Mars GaoFollow

Keywords

Neural Networks, Deep Learning, Cognitive Science

Select the category the research project fits.

Mathematical/Computational Sciences

Is this submission part of ICaP/PW (Introductory Composition at Purdue/Professional Writing)?

No

Abstract

For the current architecture of neural networks, it usually requires a high training cost in time and computation. From our perspective, the current methods in deep learn- ing might not be optimal in architecture and it fails to have an effective learning strategy. To solve these problems, in this paper, we would like to introduce the Collaborative Neural Network Group (CNNG). CNNG is a series of neural networks that work collaboratively to handle different tasks separately in the same learning system. It is evolved from a single neural network by our designed algorithm — Reflection. In this way, based on different situations extracted by the algorithm, the CNNG is able to perform different strategies when predicting the input data. In our implementation, the CNNG is combined with several relatively small neural networks. We provide a series of experiments to evaluate the performance of CNNG compared to other learning methods on three public datasets. The CNNG is able to get a higher accuracy with a much lower training cost. With CNNG by reflection, we can reduce the error rate of 74.6% by average and reach a high accuracy for many tasks, which is superior to VGG and ResNet on the tested datasets. For Fashion-MNIST and EMNIST, it can reach 98.81% and 90.88% which is the best performance currently. Moreover, the required training time is usually less than 40 minutes in our experiments. Details can be found in the experiment part.

Download

Share

COinS
 

Reflection learning with Neural Network Ensemble

For the current architecture of neural networks, it usually requires a high training cost in time and computation. From our perspective, the current methods in deep learn- ing might not be optimal in architecture and it fails to have an effective learning strategy. To solve these problems, in this paper, we would like to introduce the Collaborative Neural Network Group (CNNG). CNNG is a series of neural networks that work collaboratively to handle different tasks separately in the same learning system. It is evolved from a single neural network by our designed algorithm — Reflection. In this way, based on different situations extracted by the algorithm, the CNNG is able to perform different strategies when predicting the input data. In our implementation, the CNNG is combined with several relatively small neural networks. We provide a series of experiments to evaluate the performance of CNNG compared to other learning methods on three public datasets. The CNNG is able to get a higher accuracy with a much lower training cost. With CNNG by reflection, we can reduce the error rate of 74.6% by average and reach a high accuracy for many tasks, which is superior to VGG and ResNet on the tested datasets. For Fashion-MNIST and EMNIST, it can reach 98.81% and 90.88% which is the best performance currently. Moreover, the required training time is usually less than 40 minutes in our experiments. Details can be found in the experiment part.