one day per decade may not seem like a big deal, every day improved is important in far reaching use cases, such as for logistics planning, disaster management, agriculture and energy production. This “quiet” revolution has been tremendously valuable to society, saving lives and providing economic value across many sectors.

Now we are seeing the start of yet another revolution in weather forecasting, this time fueled by advances in machine learning (ML). Rather than hard-coding approximations of the physical equations, the idea is to have algorithms learn how weather evolves from looking at large volumes of past weather data. Early attempts at doing so go back to 2018 but the pace picked up considerably in the last two years when several large ML models demonstrated weather forecasting skill comparable to the best physics-based models. Google’s MetNet [1, 2], for instance, demonstrated state-of-the-art capabilities for forecasting regional weather one day ahead. For global prediction, Google DeepMind created GraphCast, a graph neural network to make 10 day predictions at a horizontal resolution of 25 km, competitive with the best physics-based models in many skill metrics.

Apart from potentially providing more accurate forecasts, one key advantage of such ML methods is that, once trained, they can create forecasts in a matter of minutes on inexpensive hardware. In contrast, traditional weather forecasts require large super-computers that run for hours every day. Clearly, ML represents a tremendous opportunity for the weather forecasting community. This has also been recognized by leading weather forecasting centers, such as the European Centre for Medium-Range Weather Forecasts’ (ECMWF) machine learning roadmap or the National Oceanic and Atmospheric Administration’s (NOAA) artificial intelligence strategy.

To ensure that ML models are trusted and optimized for the right goal, forecast evaluation is crucial. Evaluating weather forecasts isn’t straightforward, however, because weather is an incredibly multi-faceted problem. Different end-users are interested in different properties of forecasts, for example, renewable energy producers care about wind speeds and solar radiation, while crisis response teams are concerned about the track of a potential cyclone or an impending heat wave. In other words, there is no single metric to determine what a “good” weather forecast is, and the evaluation has to reflect the multi-faceted nature of weather and its downstream applications. Furthermore, differences in the exact evaluation setup — e.g., which resolution and ground truth data is used — can make it difficult to compare models. Having a way to compare novel and established methods in a fair and reproducible manner is crucial to measure progress in the field.

To this end, we are announcing WeatherBench 2 (WB2), a benchmark for the next generation of data-driven, global weather models. WB2 is an update to the original benchmark published in 2020, which was based on initial, lower-resolution ML models. The goal of WB2 is to accelerate the progress of data-driven weather models by providing a trusted, reproducible framework for evaluating and comparing different methodologies. The official website contains scores from several state-of-the-art models (at the time of writing, these are Keisler (2022), an early graph neural network, Google DeepMind’s GraphCast and Huawei’s Pangu-Weather, a transformer-based ML model). In addition, forecasts from ECMWF’s high-resolution and ensemble forecasting systems are included, which represent some of the best traditional weather forecasting models.

Making evaluation easier

The key component of WB2 is an open-source evaluation framework that allows users to evaluate their forecasts in the same manner as other baselines. Weather forecast data at high-resolutions can be quite large, making even evaluation a computational challenge. For this reason, we built our evaluation code on Apache Beam, which allows users to split computations into smaller chunks and evaluate them in a distributed fashion, for example using DataFlow on Google Cloud. The code comes with a quick-start guide to help people get up to speed.

Additionally, we provide most of the ground-truth and baseline data on Google Cloud Storage in cloud-optimized Zarr format at different resolutions, for example, a comprehensive copy of the ERA5 dataset used to train most ML models. This is part of a larger Google effort to provide analysis-ready, cloud-optimized weather and climate datasets to the research community and beyond. Since downloading these data from the respective archives and converting them can be time-consuming and compute-intensive, we hope that this should considerably lower the entry barrier for the community.

Assessing forecast skill

Together with our collaborators from ECMWF, we defined a set of headline scores that best capture the quality of global weather forecasts. As the figure below shows, several of the ML-based forecasts have lower errors than the state-of-the-art physical models on deterministic metrics. This holds for a range of variables and regions, and underlines the competitiveness and promise of ML-based approaches.

This scorecard shows the skill of different models compared to ECMWF’s Integrated Forecasting System (IFS), one of the best physics-based weather forecasts, for several variables. IFS forecasts are evaluated against IFS analysis. All other models are evaluated against ERA5. The order of ML models reflects publication date.

Toward reliable probabilistic forecasts

However, a single forecast often isn’t enough. Weather is inherently chaotic because of the butterfly effect. For this reason, operational weather centers now run ~50 slightly perturbed realizations of their model, called an ensemble, to estimate the forecast probability distribution across various scenarios. This is important, for example, if one wants to know the likelihood of extreme weather.

Creating reliable probabilistic forecasts will be one of the next key challenges for global ML models. Regional ML models, such as Google’s MetNet already estimate probabilities. To anticipate this next generation of global models, WB2 already provides probabilistic metrics and baselines, among them ECMWF’s IFS ensemble, to accelerate research in this direction.

