Data scientists choose models based on various tradeoffs when solving machine learning (ML) problems that involve tabular (i.e., structured) data, the most common data type within enterprises. Among such models, decision trees are popular because they are easy to interpret, fast to train, and can obtain high accuracy quickly from small-scale datasets. On the other hand, deep neural networks offer superior accuracy on larger datasets, as well as the benefits of end-to-end learning, but are black-box and difficult to interpret. TabNet, an interpretable deep learning architecture developed by Google AI, combines the best of both worlds: it is explainable, like simpler tree-based models, and can achieve the high accuracy of complex black-box models and ensembles.

We’re excited to announce that TabNet is now available in Vertex AI Tabular Workflows! Tabular Workflows provides fully managed, optimized, and scalable pipelines, making it easier to use TabNet without worrying about implementation details, and to deploy TabNet with the MLOps capabilities of Vertex. TabNet on Vertex AI Tabular Workflows is optimized for efficient scaling to massive tabular datasets. Moreover, TabNet on Vertex AI Tabular Workflows come with machine learning improvements on top of the original TabNet, yielding better accuracy for real-world data challenges.

TabNet on Vertex AI is well-suited for a wide range of tabular data tasks where model explainability is just as important as accuracy, such as financial asset price prediction, fraud/cyberattack/crime detection, retail demand forecasting, user modeling, credit/risk scoring, diagnosis from healthcare records, and product recommendations.

Overview of Tabnet

TabNet has a specially-designed architecture (overviewed in Fig. 1), based on sequential attention, that selects which model features to reason from at each step. This mechanism makes it possible to explain how the model arrives at its predictions and the thoughtful design helps with superior accuracy. TabNet not only outperforms alternative models (including neural networks and decision trees) but also provides interpretable feature attributions. More details, including results on academic benchmarks, can be found in our AAAI 2021 paper.

Since its publication, TabNet has received significant traction from various enterprises across different industries and a variety of high-value tabular data applications (most of which include the ones for which deep learning was not even used a priori). It has been used by numerous enterprises like Microsoft, Ludwig, Ravelin, and Determined. Given high customer interest on TabNet, we’ve worked on making it available on Vertex given the real-world deep learning development and productionization needs, as well as improving its performance and efficiency.

Highlights of TabNet on Vertex AI Tabular Workflows

Scaling to Very Large Datasets
Fueled by the advances in cloud technologies like BigQuery, enterprises are increasingly collecting more tabular data, and datasets with billions of samples and hundreds/thousands of features are becoming the norm. In general, deep learning models get better learning from more data samples, and more features, with the optimal methods as they can better learn the complex patterns that drive the predictions. The computational challenges become significant though when model development on massive datasets is considered. This results in high cost or very long model development times, constituting a bottleneck for most customers to fully take advantage of their large datasets. With TabNet on Tabular Workflows, we’re making it more efficient to scale to very large tabular datasets.

Key Implementation Aspects: The TabNet architecture has unique advantages for scaling: it is composed mainly of tensor algebra operations, it utilizes very large batch sizes, and it has high compute intensity (i.e., the architecture employs a high number of operations for each data byte transmitted). These open a path to efficient distributed training on many GPUs, utilized to scale TabNet training in our improved implementation.

In TabNet on Vertex AI Tabular Workflows, we have carefully engineered the data and training pipelines to maximize hardware utilization so that users can get the best return for their Vertex AI spending. The following features enable scale with TabNet on Tabular workflows:

• Parallel data reading with multiple CPUs in a pipeline optimized to maximize GPU utilization for distributed training, reflecting best practices from Tensorflow.
• Training on multiple GPUs that can provide significant speedups on large datasets with high compute requirements. Users can specify any available machine on GCP with multiple GPUs, and the model will automatically run on them with distributed training.
• For efficient data parallelism with distributed learning, we use Tensorflow mirrored distribution strategy to support data parallelism across many GPUs. Our results demonstrate >80% utilization with several GPUs on billion-scale datasets with 100s-1000s of features.

Standard implementations of deep learning models could yield a low GPU utilization, and thus inefficient use of resources. With our implementation, TabNet on Vertex, users can get the maximal return on their compute spend on large-scale datasets.

