Why does adding random pixels stop my model learning in cnn?

Question

I am using a very simple model to classify a 224x224 RGB image.

For a test, I have labelled my images (2 labels "Green" or "Red", 2,000 images of each) based on colour of a single fixed pixel from upper part of image.

My simple model achieves very high accuracy and very low loss, until I add more "random" pixels to lower part of image.

Why is my model getting confused?

The model easily copes with:

But, when I add more pixels in lower part of image - it achieves only 50% accuracy and high loss:

These project is a bit facile but I am researching a simple image structure I can control.

def VerySimpleNet():
    model = keras.Sequential([
        keras.layers.Conv2D(16, kernel_size=3, activation=tf.nn.relu, input_shape=(224, 224, 3)),
        keras.layers.Dropout(0.4),
        keras.layers.Flatten(),
        keras.layers.Dense(3, activation=tf.nn.softmax)
    ])
return model

EDIT 1

Is it possible to add a bonus to Data Science questions to encourage more answers?

EDIT 2

I would really be interested if anybody was keen to have a more interactive chat(email,txt?) with me on this question.

Answer 1

I think the problem here is that you are trying to relate what you know about the input to what the model is trying to predict. I suggest trying to interpret why your model is giving certain outputs for images.

I came across SHAP a while back when reading about model interpretation. The image plot of SHAP should be able to narrow in on why your model is interpreting a label for a specific Image.

This should lead you to figure out why your model is making the prediction.

Answer 2

One simple interpretation is that adding random noise of course can't help, and enough of it will sink your model performance. The model might eventually figure it out with enough training time, the right settings, etc, but it's having to sift through so much noise to find the signal you hid.

You have relatively few instances (2000) of relatively complex input (224x224x3 images). There are lots of opportunities for your 'random' additions to actually correlate somewhat with the label. What looks like a good feature to fit in the training set fails to generalize in your test set.

The last interpretation I'll offer is that you are using a convolutional layer here, which implies some degree of spatial invariance. What it learns about one 3x3x3 patch it learns equally for all 3x3x3 patches. But you intend it to learn something different about a particular single pixel. It really can't with a single convolutional layer with a small kernel. That is, your simple model is just too simple for this problem. Stacking more convolution + pooling layers might eventually get the spatial specialization that is required to solve your puzzle.