Where does pruning collide?

In development the brain first overproduces synapses and then removes the excess, a process now called synaptic pruning. Changeux and Danchin's 1976 Selective stabilisation of developing synapses as a mechanism for the specification of neuronal networks proposed that activity stabilises some contacts while others are removed, so a network is shaped by loss as much as by growth. Huttenlocher's 1979 study of human frontal cortex measured this rise and fall in synaptic density directly, and Paolicelli and colleagues' 2011 work showed that microglia carry out much of the removal. This meaning is a biological reduction of neural connections in a living nervous system, not a blade on a plant or a branch cut from a computed model.[1][2][3]

Unit

The loss of a synapse or connection between neurons during development

Recovery means

Not applicable

Measured by

Estimated by counting synaptic density from electron microscopy or by imaging and marker-based assays of synapse number

In machine learning, pruning is the removal of parts of a learned model, branches of a decision tree or connections and units of a neural network, to make it smaller and to curb overfitting. Breiman, Friedman, Olshen, and Stone's 1984 Classification and Regression Trees introduced cost-complexity pruning, which cuts a fully grown tree back to the subtree that best trades accuracy against size, and Quinlan's 1987 Simplifying decision trees set out pruning rules for the trees produced by his induction algorithms. This meaning is an operation on a data structure or model inside a computer, deleting elements that add cost without enough benefit, and it is not a plant cut in a garden or a synapse lost in a brain.[4][5]

Unit

One removed element of a model (a subtree, node, connection, or unit)

Recovery means

Not applicable

Measured by

Judged by change in model size and by validation accuracy or error before and after removal