Examples on real-world customer data: We have benchmarked the training time specifically for enterprise use cases where large datasets are being used and fast training is crucial. In one representative example, we used 1 NVIDIA_TESLA_V100 GPU to achieve state-of-the-art performance in ~1 hour on a dataset with ~5 million samples. In another example, we used 4 NVIDIA_TESLA_V100 GPUs to achieve state-of-the-art performance in ~14 hours on a dataset with ~1.4 billion samples.

Improving Accuracy given Real-World Data Challenges

Compared to its original version, TabNet on Vertex AI Tabular Workflows has improved machine learning capabilities. We have specifically focused on the common real-world tabular data challenges. One common challenge for real-world tabular data is numerical columns having skewed distributions, for which we productionized learnable preprocessing layers (e.g. including parametrized power transform families and quantile transformations) that improve the TabNet learning. Another common challenge is the high number of categories for categorical data, for which we adopted tunable high-dimensional embeddings. Another one is imbalance of label distribution, for which we added various loss function families (e.g. focal loss and differentiable AUC variants). We have observed that such additions can provide a noticeable performance boost in some cases.

Case studies with real-world customer data: We have worked with large customers to replace legacy algorithms with TabNet for a wide range of use cases, including recommendation, rankings, fraud detection, and estimated arrival time predictions. In one representative example, TabNet was stacked against a sophisticated model ensemble for a large customer. It outperformed the ensemble in most cases, leading to a nearly 10% error reduction on some of the key tasks. This is an impressive result, given that each percentage improvement on this model resulted in multi-million savings for the customer!

Out-of-the-box Explainability

In addition to high accuracy, another core benefit of TabNet is that, unlike conventional deep neural network (DNN) models such as multi-layer perceptrons, its architecture includes explainability out of the box. This new launch on Vertex Tabular Workflows makes it very convenient to visualize explanations of the trained TabNet models, so that the users can quickly gain insights on how the TabNet models arrive at its decisions. TabNet provides feature importance output via its learned masks, which indicate whether a feature is selected at a given decision step in the model. Below is the visualization of the local and global feature importance based on the mask values. The higher the value of the mask for a particular sample, the more important the corresponding feature is for that sample. Explainability of TabNet has fundamental benefits over post-hoc methods like Shapley values that are computationally-expensive to estimate, while TabNet’s explanations are readily available from the model’s intermediate layers. Furthermore, post-hoc explanations are based on approximations to nonlinear black-box functions while TabNet’s explanations are based on what the actual decision making is based on.

Explainability example: To illustrate what is achievable with this kind of explainability, Figure 2 below shows the feature importance for the Census dataset. The figure indicates that education, occupation, and number of hours per week are the most important features to predict whether a person can earn more than $50K/year (the color of corresponding columns are lighter). The explainability capability is sample-wise, which means that we can get the feature importance for each sample separately.

Benefits as a Fully-Managed Vertex Pipeline

TabNet on Vertex Tabular Workflows makes the model development and deployments tasks much simpler – without writing any code, one can obtain the trained TabNet model, deploy it in their application, and use the MLOps capabilities enabled by Vertex Managed Pipelines! Some of these benefits are highlighted as:

• Compatibility with Vertex AI ML Ops for implementing automated ML at scale including products like Vertex AI Pipelines and Vertex AI Experiments.
• Deployment convenience: Vertex AI prediction services, both in batch and online mode, are supported out-of-the-box.
• Customizable feature engineering to enable the best utilization of the domain knowledge of users.
• Using Google’s state-of-the-art search algorithms, automatic tuning to identify the best-performing hyperparameters, with automatic selection of the appropriate hyper-parameter search space based on dataset size, prediction type, and training budget.
• Tracking the deployed model and convenient evaluation tools.
• Easiness in comparative benchmarking with other models (such as AutoML and Wide & Deep Networks) as the user journey would be unified.
• Multi-region availability to better address international workloads.

More details

If you’re interested in trying TabNet on Vertex AI on your tabular datasets, please check out Tabular Workflow on Vertex AI and fill out this form.

