https://ift.tt/3CdvuJd The unexpected results of an AI that identifies bread Recently, I was asked the question: “Why did you decide to le...
The unexpected results of an AI that identifies bread
Recently, I was asked the question: “Why did you decide to learn Data Science?” The first things that came to mind were events like teaching myself Python during the downtime of my old security job, writing Python scripts to input and display company traveler data on a flight-map, or just how much fun I had making reports on which doors were used most frequently at my work building. However, the more I thought about it, the more I realized that these were contributing factors to my decision to study Data Science, not the primary reason. The primary reason comes from a story that has stuck with me since I first heard it.
Brain Co. Ltd.
In 2007 Japanese tech company Brain Co. Ltd. had seen moderate success selling software to big companies. For example, TEX-SIM is a user interface tool they developed for designing textile images. They also helped create software for rendering kanji characters on personal computers, and even made software for designing bridges. These talented software engineers were approached by a local bakery with a problem they were hoping to solve, and little did Brain Co. know, but their solution to this problem would come to define them, and would have a much larger impact then they ever planned.
Japan’s Pastry Problem
Japanese bakeries pride themselves on having a wide variety of freshly baked pastries. Some bakeries make hundreds of different varieties of pastries each morning. Analytical studies done at bakeries found that the total amount of pastries sold directly correlated with the total different varieties of pastries sold. A bakery that offered 100 varieties of pastries would sell twice as much as a bakery offering 30 varieties. The study also found that customers preferred pastries without wrappings, and sitting in open baskets as pictured below.
As bakeries implemented these changes based on the analytical study, a new, unforeseen problem occurred. Since the pastries needed to have no wrapping at all, there was no way to place a barcode on them. Customers would bring the pastry to the checkout clerk, who would then have to identify the pastry, and correctly remember its price. Some of these bakeries had hundreds of unique looking pastries with different prices, so this began to put a noticeable strain on the checkout clerks. Training new employees meant teaching them how to correctly identify all of the pastries and remember their prices. Bakery owners began to notice it would take weeks to train new clerks, and even the properly trained clerks still took a noticeable amount of time correctly identifying a customer’s items. So a local bakery came to Brain Co. and asked them if there was a better solution.
BakeryScan and the Trials of Brain Co.
Brain Co. had an incredible task before them. The team didn’t have access to the amount of data that would be required for a neural network so they had to come up with another solution. On top of this difficult problem, the financial crisis of 2008 occurred one year into the project. Brain Co. found that all their other business opportunities had dried up, and the survival of the company suddenly hinged on the success of this project. The Brain Co. team put everything they had into this project, which would become BakeryScan AI. James Somers of the New Yorker describes how the team developed the solution:
The state of the art in computer vision involved piecing together a pipeline of algorithms, each charged with a specific task. Suppose that you wanted to build a pedestrian-recognition system. You’d start with an algorithm that massaged the brightness and colors in your image, so that you weren’t stymied by someone’s red shirt. Next, you might add algorithms that identified regions of interest, perhaps by noticing the zebra pattern of a crosswalk. Only then could you begin analyzing image “features” — patterns of gradients and contrasts that could help you pick out the distinctive curve of someone’s shoulders, or the “A” made by a torso and legs. At each stage, you could choose from dozens if not hundreds of algorithms, and ways of combining them.
Using this “algorithm pipeline” led to many issues that were all chronicled in the team’s documentation of the development. Issues were abundant — for example, they found that the lighting in a room would significantly impact BakeryScan’s accuracy. They found that BakeryScan had trouble when two pastries were close together, and had to develop a way for it to discern when this was occurring. The same scenario occurred when the pastry was ripped or damaged. Sometimes they’d encounter problems such as the shadow a donut cast into its donut hole preventing BakeryScan from correctly identifying the item. They continuously tackled these issues with what the CEO, Hisashi Kambe, described as a “maniacal” focus. At one point, the team developed 10 prototypes over the course of two years. Finally, in 2013:
Brain Co.’s 5 year long struggle paid off. They had developed an AI that could identify over 50 different kinds of pastries with incredible accuracy. BakeryScan identifies the pastry in view of the camera, and then displays its name and price for the clerk, only taking a few seconds. BakeryScan was a hit, selling for $20,000 USD per unit, and it is now used in over 400 stores across Japan.
