Thursday, September 13 2018
12:30pm - 1:45pm
Operations Research Seminar
Image Pattern Recognition

On tasks such as the recognition of handwritten digits, traditional methods from machine learning and computer vision have always failed to beat human performance.

Inspired by the importance of diversity in biological system, we built a heterogeneous system that could achieve this goal. Our architecture could be summarized in two basic steps.

1. We generate a diverse set of classification hypothesis using Convolutional Neural Networks (CNN) and recent innovative Neural Networks (SNN).

2. Then, we include all the Machine Learning results that have already been trained in a type of parliament classifier where all the hypothesis, despite of their accuracy or methodology, are considered. The judges of the parliament are a new family of Meta Classifiers.

We have applied this new strategy of “real” Deep Learning on the very competitive MNIST handwriting benchmark data set and our method seems to be promising. It surprisingly shows that artificial diversity is a key for success in decision-making.
Speaker:Dr. Massimo Buscema
Affiliation:Semeion Research Center of Sciences of Communication, Rome
Location:SCB 4119


Download as iCalendar

Done