Welcome to June’s special edition of Eye on A.I.
It’s well known that dogs have superpowers when it comes to smell. A dog can easily sniff out the equivalent of a teaspoon of sugar dissolved in two Olympic swimming pools—about one part in 1 billion. In fact, some dogs are known to be able to detect odors at concentrations as small as one part per trillion. This is a sensitivity hundreds of thousands of times greater than humans. It is the reason that dogs have been trained to detect everything from drugs to cancer.
But this training is difficult, time-consuming and expensive, and each animal can usually only be trained to find one set of target scents. Meanwhile, our best electronic means of detecting volatile organic compounds in the air—or on surfaces—are Gas Chromatography/Mass Spectrometry (GC/MS) machines that can cost more than $100,000 and take at least 20 minutes to analyze a sample.
Now a startup called Canaery thinks it can read the neurons firing in a dog’s olfactory bulb in real-time and, with the help of machine learning, turn the animal into a detection device able to suss out a vast range of molecules, all without the animal having to be specially trained. “This does for scent what machine vision did for sight,” says Gabriel Lavella, Canaery’s founder and chief executive officer.
Canaery, which has a lab in Florida, is developing a small electrode array—Lavella says it is as thin as tissue paper and only a quarter the size of a postage stamp—that would be inserted into the dog’s nose and sit on top of the olfactory bulb. When neurons in the bulb fire, the array would pick up the signal and transmit that firing data wirelessly to a computer that can be attached to the dog’s harness or collar, where the signals are processed. The company is trying to train A.I. software to recognize the firing patterns associated with particular scents. In terms of a machine learning challenge, Lavella says that the neuron firing patterns associated with scents are actually far easier for software to pick out than many tasks involving computer vision, such as object recognition.
In some cases, for instance in learning to detect certain cancers, researchers don’t even need to know what the exact compound is that the dog is smelling. The software just has to find the right neuron firing pattern associated with a particular kind of cancer, Lavella says. It can then use the pattern itself as a biomarker of the disease.
What does a dog’s nose know? A.I. may soon tell us
Startup Canaery thinks scent is the next sense A.I. will conquer—with help from some four-footed friends.

New York-based startup Canaery thinks that an implantable electrode array and A.I. will give us the ability to use service dogs, such as this one, as detectors for a wide range of scents, able to determine the presence of everything from drugs to cancer. Currently, most detection dogs are only trained to find one thing. This one is trained to find explosives and is seen working at Paris's Orly Airport in 2018. Eric Piermont—AFP/Getty Images
Subscribe to unlock this article and get full access to Fortune.com
Already a member? Login
Welcome to June’s special edition of Eye on A.I.
It’s well known that dogs have superpowers when it comes to smell. A dog can easily sniff out the equivalent of a teaspoon of sugar dissolved in two Olympic swimming pools—about one part in 1 billion. In fact, some dogs are known to be able to detect odors at concentrations as small as one part per trillion. This is a sensitivity hundreds of thousands of times greater than humans. It is the reason that dogs have been trained to detect everything from drugs to cancer.
But this training is difficult, time-consuming and expensive, and each animal can usually only be trained to find one set of target scents. Meanwhile, our best electronic means of detecting volatile organic compounds in the air—or on surfaces—are Gas Chromatography/Mass Spectrometry (GC/MS) machines that can cost more than $100,000 and take at least 20 minutes to analyze a sample.
Now a startup called Canaery thinks it can read the neurons firing in a dog’s olfactory bulb in real-time and, with the help of machine learning, turn the animal into a detection device able to suss out a vast range of molecules, all without the animal having to be specially trained. “This does for scent what machine vision did for sight,” says Gabriel Lavella, Canaery’s founder and chief executive officer.
Canaery, which has a lab in Florida, is developing a small electrode array—Lavella says it is as thin as tissue paper and only a quarter the size of a postage stamp—that would be inserted into the dog’s nose and sit on top of the olfactory bulb. When neurons in the bulb fire, the array would pick up the signal and transmit that firing data wirelessly to a computer that can be attached to the dog’s harness or collar, where the signals are processed. The company is trying to train A.I. software to recognize the firing patterns associated with particular scents. In terms of a machine learning challenge, Lavella says that the neuron firing patterns associated with scents are actually far easier for software to pick out than many tasks involving computer vision, such as object recognition.
In some cases, for instance in learning to detect certain cancers, researchers don’t even need to know what the exact compound is that the dog is smelling. The software just has to find the right neuron firing pattern associated with a particular kind of cancer, Lavella says. It can then use the pattern itself as a biomarker of the disease.
