Weight initialization can critically affect the speed at which a neural network is able to learn. Under certain circumstances, a poor initialization of weights can prevent a neural network from learning anything.
In this series of examples we will see how a flavor of the canonical LeNet neural network is able to learn from the MNIST dataset, under various weight initialization schemes.
Caffe makes it straightforward to select one of the various supported weight initialization methods. In the prototxt definition of the network: each layer can specify how to initialize weights through a weight_filler field .
For example, in order to set the initial weights to a constant 0.2:
weight_filler {
type: "constant"
value: 0.2
}
The initialization of the bias terms can also be specified through a bias_filler field, although this tends to have a lesser impact on learning so we will focus on the initialization of weights in the following examples.
Most layers make it straightforward to initialize weights randomly from a uniform distribution over [-std, std] by calling their reset(std) method.
However the determination of the range to use has been the subject of a lot of research and we will see that this is often the key to efficient learning in a convolutional network.
The torch-toolbox project implements a handy way of quickly experimenting with different ways to compute the
std parameter to use in the call to the reset() method.
If you wish to use non-default weight initialization, you may write:
local method = 'xavier'
local model = require('weight-init')(model, method)Assuming model points to your existing model then the new model will be initialized according to the "Xavier" method.
NOTE: at the time of writing, the
torch-toolboxproject does not handle all layers (most notablycudnn.SpatialConvolution) so you might need to tailor the script to your specific needs.
This is the default initialization scheme in Caffe: weights are initialized to a constant 0. LeNet is unable to learn anything with this initialization, presumably because the constant initialization makes it difficult to break the symmetry of activations:
This is how weights are initialized in Caffe when choosing the uniform filler. By default, weights are picked randomly from a uniform distribution over [0,1].
LeNet is unable to learn anything with this initialization:
The parameters of the uniform distribution may be changed through min and max keys. For example:
weight_filler {
type: "uniform"
min: -0.5
max: 0.5
}
Under this initialization scheme the network is able to learn something:
By default Torch7 uses the initialization scheme described in LeCun, Yann A., et al. "Efficient backprop." Neural networks: Tricks of the trade. Springer Berlin Heidelberg, 2012. 9-48.for linear and convolutional layers.
Initial weights are randomly picked from a U[-1/sqrt(fan_in),1/sqrt(fan_in)] distribution.
This is what the learning curve looks like under this scheme:
This performs reasonably well, considering the fact that this is a fully automatic way of setting the initial weights which does not require hand picking the range of the uniform distribution.
This initialization was first introduced in Glorot, Xavier, and Yoshua Bengio. "Understanding the difficulty of training deep feedforward neural networks" International conference on artificial intelligence and statistics. 2010.. Initial weights are randomly picked from a U[-sqrt(3/n),sqrt(3/n)], where n is the fan-in (default), fan-out, or their average. According to the paper this initialization is particularly well suited for layers that are followed by a Sigmoid activation unit. Learning is very efficient under this scheme:
This initialization was first introduced in He, Kaiming, et al. "Delving deep into rectifiers: Surpassing human-level performance on imagenet classification." arXiv preprint arXiv:1502.01852 (2015).
This initialization scheme is well suited for layers that are followed by a ReLU non-linear unit, particularly in the case of very deep networks.
Even in this "shallow" example (which does use a ReLU non-linearity) we can see that the network is able to learn even more quickly than under the Xavier initialization scheme:
