I'm current testing a model for spiral data. After 500 epoches, loss is 0.04 but the result is still unmatch with some part of the training data. (bottom left)
The model has 2 hidden tanh x 16 units running with ml5.js. I choose tanh because it seems to be smoother than ReLU. Apart from that they're the same.
Is this thing caused by the model or by ml5 itself?