Deep learning deciphers what rats are saying
For many years, researchers knew that rodents’ squeaks told a lot about how the animals are feeling. Much like a wagging tail on a dog, certain vocalisations indicate the rodents are happy. Conversely, other vocalisations indicate the rodents are stressed, or even depressed.
But why were they interested in the rodents’ moods? These researchers wanted to understand the rodents’ responses to various stimuli. This can help researchers determine the best way to help people who are addicted or depressed. They would be able to tell if a treatment helped reduce the feelings of depression by simply analysing how the rodents chatted.
Rat chatter is difficult to decode since rodents communicate largely in ultrasonic vocalisations (USVs) that human ears cannot hear. USVs range from 20 to 115 kHz, while humans can typically hear sounds from 20 Hz to 20 kHz.
Up until now, researchers have relied heavily on time-consuming, manual analysis of rodent chatter. The vocalisations are at such a high frequency, researchers had to slow down the recordings in order to hear them. Even with specialised microphones, tagging and categorising the high-pitched squeaks in recordings is labour intensive. These methods are also vulnerable to human error and misinterpretation.
“In the past, researchers have recorded these to gain better insights into the emotional state of an animal during behaviour testing,” said Professor John Neumaier, from the Department of Psychiatry & Behavioural Sciences at the University of Washington (UW). “The problem was that manual analysis of these recordings could take 10 times longer to listen to when slowed down to frequencies that humans can hear. This made the workload exhaustive and discouraged researchers from using this natural read-out about animals’ emotional states.”
Using deep learning to analyse the USVs
Prof Neumaier turned to artificial intelligence (AI) to automate the process, working with UW postdoctoral fellow Dr Kevin Coffey and Russell Marx, a technician in the Psychiatry & Behavioural Sciences, to create DeepSqueak — deep learning software that detects and analyses USVs. Their research was recently published in Nature journal Neuropsychopharmacology.
“We can train the software to analyse these calls in a way that is much more similar to how humans learn,” said Dr Coffey. “Rather than mathematically describing what a vocalisation is, we just show it pictures and examples.”
DeepSqueak works by turning an audio problem into a visual problem. The input to DeepSqueak is an audio file (.WAV or .FLAC). DeepSqueak splits the audio files into short segments and then converts these segments into images (sonograms).
The sonograms are fed into a deep learning AI program that identifies and classifies the images, much like the AI used in self-driving cars to identify stop signs and lane markers. It first decides if a squeak is present in the sonogram and, if so, what type of squeak it is.
“DeepSqueak uses biomimetic algorithms that learn to isolate vocalisations by being given labelled examples of vocalisations and noise,” Marx said.
The team started DeepSqueak using example code, Object Detection Using Faster R-CNN Deep Learning, from the MathWorks website. From there, they developed the DeepSqueak software package and GUI in MATLAB. DeepSqueak uses Computer Vision System Toolbox, Curve Fitting Toolbox, Image Processing Toolbox, Parallel Computing Toolbox and Deep Learning Toolbox.
Technology can help develop better treatments for addiction
The research team is focused on psychiatry and behavioural science. This non-invasive research found that the rodents are happiest when anticipating a reward, such as sugar, or playing with their peers. They also found male rodents behaved differently when female rodents were around. No big surprise there.
Prof Neumaier said his goal is to develop treatments for stress disorders and addiction. DeepSqueak will help the lab get there much faster by making deciphering ultrasonic vocalisations convenient and quick.
“If scientists can understand better how drugs change brain activity to cause pleasure or unpleasant feelings, we could devise better treatments for addiction,” he said.
The team has made DeepSqueak available to all researchers so they can create their own analysis. The program can currently identify approximately 20 different USVs; the team hopes that as others identify and tag various USVs, they’ll be able to create a virtual Google Translate for rat chatter. The code is available at https://github.com/DrCoffey/DeepSqueak.
Originally published here.
Using fluorescence images from live cells, researchers have trained an artificial neural network...
When Bumrungrad International Hospital first went live with the InterSystems TrakCare electronic...
Image recognition AI has the potential to revolutionise medical diagnostics, yet its current...