Back to results
US 11,537,846 B2Granted
Neural network architecture with concurrent uncertainty output
AssigneeWisconsin Alumni Research Foundation
InventorsSeong Jae Hwang, Ronak R. Mehta, Vikas Singh
Life of this patent
Estimating the term…
Abstract
A neural net processor provides twin processing paths trainable using different moments of the input data, one moment providing a proxy for uncertainty. Subsequent operation of the trained neural net allows monitoring of the uncertainty proxy to provide real-time assessment of neural net model-based uncertainty.
Claims13
A neural network architecture comprising:
- at least one integrated circuit to implement:
- a first and second input channel;
- a first and second output channel;
- and
- a processing path from the first and second input channels to the first and second output channels through units having weights trained with respect to data having different statistical moments with the different statistical moments input on the first and second input channels, the units providing interaction between data of the first and second input channels and at least one of the units providing a set of functional elements providing a linear transformation of an input to a neuron using weights and offsets followed by a nonlinear transformer preserving a family of a moment distribution of the second input channel;
- wherein the processing path provides an output on the second output channel indicating an uncertainty of results expressed on the first input channel.
3 more elements
A method of processing data with a neural network comprising:
- at least one integrated circuit to implement:
- a first and second input channel;
- a first and second output channel;
- and
- a processing path from the first and second input channels to the first and second output channels through units having weights trained with respect to data having different statistical moments input on the first and second input channels, the units providing interaction between data of the first and second input channels and at least one of the units providing a set of functional elements providing a linear transformation of an input to a neuron using weights and offsets followed by a nonlinear transformer preserving a family of a moment distribution of the second input channel;
- the at least one integrated circuit further implements steps of:
- (a) training the units with a training set having multiple examples providing a series of data by extracting two different moments across the examples aligned with common indices of the series of data and providing the two different moments to the first and second input channels;
- (b) operating the neural network as trained in step (a) by providing a series of data to the first input channel;
- and
- (c) receiving from the second output channel an uncertainty of results expressed on the first input channel.
7 more elements
Dependencies are inferred from the claim text.