search

ijkilchenko / A_neural_network

A feedforward neural network in python
MIT License
6 stars 4 forks source link

Choose hyperparameters better for each test #2

Open ijkilchenko opened 9 years ago

ijkilchenko commented 9 years ago

Since the cost surface is not convex, random initialization can actually lead gradient descent to not convert sometimes. Some tests (and maybe the demo) should be adjusted by choosing hyperparameters better and maybe doing several random initializations...