Backpropagation is the central mechanism by which neural networks learn. It is the messenger telling the network whether or not the net made a mistake when it made a prediction.
To propagate is to transmit something (light, sound, motion or information) in a particular direction or through a particular medium. When we discuss backpropagation in deep learning, we are talking about the transmission of information, and that information relates to the error produced by the neural network when it makes a guess about data.
Untrained neural network models are like new-born babies: They are created ignorant of the world, and it is only through exposure to the world, experiencing it, that their ignorance is slowly relieved. Algorithms experience the world through data. So by training a neural network on a relevant dataset, we seek to decrease its ignorance. The way we measure progress is by monitoring the error produced by the network each time it makes a prediction.
The knowledge of a neural network with regard to the world is captured by its weights, the parameters that alter input data as its signal flows through the neural network towards the net’s final layer, which will make a decision about that input. Those decisions are often wrong, because the parameters transforming the signal into a decision are poorly calibrated; they haven’t learned enough yet. Forward propagation is when a data instance sends its signal through a network’s parameters toward the prediction at the end. Once that prediction is made, its distance from the ground truth (error) can be measured.
So the parameters of the neural network have a relationship with the error the net produces, and when the parameters change, the error does, too. We change the parameters using optimization algorithms. A very popular optimization method is called gradient descent, which is useful for finding the minimum of a function. We are seeking to minimize the error, which is also known as the loss function or the objective function.
A neural network propagates the signal of the input data forward through its parameters towards the moment of decision, and then backpropagates information about the error, in reverse through the network, so that it can alter the parameters. This happens step by step:
You could compare a neural network to a large piece of artillery that is attempting to strike a distant object with a shell. When the neural network makes a guess about an instance of data, it fires, a cloud of dust rises on the horizon, and the gunner tries to make out where the shell struck, and how far it was from the target. That distance from the target is the measure of error. The measure of error is then applied to the angle and direction of the gun (parameters), before it takes another shot.
Backpropagation takes the error associated with a wrong guess by a neural network, and uses that error to adjust the neural network’s parameters in the direction of less error. How does it know the direction of less error?
A gradient is a slope whose angle we can measure. Like all slopes, it can be expressed as a relationship between two variables: “y over x”, or rise over run. In this case, the
y is the error produced by the neural network, and
x is the parameter of the neural network. The parameter has a relationship to the error, and by changing the parameter, we can increase or decrease the error. So the gradient tells us the change we can expect in
y with regard to
To obtain this information, we must use differential calculus, which enables us to measure instantaneous rates of change, which in this case is the tangent of a changing slope expressed the relationship of the parameter to the neural network’s error. As the parameter changes, the error changes, and we want to move both variables in the direction of less error.
Obviously, a neural network has many parameters, so what we’re really measuring are the partial derivatives of each parameter’s contribution to the total change in error.
What’s more, neural networks have parameters that process the input data sequentially, one after another. Therefore, backpropagation establishes the relationship between the neural network’s error and the parameters of the net’s last layer; then it establishes the relationship between the parameters of the neural net’s last layer those the parameters of the second-to-last layer, and so forth, in an application of the chain rule of calculus.
It is of interest to note that backpropagation in artificial neural networks has echoes in the functioning of biological neurons, which respond to rewards such as dopamine to reinforce how they fire; i.e. how they interpret the world. Dopaminergenic behavior tends to strengthen the ties between the neurons involved, and helps those ties last longer.