Machine Learning

2022 MIT Beaver Works Summer Institute Autonomous Cognitive Assistant (CogWorks) Program

Summary

Learned audio, image, and language analysis techniques. My team of 5 programmed a PyTorch convolutional neural network that classifies sea animals based on image (~90% accuracy) and sound (87% accuracy).

Description

My four teammates liked my idea: an AI that classifies sea animals by image & audio. I combined ideas from articles we found to make the image & audio pipeline. I helped debug the image classifier and worked on a GUI. Teammates implementing audio classification said mel spectrograms generated from the audio were mostly black i.e. unideal to feed the CNN.

I tried plotting 20 spectrograms—all were bright. So I generated spectrograms for all the audio. Many spectrograms lacked frequencies greater than 5000 Hz, meaning the corresponding clips had been sampled at 10000 Hz, not the standard 41000 Hz. Using the sampling rate used when recording the audio, I generated spectrograms, which turned out bright with distinct patterns.

I split the spectrograms into train & test data, balanced the data, and resized & normalized the train data. I debugged the code for initializing & training the CNN.

Code: github.com/eloisezeng/classifying_sea_sounds

Additional Experience

• Machine Learning Coursera course authorized by Stanford University [COMPLETED]
• Reinforcement Learning course by David Silver [Watched 5/10 lectures - IN PROGRESS]