As mentioned above, weather forecasting has many aspects, and while the headline metrics try to capture the most important aspects of forecast skill, they are by no means sufficient. One example is forecast realism. Currently, many ML forecast models tend to “hedge their bets” in the face of the intrinsic uncertainty of the atmosphere. In other words, they tend to predict smoothed out fields that give lower average error but do not represent a realistic, physically consistent state of the atmosphere. An example of this can be seen in the animation below. The two data-driven models, Pangu-Weather and GraphCast (bottom), predict the large-scale evolution of the atmosphere remarkably well. However, they also have less small-scale structure compared to the ground truth or the physical forecasting model IFS HRES (top). In WB2 we include a range of these case studies and also a spectral metric that quantifies such blurring.

Forecasts of a front passing through the continental United States initialized on January 3, 2020. Maps show temperature at a pressure level of 850 hPa (roughly equivalent to an altitude of 1.5km) and geopotential at a pressure level of 500 hPa (roughly 5.5 km) in contours. ERA5 is the corresponding ground-truth analysis, IFS HRES is ECMWF’s physics-based forecasting model.

Conclusion

WeatherBench 2 will continue to evolve alongside ML model development. The official website will be updated with the latest state-of-the-art models. (To submit a model, please follow these instructions). We also invite the community to provide feedback and suggestions for improvements through issues and pull requests on the WB2 GitHub page.

Designing evaluation well and targeting the right metrics is crucial in order to make sure ML weather models benefit society as quickly as possible. WeatherBench 2 as it is now is just the starting point. We plan to extend it in the future to address key issues for the future of ML-based weather forecasting. Specifically, we would like to add station observations and better precipitation datasets. Furthermore, we will explore the inclusion of nowcasting and subseasonal-to-seasonal predictions to the benchmark.

We hope that WeatherBench 2 can aid researchers and end-users as weather forecasting continues to evolve.

Acknowledgements

WeatherBench 2 is the result of collaboration across many different teams at Google and external collaborators at ECMWF. From ECMWF, we would like to thank Matthew Chantry, Zied Ben Bouallegue and Peter Dueben. From Google, we would like to thank the core contributors to the project: Stephan Rasp, Stephan Hoyer, Peter Battaglia, Alex Merose, Ian Langmore, Tyler Russell, Alvaro Sanchez, Antonio Lobato, Laurence Chiu, Rob Carver, Vivian Yang, Shreya Agrawal, Thomas Turnbull, Jason Hickey, Carla Bromberg, Jared Sisk, Luke Barrington, Aaron Bell, and Fei Sha. We also would like to thank Kunal Shah, Rahul Mahrsee, Aniket Rawat, and Satish Kumar. Thanks to John Anderson for sponsoring WeatherBench 2. Furthermore, we would like to thank Kaifeng Bi from the Pangu-Weather team and Ryan Keisler for their help in adding their models to WeatherBench 2.

Posted by Vikas Bahirwani, Research Scientist, and Susan Xu, Software Engineer, Google Augmented Reality

Automatic speech recognition (ASR) technology has made conversations more accessible with live captions in remote conferencing software, mobile applications, and head-worn displays. However, to maintain real-time responsiveness, live caption systems often display interim predictions that are updated as new utterances are received. This can cause text instability (a “flicker” where previously displayed text is updated, shown in the captions on the left in the video below), which can impair users’ reading experience due to distraction, fatigue, and difficulty following the conversation.

In “Modeling and Improving Text Stability in Live Captions”, presented at ACM CHI 2023, we formalize this problem of text stability through a few key contributions. First, we quantify the text instability by employing a vision-based flicker metric that uses luminance contrast and discrete Fourier transform. Second, we also introduce a stability algorithm to stabilize the rendering of live captions via tokenized alignment, semantic merging, and smooth animation. Finally, we conducted a user study (N=123) to understand viewers’ experience with live captioning. Our statistical analysis demonstrates a strong correlation between our proposed flicker metric and viewers’ experience. Furthermore, it shows that our proposed stabilization techniques significantly improves viewers’ experience (e.g., the captions on the right in the video above).

Raw ASR captions vs. stabilized captions

Metric

Inspired by previous work, we propose a flicker-based metric to quantify text stability and objectively evaluate the performance of live captioning systems. Specifically, our goal is to quantify the flicker in a grayscale live caption video. We achieve this by comparing the difference in luminance between individual frames (frames in the figures below) that constitute the video. Large visual changes in luminance are obvious (e.g., addition of the word “bright” in the figure on the bottom), but subtle changes (e.g., update from “… this gold. Nice..” to “… this. Gold is nice”) may be difficult to discern for readers. However, converting the change in luminance to its constituting frequencies exposes both the obvious and subtle changes.

Thus, for each pair of contiguous frames, we convert the difference in luminance into its constituting frequencies using discrete Fourier transform. We then sum over each of the low and high frequencies to quantify the flicker in this pair. Finally, we average over all of the frame-pairs to get a per-video flicker.

For instance, we can see below that two identical frames (top) yield a flicker of 0, while two non-identical frames (bottom) yield a non-zero flicker. It is worth noting that higher values of the metric indicate high flicker in the video and thus, a worse user experience than lower values of the metric.

