Machine Learning at Amazon: Will Amazon Go Reinvent Retail?

Imagine a future where there are no longer queues to pay for groceries at a supermarket, order a meal at a fast food restaurant, or buy a movie ticket at a theater. As dreamlike as it sounds, these scenarios are becoming realized with the rise of computer vision, an application of machine learning. Amazon, one of the major players in the machine learning movement, has its eyes on using this technology to reinvent retail, a $5 trillion industry, with its cashier-less convenience store, Amazon Go [1]. The following link illustrates the unique shopping experience it offers.

Layout of Typical Amazon Go Store [2]

The Bigger Play: What’s in it for Amazon?

Amazon has long been known to have a desire in the offline retail space and Go is the company’s first move to have a bigger offline presence. To do this, in-store cameras and computer vision algorithms are used to enable a cashier-less checkout process that maps each item taken off the shelf to the customer that took it, and the respective account to charge for the purchase. This “just walk out technology” has allowed Amazon to offer a shopping experience that is more convenient and novel, thus gaining market share over its competitors [3].

From the company’s perspective, Amazon Go allows the company to develop a strong computer vision algorithm that can be shared in other core businesses. As more customers shop at a Go store, Amazon can collect more data to build even more robust object identification and people tracking algorithms to be leveraged in and tangential businesses.

The Blueprint to Disrupting the Retail Industry

In the short-term, Amazon is planning to scale Go stores quickly, targeting to have 3000 stores in the US by 2021 [4]. This rapid expansion allows Amazon to collect large amounts of training data for its algorithm. Compared to data from simulations or other sources, data from the actual in-store usage would be of the highest quality data from a machine learning standpoint.

Further, while not publicly disclosed, Amazon most likely has a feedback system that allows it to retrain its algorithm. After a period of deployment, if there is a mismatch between actual inventory and the inventory implied by the algorithm, then Amazon can identify the most commonly misrecognized items and have a human intervene to label these items, retraining the algorithm to improve accuracy [5].

In the medium-term, Amazon is working to overcome challenges that would further its lead from competitors. One of them is the product selection offered at the Go store. The initial set of products may have been selected for their ability to be recognized by the computer vision algorithm, which may be a factor of the item’s size and packaging uniqueness. As the algorithm becomes more accurate and able to identify items based off smaller, more nuanced features, Amazon can offer a wider selection of products that don’t depend easily recognizable packaging.

Potential Threats

Amazon is an intimidating player in the AI space because of the size of the company and its available resources. However, even with its current Go development plan, there exists two weaknesses that Amazon should work to address.

First, the just walk out technology, unlike popular belief, is not entirely based on cameras and computer vision algorithms. The technology relies on weight sensors in all the shelves, which probably help to reduce the computing power of constantly processing so many video streams of cameras by serving as triggers [6]. The weight sensors may also improve recognition accuracy by referencing the weight of an item removed from the shelf.

The implication of this reliance however, is that Amazon’s current technology can only be applied to brand new stores that are purpose-built. Old stores that can be retrofitted with this technology represent most of the retail market, but this is a segment that Amazon is giving up. Even if Amazon were not planning on selling the technology to other stores, the homogeneous nature of the data they collect and train their algorithm with will result in a less generalizable algorithm with fewer applications.

Finally, privacy implications remain a significant open question that may impact Amazon Go. While Amazon claims that it does not employ facial recognition to track people in the store, there could be widespread concern on the traceability of offline customer behavior to their online identities [7]. It would therefore be up to the consumers, regulators, and politicians to shape a future that will see more cameras and algorithms in all aspects of life.

If Amazon can overcome these challenges, then Amazon Go will surely be able to shape the future of retail.

(765 words)



[1] “Quarterly Retail E-Commerce Sales Quarter 2018 –”, U.S. Census Bureau News, 17 Aug. 2018,

[2] Tibken, Shara. You Have to Scan the Amazon Go App on Your Phone to Enter the Store. 19 Sept. 2018,×0/2018/01/19/47cdb5b1-cc88-4935-ad04-f42b14797c6c/amazon-go-badge.jpg

[3] Statt, Nick. “Why Amazon’s Future Depends on Moving from the Internet to the Physical World.” The Verge, The Verge, 2 Nov. 2018,

[4] Soper, Spencer. “Amazon Will Consider Opening Up to 3,000 Cashierless Stores by 2021.”, Bloomberg, 19 Sept. 2018,

[5] Rey, Jason Del. “Amazon’s Store of the Future Has No Cashiers, but Humans Are Watching from behind the Scenes.” Recode, Recode, 6 Jan. 2017,

[6] Coldewey, Devin. “Inside Amazon’s Surveillance-Powered, No-Checkout Convenience Store.” TechCrunch, TechCrunch, 21 Jan. 2018,

[7] Harwell, Drew, and Abha Bhattarai. “Inside Amazon Go: The Camera-Filled Convenience Store That Watches You Back.” The Washington Post, WP Company, 22 Jan. 2018,


Machine Learning Journey of Burberry: Is There Room For Machine Learning in Luxury Goods Industry?


Disney – “A Whole New World” of Machine Learning

Student comments on Machine Learning at Amazon: Will Amazon Go Reinvent Retail?

  1. Such an interesting topic. The amount of data Amazon will be able to collect is going to be massively valuable for both Amazon and retail brands – how customers compare competitive brands, lingering time at items, customer facial reactions to different packaging, real-time inventory levels, shopping patterns of individuals etc. But how Amazon can unleash the potential of those data without infringing privacy will be interesting to see.

  2. Great piece. I expect that inventory management and supply chains will see huge benefits from this model as well, given that customer demand can be sent up the chain instantaneously. To that same end, I see a lot of value for brands who could view live data of customer purchasing behaviors and use that information in their product development. Assuredly, Amazon would charge for these types of subscriptions and increase the earning potential of these stores.

    Privacy is the standout concern in my mind. Given the relative distrust of the government and its handling of domestic espionage, having facial recognition scanners in grocery stores that track everything individual customers will certainly face some public skepticism.

  3. Great article. I personally love amazon, and think it’s the way of the future. However, I wonder if they are unnecessarily investing in brick-and-mortar space. Are they spending money on a problem that people don’t want fixed? I wonder how much time people will be spending in stores once amazon truly takes over online retail. Are they using AI to fix problems in stores that people will no longer visit?

Leave a comment