An Implementation of a Support Vector Machine for Image Classification

Undergraduate #347
Discipline: Technology and Engineering
Subcategory: Computer Engineering

Romeo Kpolo - Onondaga Community College

---

Support Vector Machines (SVMs) are a class of supervised learning algorithms first conceived in the 1960s by Vladimir N. Vapnik. An SVM classifies the data by constructing an Ndimensional optimal hyperplane that separates the data into two classes. The initial use of the SVM algorithms was in their application to handwritten digit recognition. The Mixed National Institute of Standards and Technology (MNIST) database is a large database of handwritten digits used to train the SVM algorithms. Modern implementations of SVM algorithms on the MNIST database have yielded error rates as low as 1.1%. This is the same error rate found in some carefully constructed neural network algorithms. A step-by-step method for applying the SVM learning algorithm for data classification of the MNIST dataset is presented, verifying a low error recognition rate and further supporting the SVM algorithm as an important example of a kernel method depending only on the dot-products of data for classification.

Funder Acknowledgement(s): Funding was provided by NSF/ LSAMP/CSTEP.

Faculty Advisor: Dhireesha Kudithipudi, dxkeec@rit.edu