Illustration of the flicker metric between two identical frames.

Illustration of the flicker between two non-identical frames.

Stability algorithm

To improve the stability of live captions, we propose an algorithm that takes as input already rendered sequence of tokens (e.g., “Previous” in the figure below) and the new sequence of ASR predictions, and outputs an updated stabilized text (e.g., “Updated text (with stabilization)” below). It considers both the natural language understanding (NLU) aspect as well as the ergonomic aspect (display, layout, etc.) of the user experience in deciding when and how to produce a stable updated text. Specifically, our algorithm performs tokenized alignment, semantic merging, and smooth animation to achieve this goal. In what follows, a token is defined as a word or punctuation produced by ASR.

We show (a) the previously already rendered text, (b) the baseline layout of updated text without our merging algorithm, and (c) the updated text as generated by our stabilization algorithm.

Our algorithm address the challenge of producing stabilized updated text by first identifying three classes of changes (highlighted in red, green, and blue below):

1. Red: Addition of tokens to the end of previously rendered captions (e.g., “How about”).
2. Green: Addition / deletion of tokens, in the middle of already rendered captions.
   • B1: Addition of tokens (e.g., “I” and “friends”). These may or may not affect the overall comprehension of the captions, but may lead to layout change. Such layout changes are not desired in live captions as they cause significant jitter and poorer user experience. Here “I” does not add to the comprehension but “friends” does. Thus, it is important to balance updates with stability specially for B1 type tokens.
   • B2: Removal of tokens, e.g., “in” is removed in the updated sentence.
3. Blue: Re-captioning of tokens: This includes token edits that may or may not have an impact on the overall comprehension of the captions.

• C1: Proper nouns like “disney land” are updated to “Disneyland”.
• C2: Grammatical shorthands like “it’s” are updated to “It was”.

Classes of changes between previously displayed and updated text.

Alignment, merging, and smoothing

To maximize text stability, our goal is to align the old sequence with the new sequence using updates that make minimal changes to the existing layout while ensuring accurate and meaningful captions. To achieve this, we leverage a variant of the Needleman-Wunsch algorithm with dynamic programming to merge the two sequences depending on the class of tokens as defined above:

• Case A tokens: We directly add case A tokens, and line breaks as needed to fit the updated captions.
• Case B tokens: Our preliminary studies showed that users preferred stability over accuracy for previously displayed captions. Thus, we only update case B tokens if the updates do not break an existing line layout.
• Case C tokens: We compare the semantic similarity of case C tokens by transforming original and updated sentences into sentence embeddings, measuring their dot-product, and updating them only if they are semantically different (similarity < 0.85) and the update will not cause new line breaks.

Finally, we leverage animations to reduce visual jitter. We implement smooth scrolling and fading of newly added tokens to further stabilize the overall layout of the live captions.

User evaluation

We conducted a user study with 123 participants to (1) examine the correlation of our proposed flicker metric with viewers’ experience of the live captions, and (2) assess the effectiveness of our stabilization techniques.

We manually selected 20 videos in YouTube to obtain a broad coverage of topics including video conferences, documentaries, academic talks, tutorials, news, comedy, and more. For each video, we selected a 30-second clip with at least 90% speech.

We prepared four types of renderings of live captions to compare:

1. Raw ASR: raw speech-to-text results from a speech-to-text API.
2. Raw ASR + thresholding: only display interim speech-to-text result if its confidence score is higher than 0.85.
3. Stabilized captions: captions using our algorithm described above with alignment and merging.
4. Stabilized and smooth captions: stabilized captions with smooth animation (scrolling + fading) to assess whether softened display experience helps improve the user experience.

We collected user ratings by asking the participants to watch the recorded live captions and rate their assessments of comfort, distraction, ease of reading, ease of following the video, fatigue, and whether the captions impaired their experience.

Correlation between flicker metric and user experience

We calculated Spearman’s coefficient between the flicker metric and each of the behavioral measurements (values range from -1 to 1, where negative values indicate a negative relationship between the two variables, positive values indicate a positive relationship, and zero indicates no relationship). Shown below, our study demonstrates statistically significant (𝑝 < 0.001) correlations between our flicker metric and users’ ratings. The absolute values of the coefficient are around 0.3, indicating a moderate relationship.

Behavioral Measurement	Correlation to Flickering Metric*
Comfort	-0.29
Distraction	0.33
Easy to read	-0.31
Easy to follow videos	-0.29
Fatigue	0.36
Impaired Experience	0.31

Spearman correlation tests of our proposed flickering metric. *p < 0.001.

Stabilization of live captions

Our proposed technique (stabilized smooth captions) received consistently better ratings, significant as measured by the Mann-Whitney U test (p < 0.01 in the figure below), in five out of six aforementioned survey statements. That is, users considered the stabilized captions with smoothing to be more comfortable and easier to read, while feeling less distraction, fatigue, and impairment to their experience than other types of rendering.

