capsule network
In plain wordsA network design championed by Geoffrey Hinton in which groups of neurons, called capsules, each describe a part of an object and its position, so the network can check whether the parts fit together.
A neural networkneural network, n.: A kind of computer program that learns a task from examples by tuning thousands to billions of connections between simple number-crunching units, an arrangement loosely inspired by nerve cells in the brain. built from capsules, small groups of neurons that each output a vector (or matrix) describing one entity, such as an object or a part of one, instead of a single number. In the best-known version, by Sara Sabour, Nicholas Frosst and Geoffrey Hinton (2017), a vector's length gives the chance that its entity is present and its direction encodes properties such as pose, size and deformation. Each capsule predicts, through a learned transformation matrix, the pose that each candidate parent should have, and an iterative 'routing-by-agreement' step passes its output to the parents whose predictions agree, so a face is recognized only when eyes, nose and mouth sit in consistent relative positions. Capsules aimed to preserve spatial relationships that poolingpooling, n.: A step in an image network that shrinks its internal picture by summarizing each small patch with one number, usually the largest or the average, keeping what matters and discarding exact positions. discards, seeking equivarianceequivariance, n.: A property of a model whose output changes in step with its input: move the cat in a photo, and the network's internal 'cat' signal moves the same way. rather than invarianceinvariance, n.: A property of a model whose answer does not change when the input is altered in an irrelevant way, such as a photo classifier that still says 'cat' when the cat moves or the picture is flipped.. They did well on MNIST and on overlapping digits, but routing was computationally expensive and scaled poorly to large natural-image datasets, and capsule networks have seen little practical adoption.
The capsule network separated the overlapping digits that confused the CNN baseline.
Origin
Geoffrey Hinton, Alex Krizhevsky and Sida Wang introduced 'capsules' in 'Transforming Auto-encoders' (2011), after Hinton's decades of advocating part–whole representations. Sara Sabour, Nicholas Frosst and Hinton's 'Dynamic Routing Between Capsules' (arXiv, October 2017; NIPS 2017) launched capsule networks, followed by 'Matrix Capsules with EM Routing' (ICLR 2018).
Related
- convolutional neural networkconvolutional neural network, n.: A kind of neural network built for images and other grid-like data. It slides small pattern detectors across the input, finding edges first, then shapes, then whole objects.
- poolingpooling, n.: A step in an image network that shrinks its internal picture by summarizing each small patch with one number, usually the largest or the average, keeping what matters and discarding exact positions.
- equivarianceequivariance, n.: A property of a model whose output changes in step with its input: move the cat in a photo, and the network's internal 'cat' signal moves the same way.
- invarianceinvariance, n.: A property of a model whose answer does not change when the input is altered in an irrelevant way, such as a photo classifier that still says 'cat' when the cat moves or the picture is flipped.
- godfathers of AIgodfathers of AI, phr.: A media nickname for three scientists — Geoffrey Hinton, Yoshua Bengio and Yann LeCun — whose decades of work on neural networks underpins modern AI. They shared computing's top prize in 2019.
Sources
- Sabour, S., Frosst, N. & Hinton, G. E. (2017). Dynamic Routing Between Capsules. arXiv:1710.09829; NIPS 2017. arxiv.org/abs/1710.09829
Mentioned in
poolingpooling, n.: A step in an image network that shrinks its internal picture by summarizing each small patch with one number, usually the largest or the average, keeping what matters and discarding exact positions., equivarianceequivariance, n.: A property of a model whose output changes in step with its input: move the cat in a photo, and the network's internal 'cat' signal moves the same way., invarianceinvariance, n.: A property of a model whose answer does not change when the input is altered in an irrelevant way, such as a photo classifier that still says 'cat' when the cat moves or the picture is flipped.