BakeryScan’s Legacy
This story on its own is inspirational; a team of software developers with their livelihoods on the line putting all they have into a seemingly impossible task, and coming out successful. Yet, the story doesn’t end here.
In 2017, a Doctor working for Kyoto’s Louis Pasteur Center for Medical Research saw an ad for BakeryScan. After some thinking he came to a quite funny realization. He thought that the bakery items being scanned by the AI had a remarkable similarity to cancer cells that he had been studying. So, he decided to reach out to Brain Co. and see if they would be interested in testing to see if BakeryScan may be able to detect cancerous cells under a microscope.
Cyto-AiSCAN
BakeryScan was able to identify a cancerous cell under a microscope with 98% accuracy. At first, it was only able to look at one cell at a time. After some modifications and testing, the AI was able to look at an entire slide of cells under a microscope and highlight which cells were cancerous, again, with a 98% accuracy. When treating cancer, the survival rate greatly increases the earlier the cancer is detected. Using the AI dramatically increased the rate that lab technicians could sort through samples and detect the presence of cancerous cells. Brain Co. gave the modified BakerScan a new name: “Cyto-AiSCAN” and it is currently being used and trained by doctors at two major hospitals in Kobe and Kyoto, Japan.
When I heard this story I immediately thought that they must have significantly changed the way that BakerScan operated in order to detect cancer cells. Reporter James Somers for the New Yorker also thought the same thing. He asked the CEO of Brain Co., Hisashi Kambe, the same question.
I asked Kambe how it worked — did it use deep learning? “Original way,” he said. Then, with a huge smile, “Same as bread.”
Nope, same as bread.
The Impact of What One Creates
I often wonder if the team developing BakeryScan back in 2008 ever even imagined their AI would grow into this. Though I highly doubt anyone could have foreseen this outcome. A doctor realizing that pastries look almost like cancer cells under a microscope, then reaches out to the company on a whim! That single event led this AI from identifying pastries to assisting doctors in saving lives. I imagine that back in 2008, the Brain Co. team had some really frustrating days developing BakeryScan. I wish during those days somebody could have gone back in time and shown them what their little AI would become.
This is one of my favorite stories because it is so much more than a story. It’s a tale of hardship, triumph, and dedication. Who would have thought that a bread scanner could save countless lives? I can almost guarantee they didn’t. Their dedication and passion for the project led them to a destination they didn’t even know existed until they had arrived.
This story is the primary reason I want to study Data Science. Maybe it won’t be training an AI with deep learning, or creating a neural network, but I do hope that I will have the privilege of creating something impactful; something that goes beyond what I ever could have intended for it. Something I can be proud and passionate about. The field of Data Science is the place where I could best do something like that. I hope the Brain Co. team is proud of themselves, knowing that even if their little AI isn’t used forever, the impact of their creation will be remembered for a long time to come.
References
“AI Designed to Distinguish between Types of PASTRIES IDENTIFIES Cancer Cells with 99% Accuracy.” Express Digest, expressdigest.com/ai-designed-to-distinguish-between-types-of-pastries-identifies-cancer-cells-with-99-accuracy/.
“The AI PASTRY Scanner That Is Now Fighting Cancer.” The AI Pastry Scanner That Is Now Fighting Cancer — CES 2022, 20 May 2018, www.ces.tech/Articles/2021/May/The-AI-Pastry-Scanner-That-Is-Now-Fighting-Cancer.aspx.
“BakeryScan -Identify All Pastries in a Flash!-.” YouTube, YouTube, 24 Dec. 2020, www.youtube.com/watch?v=vTwxKXokVB8.
Somers, James. “The Pastry A.I. That Learned to Fight Cancer.” The New Yorker, 18 Mar. 2021, www.newyorker.com/tech/annals-of-technology/the-pastry-ai-that-learned-to-fight-cancer.
BakeryScan and Cyto-AiSCAN was originally published in Towards Data Science on Medium, where people are continuing the conversation by highlighting and responding to this story.
from Towards Data Science - Medium https://ift.tt/3kAlF21
via RiYo Analytics
ليست هناك تعليقات