Acknowledgements: We’d like to thank Nate Yoder, Yihe Dong, Dawei Jia, Alex Martin, Helin Wang, Henry Tappen and Tomas Pfister for their contributions to this blog.

Retail will forever be an industry that must constantly reinvent itself in response to, and anticipation of, ever-changing consumer demands.

Digital transformation is fueling these changes and we’ve previously spoken about how businesses including Ulta Beauty and Kohl’s are taking advantage of Google Cloud to put data at the center of what they do and deliver the best possible shopping experience and product offerings for their customers.

Leveraging Google Cloud machine learning platform, Sainsbury is able to develop predictive analytics models to spot trends and adjust inventory, providing shoppers with a better experience. 

Sainsbury’s, one of Britain’s best-known supermarkets, is another great example of a business transforming the way it engages with its customers with the cloud.

With over 150 years of service, Sainsbury’s vision is to be the most trusted retailer, where people love to work and shop. It makes customers’ lives easier, by offering great quality and service at fair prices. 

The food industry and the way that customers shop is rapidly changing. From foodie hashtags on Instagram, to the latest cooking fads, customers want to stay connected to the latest trends and Sainsbury’s is empowering them do that.

To help Sainsbury’s achieve this goal, its Commercial and Technology teams, in partnership with Accenture, are building cutting-edge machine learning solutions on Google Cloud Platform (GCP) to provide new insights on what customers want and the trends driving their eating habits.

With the help of Google Cloud Platform, we are generating new insights into how the world eats and lives, to help us stay ahead of market trends and provide an even better shopping experience for our customers.
–Phil Jordan, Group CIO, Sainsbury’s 

Sainsbury’s solution relies on data from multiple structured and unstructured sources. Using Google Cloud’s powerful cloud-based analytics tools to ingest, clean and classify that data, and a custom-built front-end interface for internal users to seamlessly navigate through a variety of filters and categories, Sainsbury’s is able to gain advanced insights in real time.

As a result, Sainsbury’s has been able to develop predictive analytics models to spot trends and adjust inventory, providing shoppers with a better experience. 

Phil Jordan, Group CIO of Sainsbury’s believes this project will have a big impact.

“The grocery market continues to change rapidly. We know our customers want high quality at great value and that finding innovative and distinctive products is increasingly important to them. With the help of Google Cloud Platform, we are generating new insights into how the world eats and lives, to help us stay ahead of market trends and provide an even better shopping experience for our customers.” 

This project is also a great example of the successes Google Cloud customers have when they work with the company’s partners.

“We’re delighted to partner with Google Cloud to help the Sainsbury’s Commercial team apply predictive analytics to the identification of new and emerging trends in grocery,” says Adrian Bertschinger, Managing Director for Retail, Accenture.

“The food sector is experiencing significant, rapid disruption, and this new, cloud-based insights platform will help Sainsbury’s identify trends much earlier and adapt their product assortment in a faster, more informed way—all for the benefit of customers.” 

Whatever the next food or shopping trend may be, Sainsbury’s is looking to the cloud to help them stay a step ahead. 

The latest round of MLPerf benchmark results have been released, and Google’s TPU v4 supercomputers demonstrated record-breaking performance at scale. This is a timely milestone since large-scale machine learning training has enabled many of the recent breakthroughs in AI, with the latest models encompassing billions or even trillions of parameters (T5, Meena, GShard, Switch Transformer, and GPT-3). 

Google’s TPU v4 Pod was designed, in part, to meet these expansive training needs, and TPU v4 Pods set performance records in four of the six MLPerf benchmarks Google entered using TensorFlow and JAX. These scores are a significant improvement over our winning submission from last year and demonstrate that Google once again has the world’s fastest machine learning supercomputers. These TPU v4 Pods are already widely deployed throughout Google data centers for our internal machine learning workloads and will be available via Google Cloud later this year.

Figure 1: Speedup of Google’s best MLPerf Training v1.0 TPU v4 submission over the fastest non-Google submission in any availability category – in this case, all baseline submissions came from NVIDIA. Comparisons are normalized by overall training time regardless of system size. Taller bars are better.¹

Let’s take a closer look at some of the innovations that delivered these ground-breaking results and what this means for large model training at Google and beyond. 