The company has just raised $4 million in seed capital in a funding round led by Breakout Ventures and including participation from Dolby Family Ventures, KdT Ventures, and SOSV. “For years we have looked at approaches to digitize scent and have never found a solution as elegant and scalable as Canaery’s technology,” Lindy Fishburne, Breakout Ventures managing partner and a Canaery board member, said in a statement.
So far, Canaery has been using rats—which also have a keen sense of smell—not dogs, for its research. Lavella says the company plans to begin field trials of its technology using rats by the fourth quarter of this year. He says it will take 18 months to two years to have a commercial product involving dogs available. He says the company’s first customers will likely be agencies that do port and border security. And he says there are huge markets for the detection of pests—from cotton weevils and bed bugs—as well as plant and animal diseases that can devastate agriculture. Medical uses of the technology, such as detecting cancer, will come later, he says, as the process to get the method approved by the FDA requires clinical trials.
Lavella says that one of the big advantages of using this kind of neuron-computer interface over simply using a trained dog is that the same animal can then be used to cover many more scents. Another advantage is that the device can determine concentrations of a scent, not merely whether it is present. He says that drug smugglers have discovered that a good way to get drugs past dogs is to sprinkle a tiny bit of the contraband everywhere in a container, for example. This way the dog alerts on everything and the customs agents can’t tell where exactly the drugs are hidden. Canaery’s device won’t be thrown off by such tactics, Lavella says. It will allow those monitoring it to tell when concentrations of a scent are rising or dissipating and help them locate a hidden stash.
Thanks to A.I., we may soon know what a dog’s nose knows.
Jeremy Kahn
@jeremyakahn
jeremy.kahn@fortune.com
Update, July 1: This story has been updated to reflect the location of Canaery’s new lab in Florida and also to clarify that in future versions of its device, data from the olfactory implant will be transmitted wirelessly to the signal processing computer.
A.I. IN THE NEWS
U.K. will change its intellectual property laws to enable more data and text mining. The country’s Intellectual Property Office (IPO) announced, following a two month-long consultation, that it will revamp Britain’s IP laws to allow data and text mining “for any purpose,” without the ability of rightsholders to prohibit data mining or force users to pay additional license fees to perform data mining, beyond whatever the users need to pay to access the data. Previously, only non-commercial data and text mining were exempted from IP protections. The new U.K. rule runs counter to the European Union’s current Copyright in the Digital Single Market directive, which also only provides an exemption for academic or scientific research. Tech Crunch has more on the proposed changes. At the same time, the IPO said it would not change current rules that allow a computer program to be awarded copyright on works of art for up to 50 years, but disallow an A.I. system from holding a patent for an invention.
FBI warns on deepfakes being used in job interviews for sensitive tech jobs. The U.S. Federal Bureau of Investigation says it has received mounting complaints about people attempting to use deepfakes—highly-realistic fake videos of people created using A.I.—in job interviews for sensitive technology jobs at corporations, according to a story in tech publication The Register. The FBI thinks the deepfakes are being used by criminal gangs hoping to gain access to sensitive databases containing personal information, including credit card numbers, and possibly also trade secrets. “In these interviews, the actions and lip movement of the person seen interviewed on-camera do not completely coordinate with the audio of the person speaking. At times, actions such as coughing, sneezing, or other auditory actions are not aligned with what is presented visually,” said the FBI in a public service announcement.
New A.I. tool is helping people find ancestors who may have perished in the Holocaust. Google engineer Daniel Patt, a descendent of four Holocaust survivors, created the A.I. software From Numbers to Names which allows a user to upload a photo of someone and then find 10 photos of Holocaust victims who could be a match—the idea being that the people in the Holocaust-era photographs are possible relatives of the person in the uploaded photo. Patt told The Jerusalem Post he was inspired to create the software following a 2016 trip to the POLIN Museum on Polish Jewish history in Warsaw. “I couldn’t shake the feeling that I had potentially walked past a photo of a family member without even knowing it,” he told the paper. The U.S. Holocaust Memorial Museum has begun using From Numbers to Names on its photograph collection. The software can also find images from other Holocaust museum collections. Geddy Lee, a member of the Canadian rock band Rush, said he used the software to find never-before-seen photos of his late mother, a Holocaust survivor, and her family, from when they were prisoners in concentration camps or displaced person camps immediately after the war.
Helping chicks in distress thanks to A.I. Scientists at the City University of Hong Kong trained an A.I. system to identify the unique alarm cries that baby chickens make when they are distressed or in trouble. The system could be used to help farmers come to the aid of the chicks helping more of them survive in chicken farms, according to a story in Science.