This is a bad attempt with MATLAB's most recent tool 'classificationLearner'.

This is a bad attempt for ‘Udacity - Self-Driving Car NanoDegree: Introduction to Neural Networks’. I got a poor result with MATLAB's most recent tool 'classificationLearner'. It has only the accuracy of 75%. The classification for students who will be accepted by a university are a mess with the accuracy of 32%.

Is that all? It should be better!

I have learned some basic ideas about Neural Networks from school, and I have excellent impression of these classifiers. Casually, I saw Machine Learning Made Easy and Signal Processing and Machine Learning Techniques for Sensor Data Analytics. So, I use the ‘classificationLearner’ with default settings. Results hurt my feelings.

There could be two reasons. One is that I have misused the toolbox of MATLAB. Another one is more advanced technologies should be used.

The teacher of Udacity uses the data in 'binary.csv' to show the basic idea of neural network. Variable 'gre', 'gpa' and 'rank' are scores of students. The variable 'admit' reflects that a student will or will not be accepted by a school. The aim of this exercise is to design a classifier which estimates the decision of the school. The teacher in Udacity suggests using dummy variable to process the variable 'rank', which means 'rank=1' will be '1 0 0', 'rank=2' will be '0 1 0' and so on. My m file 'ClassificationNN' will do this. (why not process different ranks separately?) The data and results can be download from https://github.com/BlueBirdHouse/CarND-IntroductiontoNeuralNetworks .

Thank you.

[Stephen23]

@Blue Bird: what is your question?

Answer 1

[Greg Heath]

If you have to use MATLAB to solve a classification problem,

FIRST: Use the doc and help commands to see how to use patternnet for classification

SECOND: See my simple variations by searching

      GREG  PATTERNNET  QUICKIES

THIRD: Typically, only 2 things have to be varied

        1. Number of hidden nodes, H, (the smallest successful value is best for 
stability and performance on unseen data having similar statistical characteristics)

        2. For each candidate value of H, consider 10 different designs with different 
random initial weights.

FOURTH: I have posted zillions of examples in both the NEWSREADER and ANSWERS

Hope this helps.

Thank you for formally accepting my answer

Greg