# US 12,008,478 B2 · Systems and methods for training generative models using summary statistics and other constraints

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 12,008,478 B2
- Document ID: US12008478B2
- Status: Granted
- Kind: B2
- Application number: 17/074,364
- Filed: 2020-10-19
- Published: 2021-04-22
- Granted: 2024-06-11
- Earliest claimed: 2019-10-18
- Assignees: Unlearn.AI, Inc.
- Applicants: Unlearn.AI, Inc.
- Inventors: Aaron M. Smith; Anton D. Loukianov; Charles K. Fisher; Jonathan R. Walsh
- Web page: https://patentgrep.com/patent/US12008478B2

## CPC

- G06N 3/088: Non-supervised learning, e.g. competitive learning
- G06F 18/211: Selection of the most significant subset of features
- G06F 18/24133: {Distances to prototypes}
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/045: Combinations of networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/047: Probabilistic or stochastic networks
- G06N 3/0475: Generative networks
- G06N 3/082: modifying the architecture, e.g. adding, deleting or silencing nodes or connections
- G06N 3/094: Adversarial learning
- G06N 7/08: using chaos models or non-linear system models
- G06N 20/00: Machine learning
- G06N 5/046: Forward inferencing; Production systems
- G06N 7/01: Probabilistic graphical models, e.g. probabilistic networks
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

Systems and methods for training and utilizing constrained generative models in accordance with embodiments of the invention are illustrated. One embodiment includes a method for training a constrained generative model. The method includes steps for receiving a set of data samples from a first distribution, identifying a set of constraints from a second distribution, and training a generative model based on the set of data samples and the set of constraints.

## Claims

### Claim 1

A method for training a constrained generative model, the method comprising:
receiving a set of data samples from a first distribution;
identifying a set of constraints from a second distribution; and
training a generative model based on the set of data samples and the set of constraints, wherein:
the generative model is a neural network comprising hidden units; and
training the generative model comprises:
training the generative model on the set of data samples;
updating the generative model to add a set of new hidden units and connections to the generative model; and
training the generative model based on the set of constraints by modifying weights for the set of new hidden units.

### Claim 2 (depends on 1)

The method of claim 1, wherein the set of constraints comprises a set of tolerances for the constraints, wherein training the generative model comprises training the generative model until a difference between each constraint of the set of constraints falls within a corresponding tolerance of the set of tolerances.

### Claim 3 (depends on 1)

The method of claim 1, wherein each constraint of the set of constraints is weighted to adjust for relative importance of the constraint.

### Claim 4 (depends on 1)

The method of claim 1, wherein training the generative model further comprises:
training a first model on the set of data samples; and
perturbing the first model using the set of constraints to result in the trained generative model.

### Claim 5 (depends on 4)

The method of claim 4, wherein perturbing the first model comprises minimizing a loss function that balances the set of constraints with a difference between the generative model and the perturbed model.

### Claim 6 (depends on 1)

The method of claim 1, wherein training the generative model further comprises training a model using the set of data samples and a loss function that incorporates the set of constraints as a penalty term.

### Claim 7 (depends on 1)

The method of claim 1, wherein:
the set of constraints comprises a set of functionals that map probability densities to real numbers; and
training the generative model comprises enforcing that the set of functionals attains particular values.

### Claim 8 (depends on 1)

The method of claim 1, wherein:
the set of constraints comprises at least one composite constraint; and
the at least one composite constraint is a function of a plurality of variables.

### Claim 9 (depends on 1)

The method of claim 1, wherein the generative model is at least one of the group consisting of a recurrent neural network, a Boltzmann machine, a restricted Boltzmann machine, a generative adversarial network, and a variational autoencoder.

### Claim 10 (depends on 1)

The method of claim 1, wherein modifying weights is only performed for the set of new hidden units.

### Claim 11 (depends on 1)

The method of claim 1, wherein training the generative model further comprises:
training a surrogate model to score generative models as a function of weights of the generative models; and
training the generative model based on the set of data samples, the set of constraints, and the surrogate model.

### Claim 12 (depends on 1)

The method of claim 1, further comprising generating, using the trained generative model, a set of samples based on the second distribution.

### Claim 13 (depends on 12)

The method of claim 12, wherein:
the set of data samples are from a clinical trial of a treatment for a first population segment;
the set of constraints comprises a set of summary statistics of a different second population segment; and
the generated set of samples predicts effects of the treatment for the different second population segment.

### Claim 14

A non-transitory machine-readable medium containing processor instructions for training a constrained generative model, where execution of the instructions by a processor causes the processor to perform a process that comprises:
receiving a set of data samples from a first distribution;
identifying a set of constraints from a second distribution; and
training a generative model based on the set of data samples and the set of constraints, wherein:
the generative model is a neural network comprising hidden units; and
training the generative model comprises:
training the generative model on the set of data samples;
updating the generative model to add a set of new hidden units and connections to the generative model; and
training the generative model based on the set of constraints by modifying weights for the set of new hidden units.

### Claim 15 (depends on 14)

The non-transitory machine-readable medium of claim 14, wherein the set of constraints comprises a set of tolerances for the set of constraints, wherein training the generative model comprises training the generative model until a difference between each constraint of the set of constraints falls within a corresponding tolerance of the set of tolerances.

### Claim 16 (depends on 14)

The non-transitory machine-readable medium of claim 14, wherein training the generative model comprises:
training a first model on the set of data samples; and
perturbing the first model using the set of constraints to result in the trained generative model.

### Claim 17 (depends on 16)

The non-transitory machine-readable medium of claim 16, wherein perturbing the first model comprises minimizing a loss function that balances the set of constraints with a difference between the first generative model and the perturbed model.

### Claim 18 (depends on 14)

The non-transitory machine-readable medium of claim 14, wherein training the generative model comprises training a model using the set of data samples and a loss function that incorporates the set of constraints as a penalty term.

### Claim 19 (depends on 14)

The non-transitory machine-readable medium of claim 14, further comprising generating, using the trained generative model, a set of samples based on the second distribution.

### Claim 20 (depends on 19)

The non-transitory machine-readable medium of claim 19, wherein:
the set of data samples are from a clinical trial of a treatment for a first population segment;
the set of constraints comprises a set of summary statistics of a different second population segment, and
the generated set of samples predicts effects of the treatment for the different second population segment.