User ratings from 1 (Strongly Disagree) – 7 (Strongly Agree) on survey statements. (**: p<0.01, ***: p<0.001; ****: p<0.0001; ns: non-significant)

Conclusion and future direction

Text instability in live captioning significantly impairs users’ reading experience. This work proposes a vision-based metric to model caption stability that statistically significantly correlates with users’ experience, and an algorithm to stabilize the rendering of live captions. Our proposed solution can be potentially integrated into existing ASR systems to enhance the usability of live captions for a variety of users, including those with translation needs or those with hearing accessibility needs.

Our work represents a substantial step towards measuring and improving text stability. This can be evolved to include language-based metrics that focus on the consistency of the words and phrases used in live captions over time. These metrics may provide a reflection of user discomfort as it relates to language comprehension and understanding in real-world scenarios. We are also interested in conducting eye-tracking studies (e.g., videos shown below) to track viewers’ gaze patterns, such as eye fixation and saccades, allowing us to better understand the types of errors that are most distracting and how to improve text stability for those.

Illustration of tracking a viewer’s gaze when reading raw ASR captions.

Illustration of tracking a viewer’s gaze when reading stabilized and smoothed captions.

By improving text stability in live captions, we can create more effective communication tools and improve how people connect in everyday conversations in familiar or, through translation, unfamiliar languages.

Acknowledgements

This work is a collaboration across multiple teams at Google. Key contributors include Xingyu “Bruce” Liu, Jun Zhang, Leonardo Ferrer, Susan Xu, Vikas Bahirwani, Boris Smus, Alex Olwal, and Ruofei Du. We wish to extend our thanks to our colleagues who provided assistance, including Nishtha Bhatia, Max Spear, and Darcy Philippon. We would also like to thank Lin Li, Evan Parker, and CHI 2023 reviewers.

Posted by Yujin Tang and Wenhao Yu, Research Scientists, Google

Simple and effective interaction between human and quadrupedal robots paves the way towards creating intelligent and capable helper robots, forging a future where technology enhances our lives in ways beyond our imagination. Key to such human-robot interaction systems is enabling quadrupedal robots to respond to natural language instructions. Recent developments in large language models (LLMs) have demonstrated the potential to perform high-level planning. Yet, it remains a challenge for LLMs to comprehend low-level commands, such as joint angle targets or motor torques, especially for inherently unstable legged robots, necessitating high-frequency control signals. Consequently, most existing work presumes the provision of high-level APIs for LLMs to dictate robot behavior, inherently limiting the system’s expressive capabilities.

In “SayTap: Language to Quadrupedal Locomotion”, we propose an approach that uses foot contact patterns (which refer to the sequence and manner in which a four-legged agent places its feet on the ground while moving) as an interface to bridge human commands in natural language and a locomotion controller that outputs low-level commands. This results in an interactive quadrupedal robot system that allows users to flexibly craft diverse locomotion behaviors (e.g., a user can ask the robot to walk, run, jump or make other movements using simple language). We contribute an LLM prompt design, a reward function, and a method to expose the SayTap controller to the feasible distribution of contact patterns. We demonstrate that SayTap is a controller capable of achieving diverse locomotion patterns that can be transferred to real robot hardware.

SayTap method

The SayTap approach uses a contact pattern template, which is a 4 X T matrix of 0s and 1s, with 0s representing an agent’s feet in the air and 1s for feet on the ground. From top to bottom, each row in the matrix gives the foot contact patterns of the front left (FL), front right (FR), rear left (RL) and rear right (RR) feet. SayTap’s control frequency is 50 Hz, so each 0 or 1 lasts 0.02 seconds. In this work, a desired foot contact pattern is defined by a cyclic sliding window of size L_w and of shape 4 X L_w. The sliding window extracts from the contact pattern template four foot ground contact flags, which indicate if a foot is on the ground or in the air between t + 1 and t + L_w. The figure below provides an overview of the SayTap method.

SayTap introduces these desired foot contact patterns as a new interface between natural language user commands and the locomotion controller. The locomotion controller is used to complete the main task (e.g., following specified velocities) and to place the robot’s feet on the ground at the specified time, such that the realized foot contact patterns are as close to the desired contact patterns as possible. To achieve this, the locomotion controller takes the desired foot contact pattern at each time step as its input in addition to the robot’s proprioceptive sensory data (e.g., joint positions and velocities) and task-related inputs (e.g., user-specified velocity commands). We use deep reinforcement learning to train the locomotion controller and represent it as a deep neural network. During controller training, a random generator samples the desired foot contact patterns, the policy is then optimized to output low-level robot actions to achieve the desired foot contact pattern. Then at test time a LLM translates user commands into foot contact patterns.

SayTap approach overview.

SayTap uses foot contact patterns (e.g., 0 and 1 sequences for each foot in the inset, where 0s are foot in the air and 1s are foot on the ground) as an interface that bridges natural language user commands and low-level control commands. With a reinforcement learning-based locomotion controller that is trained to realize the desired contact patterns, SayTap allows a quadrupedal robot to take both simple and direct instructions (e.g., “Trot forward slowly.”) as well as vague user commands (e.g., “Good news, we are going to a picnic this weekend!”) and react accordingly.

