This document discusses experiments with machine learning. It begins by defining machine learning as a program that improves its performance on tasks through experience. It then categorizes machine learning into supervised learning, unsupervised learning, and reinforcement learning. The document goes on to explain regression, classification, linear regression, and logistic regression. It provides an example of classifying human speech to detect whether a phrase contains the word "Google" or not. Features are extracted from audio files using Mel-Frequency Cepstral Coefficients rather than using the raw waveform. The document concludes by providing contact information for the author to ask questions.

1. Experiments with Machine Learning
Yuriy Guts
Solutions Architect

2. First Things First
What Is Machine Learning?

3. “ A computer program is said to learn from experience E
with respect to some class of tasks T and performance measure P,
if its performance at tasks in T, as measured by P,
improves with experience E.
— Tom M. Mitchell

4. Categories of Machine Learning
1. Supervised Learning.
2. Unsupervised Learning.
3. Reinforcement Learning.

7. Regression
Predict a continuous dependent variable
based on independent predictors

11. Linear Regression

15. Classification
Assign an observation to some category
from a known discrete list of categories

16. Logistic Regression
hypothesis = 1 / (1 + exp(-­‐theta' * x));

17. Logistic Regression: Cost Function
hypotheses = sigmoid(X * theta);
cost = (1 / m) * (-­‐y' * log(hypotheses) -­‐ (1 -­‐ y)' * log(1 -­‐ hypotheses));

18. Let’s classify human speech!
Decide whether a spoken phrase contains the word ‘Google’ or not

19. ‘Google’ Detector: Feature Mapping
Options for building X[ ]:
Input: Audio file (WAV, 16 bit mono, 44.1 kHz)
Output: 1 if it contains the word ‘Google’, otherwise 0
1. Use raw waveform as a feature vector.
But: will have 66150 features for a 1.5 second file.
Kinda scary, and easy to overfit.
2. Use Mel-Frequency Cepstral Coefficients (MFCC).
Believed to be closer to human auditory response.
Depending on parameters, can give about 80 features per file.

20. [cepstra, aSpectrum, pSpectrum] = MFCC(waveform);
x = [cepstra(1); cepstra(2); ...; cepstra(n)];

21. Let’s code it up
MATLAB, logistic regression with conjugate gradient optimization

22. yuriy . guts @ gmail . com
linkedin . com / in / yuriyguts
github.com/YuriyGuts/gdg-speech-classifier
Q & A