Google’s continued performance leadership

Google’s submissions for the most recent MLPerf demonstrated leading top-line performance (fastest time to reach target quality), setting new performance records in four benchmarks. We achieved this by scaling up to 3,456 of our next-gen TPU v4 ASICs with hundreds of CPU hosts for the multiple benchmarks. We achieved an average of 1.7x improvement in our top-line submissions compared to last year’s results. This means we can now train some of the most common machine learning models in a matter of seconds.

Figure 2: Speedup of Google’s MLPerf Training v1.0 TPU v4 submission over Google’s MLPerf Training v0.7 TPU v3 submission (exception: DLRM results in MLPerf v0.7 were obtained using TPU v4). Comparisons are normalized by overall training time regardless of system size. Taller bars are better. Unet3D not shown since it is a new benchmark for MLPerf v1.0.²

We achieved these performance improvements through continued investment in both our hardware and software stacks. Part of the speedup comes from using Google’s fourth-generation TPU ASIC, which offers a significant boost in raw processing power over the previous generation, TPU v3. 4,096 of these TPU v4 chips are networked together to create a TPU v4 Pod, with each pod delivering 1.1 exaflop/s of peak performance.

Figure 3: A visual representation of 1 exaflop/s of computing power. If 10 million laptops were running simultaneously, then all that computing power would almost match the computing power of 1 exaflop/s.

In parallel, we introduced a number of new features into the XLA compiler to improve the performance of any ML model running on TPU v4. One of these features provides the ability to operate two (or potentially more) TPU cores as a single logical device using a shared uniform memory access system. This memory space unification allows the cores to easily share input and output data – allowing for a more performant allocation of work across cores. A second feature improves performance through a fine-grained overlap of compute and communication. Finally, we introduced a technique to automatically transform convolution operations such that space dimensions are converted into additional batch dimensions. This technique improves performance at the low batch sizes that are common at very large scales.

Enabling large model research using carbon-free energy

Though the margin of difference in topline MLPerf benchmarks can be measured in mere seconds, this can translate to many days worth of training time on the state-of-the-art models that comprise billions or trillions of parameters. To give an example, today we can train a 4 trillion parameter dense Transformer with GSPMD on 2048 TPU cores. For context, this is over 20 times larger than the GPT-3 model published by OpenAI last year. We are already using TPU v4 Pods extensively within Google to develop research breakthroughs such as MUM and LaMDA, and improve our core products such as Search, Assistant and Translate. The faster training times from TPUs result in efficiency savings and improved research and development velocity. Many of these TPU v4 Pods will be operating at or near 90% carbon free energy. Furthermore, cloud datacenters can be ~1.4-2X more energy efficient than typical datacenters, and the ML-oriented accelerators – like TPUs – running inside them can be ~2-5X more effective than off-the-shelf systems. 

We are also excited to soon offer TPU v4 Pods on Google Cloud, making the world’s fastest machine learning training supercomputers available to customers around the world. Cloud TPUs support leading frameworks such as TensorFlow, PyTorch, and Jax, and we recently released an all-new Cloud TPU system architecture that provides direct access to TPU host machines, greatly improving the user experience. 

Want to learn more? 

Please contact your Google Cloud sales representative to request early access to Cloud TPU v4 Pods. We are excited to see how you will expand the machine learning frontier with access to exaflops of TPU computing power!

^{1. All results retrieved from www.mlperf.org on June 30, 2021. MLPerf name and logo are trademarks. See www.mlperf.org for more information. Chart uses results 1.0-1067, 1.0-1070, 1.0-1071, 1.0-1072, 1.0-1073, 1.0-1074, 1.0-1075, 1.0-1076, 1.0-1077, 1.0-1088, 1.0-1089, 1.0-1090, 1.0-1091, 1.0-1092.
2. All results retrieved from www.mlperf.org on June 30, 2021. MLPerf name and logo are trademarks. See www.mlperf.org for more information. Chart uses results 0.7-65, 0.7-66, 0.7-67, 1.0-1088, 1.0-1090, 1.0-1091, 1.0-1092.}