When output units must obey hard constraints

Patent metadata

Field	Value
Publication number	US-11521069-B2
Application number	US-201715450933-A
Country	US
Kind code	B2
Filing date	Mar 6, 2017
Priority date	Oct 31, 2016
Publication date	Dec 6, 2022
Grant date	Dec 6, 2022

Abstract

Embodiments employ an inference method for neural networks that enforces deterministic constraints on outputs without performing post-processing or expensive discrete search over the feasible space. Instead, for each input, the continuous weights are nudged until the network's unconstrained inference procedure generates an output that satisfies the constraints. This is achieved by expressing the hard constraints as an optimization problem over the continuous weights and employing backpropagation to change the weights of the network. Embodiments optimize over the energy of the violating outputs; since the weights directly determine the output through the energy, embodiments are able to manipulate the unconstrained inference procedure to produce outputs that conform to global constraints.

First claim

What is claimed is:

1. A method comprising:
training a neural network, during a training phase of the neural network, to produce a trained neural network;
after completing the training phase of the neural network:
performing inference, using the trained neural network comprising a first set of weights, on a particular unlabeled input to produce a first output;
determining whether the first output violates one or more hard constraints on output values of the trained neural network;
in response to determining that the first output violates the one or more hard constraints on output values of the trained neural network, customizing the trained neural network to produce a customized trained neural network, the customized trained neural network being configured to produce output other than the first output from performing inference on the particular unlabeled input, by:
customizing the first set of weights of the trained neural network by using backpropagation to adjust the first set of weights of the trained neural network to produce the customized trained neural network comprising a second set of weights,
wherein said customizing the first set of weights of the trained neural network is configured to lower a negative energy that the trained neural network associates with the first output;

wherein the first set of weights is different than the second set of weights;
performing inference, using the customized trained neural network, on the particular unlabeled input to produce a second output;
wherein the second output is different from the first output; and

wherein the method is performed by one or more computing devices.

2. The method of claim 1 , wherein using backpropagation to adjust the first set of weights of the trained neural network to produce the customized trained neural network is performed during a testing phase of the neural network.

3. The method of claim 1 , further comprising, after performing inference, using the customized trained neural network, on the particular unlabeled input to produce the second output:
determining whether the second output violates the one or more hard constraints on output values of the trained neural network;
in response to determining that the second output violates the one or more hard constraints on output values of the trained neural network, using backpropagation to adjust the second set of weights of the customized trained neural network to produce a second customized trained neural network comprising a third set of weights;
wherein the second customized trained neural network is configured to produce output other than the second output from performing inference on the particular unlabeled input;
wherein the third set of weights is different than both the first set of weights and the second set of weights.

4. The method of claim 1 , further comprising, after performing inference, using the customized trained neural network, on the particular unlabeled input to produce the second output:
determining whether the second output violates the one or more hard constraints;
in response to determining that the second output does not violate the one or more hard constraints on output values of the trained neural network, storing the second output as an inference result for the particular unlabeled input.

5. The method of claim 1 , further comprising:
performing inference, using the trained neural network comprising the first set of weights, on a second unlabeled input to produce a third output;
determining whether the third output violates the one or more hard constraints on output values of the trained neural network;
in response to determining that the third output violates the one or more hard constraints on output values of the trained neural network, customizing the trained neural network to produce a second customized trained neural network, the second customized trained neural network being configured to produce output other than the third output from performing inference on the second unlabeled input, by:
customizing the first set of weights of the trained neural network by using backpropagation to adjust the first set of weights of the trained neural network to produce the second customized trained neural network comprising a third set of weights,
wherein said customizing the first set of weights of the trained neural network is configured to lower another negative energy that the trained neural network associates with the third output;

wherein the third set of weights is different than both the first set of weights and the second set of weights;
performing inference, using the second customized trained neural network, on the second unlabeled input to produce a fourth output;
wherein the fourth output is different from the third output.

6. The method of claim 1 , wherein determining whether the first output violates the one or more hard constraints on output values of the trained neural network is based on a loss function that encodes the one or more hard constraints.

7. The method of claim 6 , further comprising:
receiving a definition of the one or more hard constraints; and
automatically formulating the loss function based on the definition of the one or more hard constraints.

8. The method of claim 7 , wherein the definition of the one or more hard constraints is defined using one or more of:
context-free language; or
regular language.

9. The method of claim 1 , wherein using backpropagation to adjust the first set of weights of the trained neural network to produce the customized trained neural network further comprises:
applying stochastic gradient descent to compute a gradient of a negative energy function to determine a set of changes to be applied to the first set of weights; and
applying the set of changes to the first set of weights to produce the second set of weights.

10. The method of claim 1 , wherein using backpropagation to adjust the first set of weights of the trained neural network comprises optimizing over energy of the first output.

11. One or more non-transitory computer-readable media storing one or more sequences of instructions that, when executed by one or more processors, cause:
training a neural network, during a training phase of the neural network, to produce a trained neural network;
after completing the training phase of the neural network:
performing inference, using the trained neural network comprising a first set of weights, on a particular unlabeled input to produce a first output;
determining whether the first output violates one or more hard constraints on output values of the trained neural network;
in response to determining that the first output violates the one or more hard constraints on output values of the trained neural network, customizing the trained neural network to produce a customized trained neural network, the customized trained neural network being configured to produce output other than the first output from performing inference on the particular unlabeled input, by:
customizing the first set of weights of the trained neural network by using backpropagation to adjust the first set of weights of the trained neural network to produce the customized trained neural network comprising a second set of weights,
wherein said customizing the first set of weights of the trained neural network is configured to lower a negative energy that the trained neural network associates with the first output;

wherein the first set of weights is different than the second set of weights;
performing inference, using the customized trained neural network, on the particular unlabeled input to produce a second output;
wherein the second output is different from the

Assignees

• Oracle Int Corp

Inventors

• Wick Michael
• Tristan Jean-Baptiste
• Lee Jay Yoon

Classifications

• G06N5/01

  Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/084 Primary

  Backpropagation, e.g. using gradient descent · CPC title

• G06N5/003

  Physics · mapped topic

• G06N3/0445

  Physics · mapped topic