The Bakery AI Being Used For Cancer Research

Kottke Ride Home
|

Automatic TRANSCRIPT

There's a bakery chain in japan that offers dozens of different kinds of pastries danishes. Donuts croissants sandwiches you go down a cafeteria style line. Pick what you want. Put it on a tray and when you get to the register to pain you put your tray on top of a backlit rectangle screen shows an injury of your train with bright green lines around each item it recognizes each item from among the dozens of pastries on offer at the bakery and then correctly charges. You how does it do that. You'd imagine some kind of relatively recent developments in deep learning neural networks. But in fact it's a more grown technology that got its start. In two thousand seven led by a man named hisashi kombi the new yorker recently profiled combine work with the bakery and how the technology he developed is now being deployed in many other sectors including cancer research. So let's start with the bread problem. Japanese consumers like lots of options and bread is no exception market. Research in japan showed that a bakery that offered a hundred different options sold twice as much as a bakery that offered only thirty and that unwrapped baked goods sold better because people perceived them as fresh but without the wrapping. You don't have labels or barcodes on the pastries so then you've got employees who have to memorize hundred different pastries and this slowdown lines. Because workers would often mix things up or have to take time remembering which peachtree was which not to mention. They handled each unwrapped pastry individually. So it wasn't exactly the most sanitary so this chain wanted to automate the process somehow and they turned to comedies company. Brain to help them out. Rain had already been working for twenty years on finding ways for computers to see something that was long one of the biggest challenges of artificial intelligence. The new yorker explains it really. Well quote as i write. This can look up at my shelves. They contain books and skin of yarn and a tangled cable. All inside a cabinet whose glass enclosure is reflecting leaves in the trees outside my window. I can't help. But parse the scene about a third of the neurons and my cerebral cortex are implicated in processing visual information but to a computer. It's a mess of color and brightness in shadow. A computer has never untangled. A cable doesn't get that glasses. Reflective doesn't know that trees sway in the wind. Ai researchers used to think that without some kind of model of how the world works in that was in it. A computer might never be able to distinguish the parts of complex scenes. The field of computer vision was zoo of algorithms that may do in the meantime and quotes over the last decade this has changed as deep learning and neural networks have been applied and tweaked in real world scenarios. Siri google translate and alpha. Go all rely on deep learning with layers of simulated neurons and their honed by things like tagging people in photos on social media and picking out street lights in those prove. You're not a robot tests on website forms but especially without that kind of passively crowd sourced assistance quoting again. The drawback of deep learning is that it requires large amounts of specialized data the deep learning system for recognizing faces might have to be trained on tens of thousands of portraits. And it won't recognize address unless it's also been shown. Thousands of dresses deep learning researchers therefore have learned to collect labeled data on an industrial scale and quotes and for brain the bakery chain not only with the frequency with which the bakery chain changed their offerings. Make such data required for learning. The new pastries be untenable. They were also several years too early to even consider using deep learning so they built their own system using lots and lots of algorithms by two thousand ten. They built a system with ninety nine percent. Accuracy across fifty types of bread tackling problems like different pastries that look remarkably similar and the same pastries that look different when one is baked. More or one got squished. As developing the back light the pastries have to be placed on to keep the lighting consistent even though they did have to also build a mathematical model to account for inconsistencies in color when it comes to bake times rather than showing assistant thousands of photos of each pastry as one would with a deep learning system they manually tweaked and honed the algorithms on each doughnut and danish until they got it right but their system learns to win. The system isn't sure instead of those green lines around each item. It shows yellow or red lines and prompts the user to select from some suggestions or manually input the product quoting again show bakery scan a pastry never seen on earth and it'll recognize the next one of its kind about forty percent of the time according to brain after just five examples it is ninety percent accurate and after twenty. It's nearly perfect. Moreover whereas deep learning systems are relatively inscrutable. You can't look at a neural network and say exactly why a decision emerged from it bakery scans judgments based as they are on hand engineered system or more particularly if the system. Miss identifies something. You can figure out why these days. It's unusual to develop in the way that brain developed bakery scan. The approach requires a mastery of fine details. It is in spirit artisanal. It takes years during which parameters must be tuned. Special cases accounted for deep learning relieves. You from having to understand how the seasons affect shadows in a donut hole you merely plug enough examples and the network figures it out and with deep learning. The same brain can accomplish different tasks when you feel different. Data deep mind. The alphabet subsidiary used data sets to train a single neural network to beat humans at chess shoghi and go systems that depend on domain specific knowledge as bakery scandal need not just new data but new filters new features in new algorithms before they can be used elsewhere and quotes nonetheless. The technology behind bakery scan now housed under the umbrella. Name of ai scan has gone on to be used in applications as far flung as distinguishing pills in hospitals to counting the number of people in eighteenth century woodblock prints and even spotting incorrectly wired bolts in jet engine parts but the most impressive application came in two thousand seventeen when quoting again. Dr louis pasteur's center for medical research in kyoto saw a television segment about bakery scan. He realized that cancer cells under a microscope looked kind of like bread and quotes. I love that observation but he was right. Ai scan has now been working for a few years on fine tuning their cancer cell detector. Saito a scan now. Being tested at two hospitals is able to look at an entire microscope slide and identify potentially cancerous cells with ninety nine percent accuracy based on features like the color tone size and texture of the nucleus and all round -ness of the cell as they continue to grow and build on their original system. Brain has had to bring in deep learning once recently when covid nineteen pushed bakery owners to start wrapping their pastries and individual packaging. Brain used a deep learning to help their bakery scan system still be able to identify the pastries behind the reflective plastic

Coming up next