280 likes | 475 Vues
Convolutional Networks. Extracted from NN for ML Coursera by Prof. Hinton. Weight sharing to detect the same features at different locations in the image Reduces number of free parameters (only 9 different weights here) Brings us some translation invariance
E N D
Convolutional Networks Extracted from NN for ML Coursera by Prof. Hinton
Weight sharing to detect the same features at different locations in the image • Reduces number of free parameters (only 9 different weights here) • Brings us some translation invariance • We want multiple feature maps, each of which shares weights within itself
Rectified Linear Units A smooth approximate to the rectifieris the softplusfunction. f(x) = log (1 + ex) The derivative of softplus is , i.e. the logistic function. • f(x) = max(0,x) • Gradient is defined as 0 if x < 0 1 if x > 0 ON RECTIFIED LINEAR UNITS FOR SPEECH PROCESSING, Zeiler, … Hinton et al
The advantages of using Rectified Linear Units in neural networks are: • It induces the sparsity in the hidden units. • ReLU doesn't face gradient vanishing problem as faced by sigmoid and tanh function. It has been shown that deep networks can be trained efficiently using ReLU even without pre-training. • ReLU can be used in Restricted Boltzmann machine to model real/integer valued inputs.