We demonstrate that the LLM is capable of accurately mapping user commands into foot contact pattern templates in specified formats when given properly designed prompts, even in cases when the commands are unstructured or vague. In training, we use a random pattern generator to produce contact pattern templates that are of various pattern lengths T, foot-ground contact ratios within a cycle based on a given gait type G, so that the locomotion controller gets to learn on a wide distribution of movements leading to better generalization. See the paper for more details.

Results

With a simple prompt that contains only three in-context examples of commonly seen foot contact patterns, an LLM can translate various human commands accurately into contact patterns and even generalize to those that do not explicitly specify how the robot should react.

SayTap prompts are concise and consist of four components: (1) general instruction that describes the tasks the LLM should accomplish; (2) gait definition that reminds the LLM of basic knowledge about quadrupedal gaits and how they can be related to emotions; (3) output format definition; and (4) examples that give the LLM chances to learn in-context. We also specify five velocities that allow a robot to move forward or backward, fast or slow, or remain still.

General instruction block
You are a dog foot contact pattern expert.
Your job is to give a velocity and a foot contact pattern based on the input.
You will always give the output in the correct format no matter what the input is.

Gait definition block
The following are description about gaits:
1. Trotting is a gait where two diagonally opposite legs strike the ground at the same time.
2. Pacing is a gait where the two legs on the left/right side of the body strike the ground at the same time.
3. Bounding is a gait where the two front/rear legs strike the ground at the same time. It has a longer suspension phase where all feet are off the ground, for example, for at least 25% of the cycle length. This gait also gives a happy feeling.

Output format definition block
The following are rules for describing the velocity and foot contact patterns:
1. You should first output the velocity, then the foot contact pattern.
2. There are five velocities to choose from: [-1.0, -0.5, 0.0, 0.5, 1.0].
3. A pattern has 4 lines, each of which represents the foot contact pattern of a leg.
4. Each line has a label. "FL" is front left leg, "FR" is front right leg, "RL" is rear left leg, and "RR" is rear right leg.
5. In each line, "0" represents foot in the air, "1" represents foot on the ground.

Example block
Input: Trot slowly
Output: 0.5
FL: 11111111111111111000000000
FR: 00000000011111111111111111
RL: 00000000011111111111111111
RR: 11111111111111111000000000

Input: Bound in place
Output: 0.0
FL: 11111111111100000000000000
FR: 11111111111100000000000000
RL: 00000011111111111100000000
RR: 00000011111111111100000000

Input: Pace backward fast
Output: -1.0
FL: 11111111100001111111110000
FR: 00001111111110000111111111
RL: 11111111100001111111110000
RR: 00001111111110000111111111

Input:

SayTap prompt to the LLM. Texts in blue are used for illustration and are not input to LLM.

Following simple and direct commands

We demonstrate in the videos below that the SayTap system can successfully perform tasks where the commands are direct and clear. Although some commands are not covered by the three in-context examples, we are able to guide the LLM to express its internal knowledge from the pre-training phase via the “Gait definition block” (see the second block in our prompt above) in the prompt.

Following unstructured or vague commands

But what is more interesting is SayTap’s ability to process unstructured and vague instructions. With only a little hint in the prompt to connect certain gaits with general impressions of emotions, the robot bounds up and down when hearing exciting messages, like “We are going to a picnic!” Furthermore, it also presents the scenes accurately (e.g., moving quickly with its feet barely touching the ground when told the ground is very hot).

Conclusion and future work

We present SayTap, an interactive system for quadrupedal robots that allows users to flexibly craft diverse locomotion behaviors. SayTap introduces desired foot contact patterns as a new interface between natural language and the low-level controller. This new interface is straightforward and flexible, moreover, it allows a robot to follow both direct instructions and commands that do not explicitly state how the robot should react.

One interesting direction for future work is to test if commands that imply a specific feeling will allow the LLM to output a desired gait. In the gait definition block shown in the results section above, we provide a sentence that connects a happy mood with bounding gaits. We believe that providing more information can augment the LLM’s interpretations (e.g., implied feelings). In our evaluation, the connection between a happy feeling and a bounding gait led the robot to act vividly when following vague human commands. Another interesting direction for future work is to introduce multi-modal inputs, such as videos and audio. Foot contact patterns translated from those signals will, in theory, still work with our pipeline and will unlock many more interesting use cases.

Acknowledgements

Yujin Tang, Wenhao Yu, Jie Tan, Heiga Zen, Aleksandra Faust and Tatsuya Harada conducted this research. This work was conceived and performed while the team was in Google Research and will be continued at Google DeepMind. The authors would like to thank Tingnan Zhang, Linda Luu, Kuang-Huei Lee, Vincent Vanhoucke and Douglas Eck for their valuable discussions and technical support in the experiments.

