Special Delivery: Machine Learning Tips from Amazon Web Services – Machine Design

  • Lauren
  • June 29, 2020
  • Comments Off on Special Delivery: Machine Learning Tips from Amazon Web Services – Machine Design

Every year, during Amazon’s annual planning process, leaders in every business unit are asked a pointed question: How do you plan to leverage machine learning in your business? The words “we don’t plan to” is not an acceptable answer, said Swami Sivasubramaniam, vice president of Amazon AI at Amazon Web Services. Speaking from a virtual stage at the Collision from Home conference, Sivasubramaniam told the audience that the world has already entered the golden age of artificial intelligence and machine learning. The online conference was orchestrated by the producers of the world’s largest tech conference, Web Summit, and boasted more than 30,000 attendees (June 23-25). It was arguably the ideal platform for a cloud-based, live meet-up to discuss long-term trends in the digital world and for Sivasubramaniam’s talk, “No hype: Deploying real-world machine learning.”Swami Sivasubramaniam, vice president of Amazon AI at Amazon Web Services counts down four pitfalls associated with machine learning.Machine Design ML has seen traction across industries and supply chains, and there are no shortages of examples of CIOs and CEOs professing how AI and ML are transforming business for the better, said Sivasubramaniam. His list of illustrations included the financial sector, where Intuit uses machine learning to forecast their contact center volume as a way to staff up; the medical field, where Aidoc leverages AI and computer vision to build systems that assist radiologists with image scans; and the public sector, where agencies such as NASA use ML algorithms to explore extreme conditions associated with superstorms. (NASA partners with AWS to detect solar flares based on signal anomalies that occur in space.) Amazon’s own journey goes back more than 20 years, when the digital conglomerate (today worth $1.3 trillion) started using machine learning technology for its supply chain, fulfillment centers and last-mile delivery. Amazon’s Echo voice-controlled smart speaker device is just a recent example of a product with development roots inside Amazon Lab126 dating back 10 years. Back then, Jeff Bezos and his leadership team began to realize that machine learning was about to go through a pivotal moment. “With the advent of new technologies, such as deep learning on the horizon, they started realizing that every line of business is going to need to have a machine learning strategy,” said Sivasubramaniam. Swami Sivasubramaniam, vice president of Amazon AI at Amazon Web Services.AmazonSince then, explained Sivasubramaniam, every line of business at Amazon—irrespective of whether they are running technology, research, human resources, finance or supply chain—is asked to consider how it can enhance the customer experience using machine learning in a meaningful way. Along the way Amazon learned a few lessons about applying ML successfully, all of which Sivasubramaniam demarcates as four pitfalls. Get data in order. Understand where to apply machine learning. Address the skills gap. Don’t do the undifferentiated heavy lifting. Four Machine Learning Pitfalls At Amazon, Sivasubramaniam’s professional repertoire extends to bootstrapping the NoSQL database ecosystem, as well as the AWS stack: ML algorithms (deep learning frameworks and ML algorithms); ML platform services; and AI application services such as Lex (rich conversational experiences), Polly (text to speech service) and Rekognition (image processing service). Culled from his online presentation and edited for clarification, the following tactical pointers can be applied to avoid obstacles. 1. Get Data in Order When data scientists are asked to name the biggest impediment when it comes to machine learning, they say “data,” asserts Sivasubramaniam. “More than 50% of data scientists spend their time in data wrangling, annotation, ETL and so forth,” he said, noting that the way to avoid this and accelerate machine learning, is to ask three questions: What data is available today? How can it be made easily available so that you can get started? And in a year’s time, what data will we wish we had so that we can start collecting today, and so that we continue to build a durable advantage for years to come? 2. Understand Where to Apply Machine Learning Picking the right business problem is important, said Sivasubramaniam, who compartmentalizes them along three dimensions: data readiness, business impact and machine learning applicability. “Machine learning algorithms and research have come a long way in solving the problems,” he said. “If you pick a problem where the data is not ready and machine learning research hasn’t been developed enough to solve this problem, but it is high business impact, you can throw a lot of resources at it. “But if you force a deadline, it’s going to lead to frustrated data scientists. On the other hand, if you pick a low-business-impact problem but high data and machine learning applicability, it could be a good prototype to build experience. Ideally, what you want is a problem that scores high on these three dimensions because it’s a great place to start.” Sivasubramaniam advised against building a group of technical experts in machine learning and placing them in a separate team without any contact with domain experts. “What typically happens is that technical experts tend to build interesting proof-of-concepts with no take off from business,” said Sivasubramaniam, adding that the ideal team comprised of domain experts and technical experts will “work backwards from the customer and build something meaningful.” 