Neural decoding of inferior colliculus multiunit activity for sound category identification with temporal correlation...

Area of Science:

• Neuroscience
• Bioacoustics
• Computational Auditory Neuroscience

Background:

• Neural transformations for sound perception are not fully understood.
• The inferior colliculus (IC) is hypothesized to process temporal sound features for natural sound identification.
• Multi-unit activity (MUA) in the IC may encode information about natural sounds.

Purpose of the Study:

• To predict natural sounds from MUA signals in the IC.
• To investigate the role of temporal characteristics in MUA for sound identification.
• To evaluate the effectiveness of deep learning models for classifying natural sounds based on neural activity.

Main Methods:

• Collected publicly accessible MUA data from the IC.
• Converted temporal correlation values of MUA signals into image representations.
• Utilized denoising techniques and varying segment sizes to create data subsets.
• Applied transfer learning from pre-trained Convolutional Neural Networks (CNNs) (Alexnet, Googlenet, Squeezenet) for feature extraction.
• Employed classifiers including Support Vector Machines (SVM), k-nearest neighbour (KNN), Naive Bayes, and Ensemble methods.
• Evaluated performance using accuracy, sensitivity, specificity, precision, and F1 score.

Main Results:

• Classification accuracy significantly improved after applying denoising methods.
• Deep learning models, particularly CNNs, demonstrated effectiveness in classifying sounds from neural data.
• Transfer learning approaches facilitated robust feature extraction for sound classification.
• Noise reduction positively impacted the performance of all tested classifiers.

Conclusions:

• The temporal characteristics of MUA in the IC contain predictable information about natural sounds.
• Denoising MUA signals enhances the accuracy of sound classification, supporting the role of the IC in auditory perception.
• CNNs combined with transfer learning offer a powerful framework for analyzing neural correlates of sound perception.