Posted by Dahun Kim and Weicheng Kuo, Research Scientists, Google

The ability to detect objects in the visual world is crucial for computer vision and machine intelligence, enabling applications like adaptive autonomous agents and versatile shopping systems. However, modern object detectors are limited by the manual annotations of their training data, resulting in a vocabulary size significantly smaller than the vast array of objects encountered in reality. To overcome this, the open-vocabulary detection task (OVD) has emerged, utilizing image-text pairs for training and incorporating new category names at test time by associating them with the image content. By treating categories as text embeddings, open-vocabulary detectors can predict a wide range of unseen objects. Various techniques such as image-text pre-training, knowledge distillation, pseudo labeling, and frozen models, often employing convolutional neural network (CNN) backbones, have been proposed. With the growing popularity of vision transformers (ViTs), it is important to explore their potential for building proficient open-vocabulary detectors.

The existing approaches assume the availability of pre-trained vision-language models (VLMs) and focus on fine-tuning or distillation from these models to address the disparity between image-level pre-training and object-level fine-tuning. However, as VLMs are primarily designed for image-level tasks like classification and retrieval, they do not fully leverage the concept of objects or regions during the pre-training phase. Thus, it could be beneficial for open-vocabulary detection if we build locality information into the image-text pre-training.

In “RO-ViT: Region-Aware Pretraining for Open-Vocabulary Object Detection with Vision Transformers”, presented at CVPR 2023, we introduce a simple method to pre-train vision transformers in a region-aware manner to improve open-vocabulary detection. In vision transformers, positional embeddings are added to image patches to encode information about the spatial position of each patch within the image. Standard pre-training typically uses full-image positional embeddings, which does not generalize well to detection tasks. Thus, we propose a new positional embedding scheme, called “cropped positional embedding”, that better aligns with the use of region crops in detection fine-tuning. In addition, we replace the softmax cross entropy loss with focal loss in contrastive image-text learning, allowing us to learn from more challenging and informative examples. Finally, we leverage recent advances in novel object proposals to enhance open-vocabulary detection fine-tuning, which is motivated by the observation that existing methods often miss novel objects during the proposal stage due to overfitting to foreground categories. We are also releasing the code here.

Region-aware image-text pre-training

Existing VLMs are trained to match an image as a whole to a text description. However, we observe there is a mismatch between the way the positional embeddings are used in the existing contrastive pre-training approaches and open-vocabulary detection. The positional embeddings are important to transformers as they provide the information of where each element in the set comes from. This information is often useful for downstream recognition and localization tasks. Pre-training approaches typically apply full-image positional embeddings during training, and use the same positional embeddings for downstream tasks, e.g., zero-shot recognition. However, the recognition occurs at region-level for open-vocabulary detection fine-tuning, which requires the full-image positional embeddings to generalize to regions that they never see during the pre-training.

To address this, we propose cropped positional embeddings (CPE). With CPE, we upsample positional embeddings from the image size typical for pre-training, e.g., 224×224 pixels, to that typical for detection tasks, e.g., 1024×1024 pixels. Then we randomly crop and resize a region, and use it as the image-level positional embeddings during pre-training. The position, scale, and aspect ratio of the crop is randomly sampled. Intuitively, this causes the model to view an image not as a full image in itself, but as a region crop from some larger unknown image. This better matches the downstream use case of detection where recognition occurs at region- rather than image-level.

For the pre-training, we propose cropped positional embedding (CPE) which randomly crops and resizes a region of positional embeddings instead of using the whole-image positional embedding (PE). In addition, we use focal loss instead of the common softmax cross entropy loss for contrastive learning.

We also find it beneficial to learn from hard examples with a focal loss. Focal loss enables finer control over how hard examples are weighted than what the softmax cross entropy loss can provide. We adopt the focal loss and replace it with the softmax cross entropy loss in both image-to-text and text-to-image losses. Both CPE and focal loss introduce no extra parameters and minimal computation costs.

Open-vocabulary detector fine-tuning

An open-vocabulary detector is trained with the detection labels of ‘base’ categories, but needs to detect the union of ‘base’ and ‘novel’ (unlabeled) categories at test time. Despite the backbone features pre-trained from the vast open-vocabulary data, the added detector layers (neck and heads) are newly trained with the downstream detection dataset. Existing approaches often miss novel/unlabeled objects in the object proposal stage because the proposals tend to classify them as background. To remedy this, we leverage recent advances in a novel object proposal method and adopt the localization quality-based objectness (i.e., centerness score) instead of object-or-not binary classification score, which is combined with the detection score. During training, we compute the detection scores for each detected region as the cosine similarity between the region’s embedding (computed via RoI-Align operation) and the text embeddings of the base categories. At test time, we append the text embeddings of novel categories, and the detection score is now computed with the union of the base and novel categories.

The pre-trained ViT backbone is transferred to the downstream open-vocabulary detection by replacing the global average pooling with detector heads. The RoI-Align embeddings are matched with the cached category embeddings to obtain the VLM score, which is combined with the detection score into the open-vocabulary detection score.

Results

We evaluate RO-ViT on the LVIS open-vocabulary detection benchmark. At the system-level, our best model achieves 33.6 box average precision on rare categories (AP_r) and 32.1 mask AP_r, which outperforms the best existing ViT-based approach OWL-ViT by 8.0 AP_r and the best CNN-based approach ViLD-Ens by 5.8 mask AP_r. It also exceeds the performance of many other approaches based on knowledge distillation, pre-training, or joint training with weak supervision.