3. Address the Skills Gap There are not enough people who know machine learning, said Sivasubramaniam. He pointed to World Economic Forum data, which shows that jobs such as artificial intelligence and machine learning specialists or data scientists are forecasted to be among the most in-demand roles across most industries by 2022. Amazon started addressing this demand through its Machine Learning University about six years ago when it started training engineers and product managers, said Sivasubramaniam. 4. Don’t do the Undifferentiated Heavy Lifting The final pitfall, according to Sivasubramaniam, is that organizations become excited about solving undifferentiated heavy lifting and, by extension, fall prey to the idea of building a machine learning platform, a translation engine or a contributors’ engine. In a 2006 speech, “We Build Muck, So You Don’t Have To,” Jeff Bezos defined “undifferentiated heavy lifting” as server hosting, bandwidth management, contract negotiation, scaling and managing physical growth, as well as dealing with the accumulated complexity of heterogeneous hardware and co-ordinating large teams to manage each of these areas. Sivasubramaniam offers this advice: “What you ideally want is your engineers to focus on things that matter to the business and leverage things from clouds, such as AWS, or from open-source technologies, and solve the purely differentiated business problem.” Avoiding these pitfalls will set enterprises up for a future of machine learning where “businesses move from being reactive to proactive, to automate their processes from manual to automated processing, and from generalized customer experience to personalized experiences, and to taking technology from being obscure to being accessible,” he said. A tricky problem for manufacturers interested in benefiting from artificial intelligence is knowing where to start the adoption and how to scale up deployment. For vision AI software company Neurala, Inc., the launching point for manufacturers may very well be at the site of the product line. The Boston-based company recently introduced its VIA (vision inspection automation) software, which enables manufacturers that have not worked with AI before to train and use vision AI to identify defects in products or packaging on the production line.Heather Ames Versace, co-founder and COO, Neurala.NeuralaThe anomaly recognition technology identifies products that deviate from “acceptable” images without collecting images of defective parts. The system builds on Neurala’s Brain Builder platform that allows quality control managers and vision specialists to deploy vision AI for inspection and allows manufacturing plants to integrate legacy hardware on the floor. “We’re really trying to create a scalable software product that allows you to do those functions remotely,” said Heather Ames Versace, Neurala’s co-founder and COO. “Part of our DNA is that, whatever we develop as a software solution, it doesn’t have a dependency on making massive capital purchases, as well as making it scalable.” Hardware AgnosticOne way in which Neurala substantiates this claim is by ensuring customers can set up their systems with off-the-shelf components. The VIA software appeals to manufacturers reluctant to rely on connectivity to the cloud, Ames Versace explained, as it allows a particular facility to keep data on the factory site without the privacy or lag time concerns typically associated with cloud deployments.Neurala’s VIA software automates quality inspection processes that were previously not viable—improving inspection rates, decreasing human intervention and allowing smaller batches to be inspected.NeuralaInspecting Baked GoodsTake as an example a commercial bakery where inspections are typically done by hand, and where one in 100 trays might be pulled off the line for inspection. The challenges are that these inspections require a full-time person and consistency must be maintained when a different employee is assigned to the same task. But, in keeping with the times to have fewer people on the floor as a way to meet hygiene guidelines, Neurala’s inspections technology can be performed without human intervention while increasing uptime. “When baked goods come out of the oven on the conveyor, they need to be inspected for such criteria as color, shape and whether ingredients or toppings were distributed evenly,” said Ames Versace, adding that while inspecting baked goods can be a subjective assessment, the process is integral to maintaining brand equity. Neurala’s automated quality inspection software catches defects early and has the ability to train and operate multiple AI models. The VIA software is compatible with GigE cameras that can view baking trays before baking and after exiting the oven. In a recent use case, a bakery that used Ethernet/IP communication required a gateway to transfer the Modbus TCP communication to Ethernet/IP to talk to the PLCs. This solution is adaptive and allows the operator to refine decision-making. “It’s a constant learning situation where both subjective measurements are now supplemented with an AI tool,” said Ames Versace. The bakery’s resulting productivity improvements included higher throughput, lower waste costs and fewer staff, although bottom-line results differ by industry.In another example, a manufacturer of plastic molding parts for the electronics assembly had minimal knowledge of machine vision but needed to rapidly increase quality inspection protocols. Installing a GigE camera on the assembly line to inspect the parts after deburring allowed an “inspector”—an interface between a Modbus TCP and OPC UA to the PLC—to signal defects. A human operator would then inspect the defective part and determine whether it should be scrapped or reworked. According to Neurala, the benefit of using AI to increase uniformity and anomaly recognition, in addition to human inspection, is that it has the potential to ramp up production of products that have stronger quality control requirements, as well as providing the impetus for a higher mix of products.
Source: https://www.machinedesign.com/automation-iiot/article/21135436/special-delivery-machine-learning-tips-from-amazon-web-services