Weight initialization is an important consideration in the design of a neural network model.
The nodes in neural networks are composed of parameters referred to as weights used to calculate a weighted sum of the inputs.
Neural network models are fit using an optimization algorithm called stochastic gradient descent that incrementally changes the network weights to minimize a loss function, hopefully resulting in a set of weights for the mode that is capable of making useful predictions.
This optimization algorithm requires a starting point in the space of possible weight values from which to begin the optimization process. Weight initialization is a procedure to set the weights of a neural network to small random values that define the starting point for the optimization (learning or training) of the neural network model.
… training deep models is a sufficiently difficult task that most algorithms are strongly affected by the choice of initialization. The initial point can determine whether the algorithm converges at all, with some initial points being so unstable that the algorithm encounters numerical difficulties and fails altogether.
— Page 301, Deep Learning, 2016.
Each time, a neural network is initialized with a different set of weights, resulting in a different starting point for the optimization process, and potentially resulting in a different final set of weights with different performance characteristics.
For more on the expectation of different results each time the same algorithm is trained on the same dataset, see the tutorial:
We cannot initialize all weights to the value 0.0 as the optimization algorithm results in some asymmetry in the error gradient to begin searching effectively.
Historically, weight initialization follows simple heuristics, such as:
- Small random values in the range [-0.3, 0.3]
- Small random values in the range [0, 1]
- Small random values in the range [-1, 1]
These heuristics continue to work well in general.
We almost always initialize all the weights in the model to values drawn randomly from a Gaussian or uniform distribution. The choice of Gaussian or uniform distribution does not seem to matter very much, but has not been exhaustively studied. The scale of the initial distribution, however, does have a large effect on both the outcome of the optimization procedure and on the ability of the network to generalize.
— Page 302, Deep Learning, 2016.
Nevertheless, more tailored approaches have been developed over the last decade that have become the defacto standard given they may result in a slightly more effective optimization (model training) process.
These modern weight initialization techniques are divided based on the type of activation function used in the nodes that are being initialized, such as “Sigmoid and Tanh” and “ReLU.”