RO-ViT outperforms both the state-of-the-art (SOTA) ViT-based and CNN-based methods on LVIS open-vocabulary detection benchmark. We show mask AP on rare categories (APr) , except for SOTA ViT-based (OwL-ViT) where we show box AP.

Apart from evaluating region-level representation through open-vocabulary detection, we evaluate the image-level representation of RO-ViT in image-text retrieval through the MS-COCO and Flickr30K benchmarks. Our model with 303M ViT outperforms the state-of-the-art CoCa model with 1B ViT on MS COCO, and is on par on Flickr30K. This shows that our pre-training method not only improves the region-level representation but also the global image-level representation for retrieval.

We show zero-shot image-text retrieval on MS COCO and Flickr30K benchmarks, and compare with dual-encoder methods. We report recall@1 (top-1 recall) on image-to-text (I2T) and text-to-image (T2I) retrieval tasks. RO-ViT outperforms the state-of-the-art CoCa with the same backbone.

RO-ViT open-vocabulary detection on LVIS. We only show the novel categories for clarity. RO-ViT detects many novel categories that it has never seen during detection training: “fishbowl”, “sombrero”, “persimmon”, “gargoyle”.

Visualization of positional embeddings

We visualize and compare the learned positional embeddings of RO-ViT with the baseline. Each tile is the cosine similarity between positional embeddings of one patch and all other patches. For example, the tile in the top-left corner (marked in red) visualizes the similarity between the positional embedding of the location (row=1, column=1) and those positional embeddings of all other locations in 2D. The brightness of the patch indicates how close the learned positional embeddings of different locations are. RO-ViT forms more distinct clusters at different patch locations showing symmetrical global patterns around the center patch.

Each tile shows the cosine similarity between the positional embedding of the patch (at the indicated row-column position) and the positional embeddings of all other patches. ViT-B/16 backbone is used.

Conclusion

We present RO-ViT, a contrastive image-text pre-training framework to bridge the gap between image-level pre-training and open-vocabulary detection fine-tuning. Our methods are simple, scalable, and easy to apply to any contrastive backbones with minimal computation overhead and no increase in parameters. RO-ViT achieves the state-of-the-art on LVIS open-vocabulary detection benchmark and on the image-text retrieval benchmarks, showing the learned representation is not only beneficial at region-level but also highly effective at the image-level. We hope this study can help the research on open-vocabulary detection from the perspective of image-text pre-training which can benefit both region-level and image-level tasks.

Acknowledgements

Dahun Kim, Anelia Angelova, and Weicheng Kuo conducted this work and are now at Google DeepMind. We would like to thank our colleagues at Google Research for their advice and helpful discussions.

Posted by Susanna Ricco and Utsav Prabhu, co-leads, Perception Fairness Team, Google Research

Google’s Responsible AI research is built on a foundation of collaboration — between teams with diverse backgrounds and expertise, between researchers and product developers, and ultimately with the community at large. The Perception Fairness team drives progress by combining deep subject-matter expertise in both computer vision and machine learning (ML) fairness with direct connections to the researchers building the perception systems that power products across Google and beyond. Together, we are working to intentionally design our systems to be inclusive from the ground up, guided by Google’s AI Principles.

Perception Fairness research spans the design, development, and deployment of advanced multimodal models including the latest foundation and generative models powering Google’s products.

Our team’s mission is to advance the frontiers of fairness and inclusion in multimodal ML systems, especially related to foundation models and generative AI. This encompasses core technology components including classification, localization, captioning, retrieval, visual question answering, text-to-image or text-to-video generation, and generative image and video editing. We believe that fairness and inclusion can and should be top-line performance goals for these applications. Our research is focused on unlocking novel analyses and mitigations that enable us to proactively design for these objectives throughout the development cycle. We answer core questions, such as: How can we use ML to responsibly and faithfully model human perception of demographic, cultural, and social identities in order to promote fairness and inclusion? What kinds of system biases (e.g., underperforming on images of people with certain skin tones) can we measure and how can we use these metrics to design better algorithms? How can we build more inclusive algorithms and systems and react quickly when failures occur?

Measuring representation of people in media

ML systems that can edit, curate or create images or videos can affect anyone exposed to their outputs, shaping or reinforcing the beliefs of viewers around the world. Research to reduce representational harms, such as reinforcing stereotypes or denigrating or erasing groups of people, requires a deep understanding of both the content and the societal context. It hinges on how different observers perceive themselves, their communities, or how others are represented. There’s considerable debate in the field regarding which social categories should be studied with computational tools and how to do so responsibly. Our research focuses on working toward scalable solutions that are informed by sociology and social psychology, are aligned with human perception, embrace the subjective nature of the problem, and enable nuanced measurement and mitigation. One example is our research on differences in human perception and annotation of skin tone in images using the Monk Skin Tone scale.

