Markets 
To some, it’s just annoying, but to others, it marks the chilling call of the world’s deadliest creature: the buzz of a mosquito.
In an effort to slash the extraordinary death toll wreaked by the insects, a University of Wollongong researcher has harnessed a super-efficient, lo-fi form of AI that can discern a disease-carrying mosquito from a harmless one by tuning in to this buzz.
The detection device is one of a raft of new inventions using tiny machine learning, or TinyML, a low-power type of AI serving as the antithesis to the voracious large language models that require energy-guzzling data centres.
First, Associate Professor Kiran Trivedi wanted to see if TinyML could identify the bird calls floating through his window.
Trivedi, a computer science expert from the university’s Indian campus in Ahmedabad, downloaded bird call data from Google and created a small model deployed onto a tiny chip that could identify the calls of species such as the Indian cuckoo.
Once that was working, he turned to babies. Trivedi’s brother, a pediatrician who works in a neonatal ward, told him anxious new parents often brought in their infants to ask the doctor why they were crying.
“So again, I collected the data for babies’ crying sounds, and created a model for that.”
The result was another experimental device designed to deduce whether a baby was bawling because of hunger, sleepiness or discomfort.
“That is how it started,” Trivedi said. But crying babies, he thought, were hardly the world’s biggest problem. So he pivoted to mosquitoes, which kill 700,000 people per year via malaria, dengue and Zika.
Each species has a unique wing beat. The malaria-carrying kind, Anopheles, has a low-pitched buzz. Culex mosquitoes carry Japanese encephalitis and Ross River virus; they buzz at medium pitch. Aedes aegypti, a dengue-carrying species also found in Australia, flies at the highest pitch.
Trivedi used 40 gigabytes of mosquito sound data, equivalent to 10,000 stored photos, to make a model that was 150 megabytes – still too large for a small chip.
To compress the model, he converted the audio recordings into images called “spectrograms”, which are visual representation of soundwaves. The AI learned to pick up patterns in the images invisible to the human eye.
Then, through a process called “quantisation”, he stripped back and simplified the model until only the information essential for making accurate predictions remained.
The resulting model was a tiny 19 kilobytes, about the size of a five-page Word document.
“I deployed that onto my device, connected a display and some batteries, and it was, surprisingly, a perfectly working device.”
He caught a mosquito in his room in Ahmedabad, in western India, and let it fly near the device. It was detected as an Anopheles mosquito; by far the deadliest species. The device is currently 88.3 per cent accurate.
This is TinyML at work, a form of AI that can operate off-grid, independently, on a tiny chip with little power. With the energy an LLM uses to generate a paragraph of text, a TinyML device can run for weeks.
“For AI, usually we need massive [amounts of] data, a powerful machine, and an internet network or Wi-Fi,” said Associate Professor Jianlong Zhou, an AI researcher at the University of Technology Sydney not involved in Trivedi’s research. “TinyML is different – we can use it in places with no power, no internet, no network.
“TinyML collects data locally, processes it locally, and gives feedback locally, so there’s no privacy issues.”
Zhou has researched TinyML applications in beekeeping, where TinyML nanosensors can monitor a hive’s environment down to humidity, CO2, pheromone imbalances in the queen bee, early chemical signs of Varroa mite infestations, and the vibrations of bees which signal an impending swarming event.
Smart agriculture, as well as wearable health sensors, will be major drivers of TinyML technology in the near future, he says. By one estimate, the number of TinyML devices in operation will spike to 2.5 billion by 2030.
As for Trivedi’s device, he dreams the sensors could be mass-produced for 500 rupees, or about $7.50, and become part of a citywide detection network.
“It’s all about early detection,” he said. “Usually, what happens where I live is that you read in the newspaper that there are a lot of mosquitos breeding, and dengue has been found. But the outbreak has already happened and a lot of people have already been infected.”
The technology he’s working on could catch mosquito outbreaks as they begin, leading to targeted early control measures such as insecticide fogging. He also imagines the model built into smart speakers such as Alexa and alerting people when a disease-causing mosquito is buzzing nearby.
Trivedi displayed his technology at the UN’s AI for Good summit this week.
“If we can identify the mosquito species and help society, it could be a great success for me,” he said.
The Examine newsletter explains and analyses science with a rigorous focus on the evidence. Sign up to get it each week.
Angus Dalton is the science reporter for The Sydney Morning Herald.Connect via X or email.From our partners
Read the full article here
