Week 2 (June 3 – June 9) Progress

Added the margin method to the classdef ClassificationSVM.

This week, I completed my work on the margin method. I used the following resources:

  • Libsvm PDF by CSIE: It contains the mathematical formulations implemented in LIBSVM, which I used to cross-check against MATLAB's documentation.
  • README provided with the LIBSVM package for MATLAB/Octave.
  • Matlab's Documentation on the margin method.

Note:

  • The output of the margin method is not matching because of differences in the internal optimizations of MATLAB's implementation and LIBSVM.
  • In the MATLAB documentation, I only found the formula for binary classification. However, LIBSVM supports multi-class classification as well. Therefore, I had to add a condition in the margin method to restrict it to binary class SVM objects. Training and prediction can still be done for multi-class classification.
  • MATLAB's ClassificationSVM only supports binary or one-class SVM. However, my classdef supports multi-class SVM as well. I will be discussing this with Andreas in the meeting on 11th June 2024 to decide whether to stick to MATLAB compatibility and restrict usage to binary classes only or provide extra options (multi-class classification) to users.

In the next week, i.e., Week 3 (June 10 – June 16), I will be working on the Loss method, which, according to the timeline, was targeted for Week 4. I am trying to create a frontlog, as a frontlog is always better than a backlog. 😉

Comments

Post a Comment

Popular posts from this blog

My Plans for Google Summer of Code 2024: A Brief Overview of My Project and Timeline.

  My Plans for Google Summer of Code 2024: A Brief Overview of My Project and Timeline. Hello everyone! I am Pallav Purbia from Rajasthan, India currently pursuing my Btech in Electrical and Electronics Engineering with a minor degree in Artificial Intelligence and Machine learning from National Institute of Technology Delhi. I am excited to share that I will be contributing to the  Statistics package  of  GNU Octave  this summer.  Andreas Bertsatos  will be mentoring me throughout this project. I feel incredibly fortunate to be working with him. My Fork of the repository can be found  here  where you can check out the progress of the project. Weekly updates and the project’s overall progress will be documented in  my blog . Overview of the Project: Octave is a pioneer in open-source mathematical-oriented programming language for scientific computing and numerical computation and...
Read more

Week 1 (May 27 - June 2) Progress

Week 1 (May 27 - June 2) Progress Implemented the constructor of the ClassificationSVM and added the predict method to the classdef . This week, I started my work by creating the necessary files ( ClassificationSVM , ClassificationNeuralNetwork , fitcsvm , fitcnet ) and placing them in the appropriate folders. Note: I have added ClassificationNeuralNetwork and fitcnet to my fork, but not much work has been done on them yet. According to the timeline, I will be focusing on these files at the end of July and August. I wrote the help documentation for ClassificationSVM and fitcsvm in Texinfo, within their respective files. The fitcsvm Texinfo contains help related to the expected inputs from the user. The ClassificationSVM Texinfo contains help related to the returned object properties. I studied the svmtrain and svmpredict functions from LIBSVM for Octave. I will be using them for the background algorithm. I was implementing the following tasks in the classdef : Taking the inp...
Read more

Week 7 (July 8 –July 14) Progress

Started studying algorithms and mathematics for implementing the ClassificationNeuralNetwork class This week, I started work on the ClassificationNeuralNetwork.m and fitcnet.m files. I also began studying the mathematics behind neural network classifiers. After reviewing the  MATLAB documentation , I identified the following components for the ClassificationNeuralNetwork class: Weights initializer: Layer weights are initialized using either 'glorot' or 'he'. Biases initializer: Biases are initialized using either 'zeros' or 'ones'. Loss function: Cross entropy. Solver used: L-BFGS (Limited-memory quasi-Newton code for bound-constrained optimization). Initially, I am focusing on getting the forward propagation to work properly. You can see my work  here . Resources I found helpful include: " Neural Network from scratch in Python " by Sentdex (unfortunately, the playlist was never completed). Andrew Ng's  lecture , where he mentions t...
Read more