MNIST HANDWRITING RECOGNITION BY NEURAL NETWORK

In this project, most modern machine learning algorithm “neural network” is implemented on the MNIST dataset of handwritten digits and how training accuracy and testing accuracy varies with the number of hidden layers used is mentioned.  First, the introduction of the neural network is given in section-I. Then underlying theories of neural network in explained in section-II. Section-III describes the whole step-by-step explanation of whole code used in this report. All the computational results are shown in section-IV and the report is summarized and concluded in section-V.

​

Key Words: Machine Learning, Neural Network, Deep Learning, MNIST Data Set, Feature Extraction

DYNAMIC MODE DECOMPOSITION FOR REAL-TIME BACKGROUND/FOREGROUND SEPARATION IN VIDEO STREAM

This report contains the method of dynamic mode decomposition (DMD) for separating video frames into the background (low-rank) and foreground (sparse) components in real-time. DMD terms with frequencies near zero are interpreted as background (low-rank) portions of the given video frames, and the terms with frequencies far from zero are sparse foreground components. An approximate low-rank/sparse separation is achieved at the computational cost of just one singular value decomposition and one linear equation solve. The DMD method that is shown here is working robustly in real-time without any parameter tuning which is ideal for video surveillance and recognition applications. Also, an algorithm is implemented to the video which is not perfectly steady and where the background is also moving a little bit to see how the algorithm performs.

​

Key Words: Video Processing, Dynamic Mode Decomposition, Singular Value Decomposition, Feature Extraction

MUSIC GENRE CLASSIFICATION

This report contains genre classification using supervised machine learning algorithm. Three different tests are performed which are; band classification from three different bands having different genres, band classification from three different bands having the same genre, genre classification from three different genres. To do this, SVD on the real part of the spectrogram version of randomly chosen five seconds clip from 100 songs per genre is done to make training data set. Then 80%-20% - cross-validation is done to check training accuracy and then outside data is fed to Matlab to check its testing accuracy. Many iterations are done to tune parameters and results of those are shown in section five. All the training sets are showing good, ~80-90%, accuracy in cross-validation and there is testing accuracy of ~70% for the test-1, ~55% for the test-2 and ~70% for the test-3 is obtained.

​

Key Words: Machine Learning, Naive Bayes, Gaussian Mixture Model, Support Vector Machine,  Spectrogram, Audio Processing, Feature Extraction

PRINCIPAL COMPONENT ANALYSIS BY EIGEN VALUE DECOMPOSITION

This report contains a principal component analysis of the simple harmonic motion captured by three cameras from a different angle. The main goal of this project is to understand how Principal Component Analysis work and how much data redundancy it can remove when we implement it. The one more major objective is to learn what happens when we perform PCA on noisy data, part-2, and the data which has experimental errors, part - 3 & 4. In this report, data redundancy is removed by diagonalizing, eigen-decomposition of, the covariance matrix and the weight of eigenvalues are plotted to see how many independent data really exist. At the end comparison between PCA of the ideal case, noisy case and inaccurate experiment are discussed.

​

Key Words: Motion Tracking, Video Processing, Principal Component Analysis, Eigen Value Decomposition, Feature Extraction