Deep Reservoir Computing for CV

Deep ESN architectures for Object Detection

artificial intelligence, machine learning, deep learning, dreams, publications

Deep Reservoir Computing for Computer Vision

Introduction

The field of computer vision is a domain of artificial intelligence whose main purpose is to enable machine to analyze, process and understand one or more images taken by an acquisition system. This project seeks to apply simple and deep reservoir computing models to image analysis and more exotic models such as echo state models with slow feature analysis-based outputs.

Project on Research Gate : https://www.researchgate.net/project/Deep-Reservoir-Computing-for-Image-and-Video-classification

Tasks

  • MNIST handwritten digits recognition ;
  • Object recognition ;
  • Unsupervised object recognition ;

Goals

  • Explore the capacities of deep reservoir computing methods on compter vision tasks ;
  • Determine if exotic reservoir computing models can be used to identify objects on an unsupervised way ;
  • Test Conceptors and neuro-symbolic system to reason on images ;

Datasets

MNIST handwritten digits

The MNIST database (Modified National Institute of Standards and Technology database) is a large database of handwritten digits that is commonly used for training various image processing systems. The database is also widely used for training and testing in the field of machine learning. It was created by “re-mixing” the samples from NIST’s original datasets. The creators felt that since NIST’s training dataset was taken from American Census Bureau employees, while the testing dataset was taken from American high school students, it was not well-suited for machine learning experiments. Furthermore, the black and white images from NIST were normalized to fit into a 28×28 pixel bounding box and anti-aliased, which introduced grayscale levels.

The MNIST database contains 60,000 training images and 10,000 testing images. Half of the training set and half of the test set were taken from NIST’s training dataset, while the other half of the training set and the other half of the test set were taken from NIST’s testing dataset. There have been a number of scientific paperson attempts to achieve the lowest error rate; one paper, using a hierarchical system of convolutional neural networks, manages to get an error rate on the MNIST database of 0.23%. The original creators of the database keep a list of some of the methods tested on it. In their original paper, they use a support vector machine to get an error rate of 0.8%. An extended dataset similar to MNIST called EMNIST has been published in 2017, which contains 240,000 training images, and 40,000 testing images of handwritten digits and characters.

(Wikipedia)

Models

Model

Type

Task

RC-aESN-C

Vanilla ESN

MNIST

RC-aESN-JS

Vanilla ESN

MNIST

RC-aESN-LS

Vanilla ESN

MNIST

RC-aESN-MS

Vanilla ESN

MNIST

RC-aESN-MTS

Vanilla ESN

MNIST

Results

MNIST

Classifier

Error rate

Add/mult

LeNet-1 (deep learning)

1.7%

1.6 x 10e5

RC-aESN-MTS 1,200 neurons

1.68%

2.5 x 10e5

RC-aESN-JS 1,200 neurons

1.42%

4.5 x 10e6

21 RC-aESN-JS 1,000 neurons

1.25%

15 x 10e6

SVM degree 4

1.1%

14 x 10e6

LeNet-4

1.1%

1.6 x 10e5

LeNet-5

0.95%

4.0 x 10e5

RC-aESN 4,000 neurons

0.93%

Images

   

Publications

Related posts

Client:Self-initiated

0 Comments

Leave a reply

Your email address will not be published. Required fields are marked *

*