zzkzzk1996 / cs520

CS520

Contributors:

• Zekun Zhang  [email protected]
• Ziqi Wang  [email protected]

Project 1

Maze Runner

Project 2

Mine Sweeper

cs520's Issues

direction of CNN

inspired by one hot encode, I think we can use multiple kernel at ONE layer to achieve better performance as a single kernel's learning ability is strongly limited

Activation function choose

I use a Relu function here. But considering that the gradient on 0 does not exist, we can try to use different activation functions such as tanh or sigmoid. it can be a trading off between speed and accuarcy

New structure of CNN

Do we need multiple convolutional layers to reduce the loss further?
under the condition of normalization and current structure, I found that the tanh seems to have the best performance, the lowest loss can reach 6000(randomly generated numbers, 255 x 255, lr=1e-7).