AI Case Study
University of Washington researchers built the DeepSqueak neural net program to analyse mouse noises
Researchers at the University of Washington have developed a free open use software program to analyse the different noises emitted by mice. The program is intended to help other researchers analyse mice noises in laboratory conditions, a usually difficult and and laborious process, quicker and more effectively.
Industry
Public And Social Sector
Education And Academia
Project Overview
To address issues in USV research, researchers at the University of Washington have created the freely available Deepsqueak software package, which "uses regional convolutional neural networks (Faster-RCNN) to increase detection rate, reduce false positives, reduce analysis time, classify calls, and perform syntax analysis automatically."
Reported Results
When compared with other USV analysis software, DeepSqueak had the highest precision and recall rate under different noise settings. When applied to a dataset of mice noises under different conditions, DeepSqueak found that "males use a simpler syntax when vocalizing around other males, predominantly producing short simple USVs, and transition from complex USVs back to simple USVs. Male mice produced more complex patterns of USVs when exposed to females, with the most complex syntax occurring during exposure to female urine."
Technology
The DeepSqueak system allows for both unsupervised clustering as classification of USVs as well as supervised classification where users can manually create categories and label data which can be then used to train the system.
"Contours extracted from these USVs were passed through DeepSqueak’s “elbow optimized” k-means based clustering algorithm with 4 different input parameters...
"We have also attempted to optimize unguided clustering but included a network that highlights DeepSqueak’s ability to perform user- guided neural network-based classification. To produce this network, DeepSqueak was used to detect ~56,000 USVs from B6D2F1 mouse recordings obtained from Mouse Tube. These USVs were clustered with k-means clustering and the previously mentioned five categories of USVs were isolated, manually reviewed, labeled, and used to train the classification neural network. All USVs from the aforementioned dataset were then classified using this neural network and an analysis of male mouse syntax during exposure to male mice, female mice, and female urine was performed"
Function
R And D
Core Research And Development
Background
"Rodents engage in social communication through a rich repertoire of ultrasonic vocalizations (USVs). Recording and analysis of USVs has broad utility during diverse behavioral tests and can be performed noninvasively in almost any rodent behavioral model to provide rich insights into the emotional state and motor function of the test animal. Despite strong evidence that USVs serve an array of communicative functions, technical and financial limitations have been barriers for most laboratories to adopt vocalization analysis... Investigator analysis of USV recordings is slow and laborious, while existing automated analysis software are vulnerable to broad spectrum noise routinely encountered in the testing environment."
Benefits
Data
"To generate the default detection networks, hundreds of rat and mouse USVs were manually isolated from our own recordings and from external labs." For the supervised learning classification method, 56,000 USVs taken from the database Mouse Tube.
Audio recordings are converted into sonograms by the program.