Download presentation
Presentation is loading. Please wait.
1
Mihir Patel and Nikhil Sardana
Neural Networks Pt 2 Mihir Patel and Nikhil Sardana
2
Synopsis
3
Data Techniques Expanding datasets Weight initialization Batching
Narrow gaussian Batching
4
Learning Rate Step size for backpropagation
5
Activation Functions: tanh
y = (ex-e-x)/(ex + e-x) Benefits Greater derivative = faster learning -1 vs. 0 prevents stagnation
6
Cost Function: Cross-Entropy
Saturation Needs Approaches 0 when correct Always has positive sign Benefits Doesn’t give 0 when incorrect
7
Overfitting Over-matching dataset
8
Solutions to Overfitting
Weight Decay Punish larger weights Dropout
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.