Is this image too complex for the shallow NN classifier?

Question

I am trying to classify a series of images like this, with each class containing images taken from a similar cellular structure:

I built a simple network in Keras to do this, structured as:

1000 - 10 

An unchanged network achieves very high (> 90%) accuracy according to the MNIST classification, but almost never exceeds 5% for these types of images. Is it because they are too complicated? My next approach would be to try and stack deep autocoders.

Answer 1

Seriously - I do not expect any non-revolutionary model to work well with this data type.

• The non-conveyor network for MNIST works well because the data is well processed (it is centered in the middle and resized to a certain size). Your images are missing.

• You may notice (in your photographs) that some motifs are re-created again - like these dark dots - with different positions and sizes - if you don’t use a convolution model, you won’t be able to capture it effectively (for example, you have to admit that the dark dot is a little moved to the image as a completely different object).

Because of this, I think you should try the MNIST convolution model instead of the classic one, or just try to create your own.

Answer 2

First question: do you have a longer training session, do you get better accuracy? You may not have prepared enough.

Also, what is the accuracy of the training data and what is the accuracy of the test data? If both are tall, you can work longer or use a more complex model. If training accuracy is better than testing accuracy, you are essentially within your data. (i.e. scaling brute force to fit the model will not help, but smart improvements can, i.e. use convolution networks)

Finally, complex and noisy data may require a lot of data to make a reasonable classification. Therefore, you need many, many images.

Deep stackable autocoders, as I understand that this is an uncontrolled method that is not suitable for classification.