1
$\begingroup$

I wanted to build a digit recognition neural network using MATLAB ANFIS kit.

I started by using the MNIST database and I figured out it's almost impossible to classify 784 dimension data using ANFIS. So, I reduced my data dimension from 784 to 13, using an autoencoder in Python. With the new data, I had about 80 percent accuracy in classification using a sequential model. I implemented my data in MATLAB too.

Since MATLAB treats the problem as a regression problem, I had about 1.5 RMSE after 10 epoch of learning, in both grid partitioning and subtractive clustering, and also the error curve almost seems constant in the process.

Is there any way that I can have less error?

$\endgroup$
2
  • $\begingroup$ ANFIS is too shallow to learn much better. What do you see as state of the art (SOA) results with ANFIS? It is not a modern solution. Why are you considering it? This paper, "A Deep neuro-fuzzy network for image classification" (arxiv.org/pdf/2001.01686v1.pdf), shows that a deeper variant of ANFIS can achieve an accuracy of 99.58%. $\endgroup$ – Brian O'Donnell Jul 29 '20 at 11:46
  • $\begingroup$ @BrianO'Donnell I really don't want to work on it that much but I have to,it's my final course project ,is there any tool for deep ANFIS that I don't have to implement all the stuff by myself? $\endgroup$ – Mohamadreza Abasian Jul 29 '20 at 13:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.