THE ACCURACY OF RHYTHM RECOGNITION WITH CONVOLUTIONAL NEURAL NETWORKS ON TRUENORTH PROCESSOR

Open Access
Author:
Galante, Eric Andrew
Area of Honors:
Computer Engineering
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
  • Vijaykrishnan Narayanan, Thesis Supervisor
  • Vijaykrishnan Narayanan, Honors Advisor
  • John Morgan Sampson, Faculty Reader
Keywords:
  • Convolutional Neural Networks
  • TrueNorth
  • Rhythm Recognition
Abstract:
Convolutional Neural Networks (CNNs) are widely used for image processing and have shown a high accuracy in recognizing properties of inputs in order to classify them [5]. However, these CNNs can also be altered to work with other data inputs, such as sound files. In this research, we explore the accuracy and efficiency of training a CNN to recognize 7 classes of drum rhythms. In particular, we explore altering the existing Eedn training code to work with a dataset comprised of 15 recordings of each class and training the network accordingly. We demonstrate that Eedn network can recognize the drum rhythms with an accuracy that increases in conjunction with the size of the data width. Because of the nature of a drum, each hit is basically indistinguishable from another, providing a difficult dataset to learn. Therefore, as the data width increases, each data point represents a longer amount of time for each audio sample and the accuracy increases because the network has the ability to recognize more patterns in the rhythm. However, this increase in data size also increases the memory required to handle the dataset, with the ideal size exceeding well over 100GB. This leads to the inability to test the network to its full extent, which would result in a fully trained rhythm recognition network.