Word
amortized inference
Definition
n. training a neural network once so that future predictions become much faster and cheaper to compute.
n. an approximation technique where the computational cost of per-instance latent variable inference is shared by training an inference model (encoder) across all data points.
Examples
“Using amortized inference allows the image model to reconstruct sketches in a split second.”
“Variational autoencoders rely on amortized inference to bypass slow Markov chain sampling during generation.”
“While amortized inference provides substantial speedups at test time, it introduces an amortization gap relative to exact per-sample gradient optimization.”
Examples
simple
“Using amortized inference allows the image model to reconstruct sketches in a split second.”
contextual
“Variational autoencoders rely on amortized inference to bypass slow Markov chain sampling during generation.”
complex
“While amortized inference provides substantial speedups at test time, it introduces an amortization gap relative to exact per-sample gradient optimization.”
Etymology
From financial amortization (from Old French amortir, 'to extinguish a debt', from Latin mort- 'death') applied to computation by Gershman and Goodman (2014) to denote spreading inference costs over many evaluations.
Etymology adapted from Wiktionary, available under CC BY-SA 4.0.
Domains
Scan code
englishreference.com/q/amortized-inference