Scientific Task – Neural Playground

A written scientific task set around TensorFlow Playground, the interactive web interface for experimenting with small neural networks. The fictional John Intelligent has collected five Playground configurations whose behaviour he cannot explain; for each, the contestant explains the phenomenon and suggests a remedy, supplying a Playground link with the modified settings.

Question 1 (4 points): on the spiral dataset the test loss does not decrease; explain why (2) and fix it by changing a single parameter (2). Question 2 (4 points): the training loss drops sharply and then stagnates; explain (2) and remove the stagnation with a minimal change (2). Question 3 (2 points): with linear activations, adding neurons to the hidden layers does not improve the model; explain why. Question 4 (4 points): the test loss does not fall below 0.1 within at most 1000 epochs; explain (2) and fix it without changing the number of hidden layers or neurons (2).

Question 5 (6 points): starting from a configuration with linear activation, batch size 2, learning rate 0.00001, one hidden layer of 7 neurons and only the X1X2 input, reach a test loss of at most 0.1 within at most 1000 epochs with as few changes as possible, without changing the hidden layer count or width, without regularisation, keeping the classification problem type and keeping the X1X2 input; every change must be justified step by step.