CONFUSION MATRIX RESULTS INFORMATION FOR PRETRAINED NEURAL NETWORK

I got this confusion matrix for pretrained vgg19 model after the validation process, what are all the information obtained by this confusion matrix ?

how can we identified that this model is good for classification of multiclass based on this confusion matrix?

Thank you for your response in advance.

Answer 1

[Monisha Nalluru]

Hi,

In a confusion matrix the rows correspond to predicted class (Output class) and columns corresponds to true class (Target class). The diagonal cells correspond to observations that are correctly classified. The off diagonal cells correspond to incorrectly classified observations. Both the number of observations and percentage of total number of observations are shown in each cell.

In your case ‘a’ is a classified as ‘a’ for 43 times, ‘c’ is classified as ‘a’ for 0 times, ‘d’ is classified as ‘a’ for 4 time and similarly all other values.

In order to identify whether it is good classification model it depends on accuracy which you. In your case it is 84% which is good for few fields and not very good for medical field datasets.

Also refer the following link: https://www.mathworks.com/help/deeplearning/ref/plotconfusion.html

[Deepika B]

thank you mam

[Deepika B]

the accuracy is not improve for my fundus image classification problem mam.

how do i improve the accuracy?

[Monisha Nalluru]

Try using different algorithms or networks.

[Deepika B]

Mam Is alexnet is used for fundus image classification problem? Is there any contrain to use pretrained network for our own tasks? If any features already available in the alexnet for imagenet dataset is not match with our own problem's feature means, we are not even try this network ? why?

[Monisha Nalluru]

Hello,

You can give a try to alexnet and check if it works for your dataset or not.

The constraints can be the size of input dataset and number of target classes which can be changed as per custom dataset.

Moreover, go through the literature of fundus image classification and check if those algorithms is working to your dataset or not.

[Deepika B]

Thank you for your kind response and time.