I used Tensorflow code and NN model from this article.
Network Details: 2 inputs, 1 hidden layer with 2 neurons, and 1 output layer with 1 neuron. Sigmoid activation function is used. The cost function is the "average over all the training examples" according to the link: [ y * log(y_hat) - (1 - y) * log( 1-y_hat ) ]. Gradient Descent with learning rate of 0.01 is used to train the algorithm. Weights initialized from uniform distribution between -1 and 1, biases initialized to 0.
Questions: (1) Why is the classifier not working very well in the 15-point case? It seems to be satisfied with classifying that corner and disregards the 2 points in the middle. (2) Is there a technique I can use to sort of "nudge it" towards the right solution? Different initialization?