Fast, Simple and Accurate Handwritten Digit Classification by Training Shallow Neural Network Classifiers with the...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Deep neural networks dominate complex tasks like image recognition.
• Shallow neural networks are often overlooked for challenging benchmarks.

Purpose of the Study:

• To demonstrate shallow neural networks can achieve high accuracy on benchmarks like MNIST.
• To showcase the Extreme Learning Machine (ELM) approach for efficient neural network training.

Main Methods:

• Utilized the Extreme Learning Machine (ELM) algorithm for training shallow neural networks.
• Implemented enhanced ELM with random receptive field sampling for sparse weight matrices.
• Combined ELM with limited backpropagation for reduced hidden unit requirements.

Main Results:

• Achieved error rates below 1% on the MNIST handwritten digit benchmark.
• Attained error rates below 5.5% on the NORB image database.
• Demonstrated rapid training times of approximately 10 minutes using the ELM approach.

Conclusions:

• Shallow neural networks with enhanced ELM offer competitive accuracy and efficiency.
• ELM is a viable alternative for simpler problems and a potential component in deep learning architectures.