Our tools are also used to study representation in large-scale content collections. Through our Media Understanding for Social Exploration (MUSE) project, we’ve partnered with academic researchers, nonprofit organizations, and major consumer brands to understand patterns in mainstream media and advertising content. We first published this work in 2017, with a co-authored study analyzing gender equity in Hollywood movies. Since then, we’ve increased the scale and depth of our analyses. In 2019, we released findings based on over 2.7 million YouTube advertisements. In the latest study, we examine representation across intersections of perceived gender presentation, perceived age, and skin tone in over twelve years of popular U.S. television shows. These studies provide insights for content creators and advertisers and further inform our own research.

An illustration (not actual data) of computational signals that can be analyzed at scale to reveal representational patterns in media collections. [Video Collection / Getty Images]

Moving forward, we’re expanding the ML fairness concepts on which we focus and the domains in which they are responsibly applied. Looking beyond photorealistic images of people, we are working to develop tools that model the representation of communities and cultures in illustrations, abstract depictions of humanoid characters, and even images with no people in them at all. Finally, we need to reason about not just who is depicted, but how they are portrayed — what narrative is communicated through the surrounding image content, the accompanying text, and the broader cultural context.

Analyzing bias properties of perceptual systems

Building advanced ML systems is complex, with multiple stakeholders informing various criteria that decide product behavior. Overall quality has historically been defined and measured using summary statistics (like overall accuracy) over a test dataset as a proxy for user experience. But not all users experience products in the same way.

Perception Fairness enables practical measurement of nuanced system behavior beyond summary statistics, and makes these metrics core to the system quality that directly informs product behaviors and launch decisions. This is often much harder than it seems. Distilling complex bias issues (e.g., disparities in performance across intersectional subgroups or instances of stereotype reinforcement) to a small number of metrics without losing important nuance is extremely challenging. Another challenge is balancing the interplay between fairness metrics and other product metrics (e.g., user satisfaction, accuracy, latency), which are often phrased as conflicting despite being compatible. It is common for researchers to describe their work as optimizing an “accuracy-fairness” tradeoff when in reality widespread user satisfaction is aligned with meeting fairness and inclusion objectives.

We built and released the MIAP dataset as part of Open Images, leveraging our research on perception of socially relevant concepts and detection of biased behavior in complex systems to create a resource that furthers ML fairness research in computer vision. Original photo credits — left: Boston Public Library; middle: jen robinson; right: Garin Fons; all used with permission under the CC- BY 2.0 license.

To these ends, our team focuses on two broad research directions. First, democratizing access to well-understood and widely-applicable fairness analysis tooling, engaging partner organizations in adopting them into product workflows, and informing leadership across the company in interpreting results. This work includes developing broad benchmarks, curating widely-useful high-quality test datasets and tooling centered around techniques such as sliced analysis and counterfactual testing — often building on the core representation signals work described earlier. Second, advancing novel approaches towards fairness analytics — including partnering with product efforts that may result in breakthrough findings or inform launch strategy.

Advancing AI responsibly

Our work does not stop with analyzing model behavior. Rather, we use this as a jumping-off point for identifying algorithmic improvements in collaboration with other researchers and engineers on product teams. Over the past year we’ve launched upgraded components that power Search and Memories features in Google Photos, leading to more consistent performance and drastically improving robustness through added layers that keep mistakes from cascading through the system. We are working on improving ranking algorithms in Google Images to diversify representation. We updated algorithms that may reinforce historical stereotypes, using additional signals responsibly, such that it’s more likely for everyone to see themselves reflected in Search results and find what they’re looking for.

This work naturally carries over to the world of generative AI, where models can create collections of images or videos seeded from image and text prompts and can answer questions about images and videos. We’re excited about the potential of these technologies to deliver new experiences to users and as tools to further our own research. To enable this, we’re collaborating across the research and responsible AI communities to develop guardrails that mitigate failure modes. We’re leveraging our tools for understanding representation to power scalable benchmarks that can be combined with human feedback, and investing in research from pre-training through deployment to steer the models to generate higher quality, more inclusive, and more controllable output. We want these models to inspire people, producing diverse outputs, translating concepts without relying on tropes or stereotypes, and providing consistent behaviors and responses across counterfactual variations of prompts.

Opportunities and ongoing work

Despite over a decade of focused work, the field of perception fairness technologies still seems like a nascent and fast-growing space, rife with opportunities for breakthrough techniques. We continue to see opportunities to contribute technical advances backed by interdisciplinary scholarship. The gap between what we can measure in images versus the underlying aspects of human identity and expression is large — closing this gap will require increasingly complex media analytics solutions. Data metrics that indicate true representation, situated in the appropriate context and heeding a diversity of viewpoints, remains an open challenge for us. Can we reach a point where we can reliably identify depictions of nuanced stereotypes, continually update them to reflect an ever-changing society, and discern situations in which they could be offensive? Algorithmic advances driven by human feedback point a promising path forward.

Recent focus on AI safety and ethics in the context of modern large model development has spurred new ways of thinking about measuring systemic biases. We are exploring multiple avenues to use these models — along with recent developments in concept-based explainability methods, causal inference methods, and cutting-edge UX research — to quantify and minimize undesired biased behaviors. We look forward to tackling the challenges ahead and developing technology that is built for everybody.

Acknowledgements

We would like to thank every member of the Perception Fairness team, and all of our collaborators.