The embattled U.S. Postal Service took the first, tentative steps toward embracing AI technology in 2019 with its decision to deploy an image analysis system using GPU-supported servers and deep learning software to train package-sorting algorithms.

Requirements for this implementation of robotic process automation are enormous: USPS processes 40 percent of the world’s mail, including 7.3 billion packages a year. Among the goals of the AI initiative dubbed Edge Compute Infrastructure Program (ECIP) was deploying edge AI servers in the postal service’s 195 mail-processing centers to scan, track and deliver the estimated 231 packages handled per second.

Cameras mounted on the sorting machines capture addresses, barcodes and other data such as hazardous materials symbols. (Source: U.S. Postal Service)

USPS contracted with Accenture Federal Services and Hewlett Packard Enterprise to provide ECIP infrastructure. Nvidia, a supplier for the AI project, touted the new machine vision application this week as an enterprise-scale AI deployment unique to government agencies. The GPU leader is providing its , each consisting of four V100 GPUs running on HPE servers used to train package-sorting algorithms.

Combined with Nvidia’s inference server that delivers models to USPS processing centers, the algorithms are used for image classification and object detection applications. For example, a model could be trained to spot damaged bar codes.

Those models would help accelerate steps like determining the size, weight and postage requirements for packages as well as tracking down lost parcels.

ECIP was originally scheduled to hit the ground running last spring. A USPS spokesperson said an operational ECIP system went online in late August 2020.

Nvidia estimates the sprawling USPS distribution network would have required a network of servers requiring 800 CPUs to implement a traditional computer vision platform — the equivalent of a data center. The implementation also would have been limited in its ability to access cloud resources.

“That’s absolutely a non-starter,” said Anthony Robbins, vice president of Nvidia’s federal business unit.

The USPS initiative instead uses Hewlett Packard 6500 servers, each equipped with four Nvidia V100 Tensor Core GPUs. The combination was used to train algorithms to perform equivalent scanning tasks in 20 minutes, Nvidia claims.

The operational ECIP platform currently processes about 20 terabytes of image data daily gathered by more than 1,000 processing machines.

“The overall design here is to continue to enhance and build a database for packages so that [USPS] can over time improve package processing and efficiency, and build from this model [for] the full range of mail processing, which is 129 billion piece of mail a year,” Robbins said.

In response to our query about ECIP error rates, the USPS spokesperson said: “The current usage of the system isn’t producing data that would have a constituted ‘error rate’.” Rather, the spokesperson added, “The main usage of the system is for troubleshooting purposes of anomalies in package sorting and processing.”

According to a by Nvidia, USPS released a requirement for an optical character recognition (OCR) capability for ECIP that would streamline its imaging workflow. The goal was to replace hardware and software infrastructure and inefficient cloud access with an AI-based machine vision system.

The new OCR workflow runs as a container-based deep learning model managed by the Kubernetes cluster orchestrator and served by Nvidia’s . Triton is designed to automate delivery of different AI models to systems that vary by GPU and CPU versions supporting deep learning frameworks.

“People in our organization are thinking of new ways to apply machine learning to new facets of robotics, data processing and image handling,” said Todd Schimmel, the USPS manager for ECIP.

Nvidia stressed the AI project’s scale. “There are not many enterprise-wide AI, [machine learning], computer vision projects that have been deployed at this scale, especially not in the case of government,” Robbins asserted.

AI-based deployments like ECIP have prompted fresh assessments of how best to oversee the scaling of automated systems. As AI algorithms are deployed in real-world applications, “It’s important that we devise a way in which we can ensure AI systems do not have runaway effects on decisioning systems,” said , head of KPMG’s Data Engineering and Innovation-Driven AI Transformation unit.

The management consultant touts an “AI parenting” framework serving as an extension of traditional “human-in-the-loop” pipelines. “While training processes have historically required a less-skilled human workforce, implementing a more concrete framework to incorporate cause and effect in the training process will help advance AI systems to the next level,” Krishna added.

The post appeared first on .