Neural network
A model made of layers of computing units connected by weights fitted to the data, able to learn very complex relationships, dominant for images and text, more debated for pricing on tabular data.
An artificial neural network is a model made of successive layers of computing units, each combining the outputs of the previous layer according to weights, then applying a nonlinear transformation. Training adjusts these weights, sometimes billions of them, to minimize error on the data, through gradient backpropagation. When the network has many layers, it is called deep learning. These models can approximate relationships of a complexity that classical methods do not capture, and since the early 2010s they have transformed image and speech recognition and text processing. In insurance their use is clear for unstructured data: damage photos, documents, call audio, text. For pricing on tabular data, their advantage over tree methods is more debated, and their opacity creates justification difficulties. Actuarial research has proposed architectures combining a neural network with the readable structure of a generalized linear model, to keep variable-by-variable interpretation while capturing complex effects.
In 2012, a deep network won the ImageNet image recognition competition with an error rate of about 15 percent, against about 26 percent for the best competitor using classical methods, a gap that triggered the general adoption of deep learning, from which today's photo-based damage estimation tools derive.
neural network, deep learning, artificial neural network
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.