# US 2025/0165770 A1 · System, network and method for selective activation of a computing network involving super-imposable stochastic graphs

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

- Number: US 2025/0165770 A1
- Document ID: US20250165770A1
- Status: Published application
- Kind: A1
- Application number: 18/514,467
- Filed: 2023-11-20
- Published: 2025-05-22
- Earliest claimed: 2023-11-20
- Assignees: Silvretta Research, Inc.
- Applicants: Silvretta Research, Inc.
- Inventors: Giuseppe G. Nuti
- Web page: https://patentgrep.com/patent/US20250165770A1

## CPC

- G06N 3/08: Learning methods
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Embodiments of the present disclosure implement a stochastic neural network (SNN) where nodes are selectively activated depending on the inputs and which can be trained on multiple objectives. A system and/or network can include one or more nodes and one or more synapses, wherein each synapse connects a respective pair of the plurality of nodes. The system and/or network can further include one or more processing elements, wherein each of the processing elements is embedded in a respective synapse, and wherein each of the processing elements is adapted to receive an input and generate an output based on the input. In various embodiments, a super-imposable stochastic graph is employed with training, regularization and load balancing.

## Claims

### Claim 1

A method of operating a stochastic neural network, comprising:
providing a stochastic neural network comprising a plurality of nodes, a plurality of synapses and a plurality of processing elements, wherein each of the plurality of synapses comprises a connection between a respective pair of the plurality of nodes, and wherein each of the plurality of processing elements is embedded in a respective synapse of the plurality of synapses, wherein each of the plurality of processing elements is adapted to receive an input and generate an output based on the input, wherein the output of a first synapse of the one or more of the plurality of synapses feeds into a first subset of the plurality of nodes and comprises a first synaptic value computed according to first output of a first node, wherein the output of a second synapse of the one or more of the plurality of synapses comprises a second synaptic value computed according to a second output of the first node, wherein the second output of the first node is distinct from the first output of the first node, wherein, upon receipt of a first problem input, at least a portion of the stochastic neural network is selectively activated by one or more of the plurality of processing elements embedded in one or more of the plurality of synapses; and
balancing a computational load of the stochastic neural network based on enhancing diverse activation of the plurality of nodes.

### Claim 2 (depends on 1)

The method of claim 1, wherein the diverse activation of the plurality of nodes comprises minimizing auto-covariance of outputs associated with each of the plurality of synapses.

### Claim 3 (depends on 1)

The method of claim 1, wherein the diverse activation of the plurality of nodes comprises targeting a firing rate of a first node of the first subset of the plurality of nodes.

### Claim 4 (depends on 1)

The method of claim 1, wherein the diverse activation of the plurality of nodes comprises determining a resource usage value for each of the plurality of synapses and where the resource usage value is a function of one or more recent activations or one or more activation attempts.

### Claim 5 (depends on 1)

The method of claim 1, wherein the diverse activation of the plurality of nodes comprises regularizing the stochastic neural network based on an action affecting at least one of the plurality of nodes.

### Claim 6

A method of operating a stochastic neural network, comprising:
providing a stochastic neural network comprising a plurality of nodes, a plurality of synapses and a plurality of processing elements, wherein each of the plurality of synapses comprises a connection between a respective pair of the plurality of nodes, and wherein each of the plurality of processing elements is embedded in a respective synapse of the plurality of synapses, wherein each of the plurality of processing elements is adapted to receive an input and generate an output based on the input, wherein the output of a first synapse of the one or more of the plurality of synapses feeds into a first subset of the plurality of nodes and comprises a first synaptic value computed according to first output of a first node, wherein the output of a second synapse of the one or more of the plurality of synapses comprises a second synaptic value computed according to a second output of the first node, wherein the second output of the first node is distinct from the first output of the first node, wherein the first subset of the plurality of nodes comprises a first subset of activated nodes, wherein, upon receipt of a first problem input, at least a portion of the stochastic neural network is selectively activated by one or more of the plurality of processing elements embedded in one or more of the plurality of synapses; and
training the stochastic neural network.

### Claim 7 (depends on 6)

The method of claim 6, wherein the training comprises updating the synaptic value of the first synapse and updating the first subset of activated nodes.

### Claim 8 (depends on 6)

The method of claim 6, wherein the training comprises back propagating from each of the first subset of the plurality of nodes to one or more input nodes along a path.

### Claim 9 (depends on 8)

The method of claim 8, wherein the path comprises one or more paths that activated the first subset of the plurality of nodes.

### Claim 10 (depends on 6)

The method of claim 6, wherein the training comprises training a subset of computed but not activated synapses from the plurality of synapses.

### Claim 11 (depends on 6)

The method of claim 6, further comprising exploring the stochastic neural network.

### Claim 12 (depends on 11)

The method of claim 11, wherein the first synaptic value is further computed according to a path preceding the first synapse, wherein the path comprises a plurality of paths and wherein exploring the stochastic neural network comprises determining a likelihood of exploration based on a count of paths that have been initiated from the plurality of paths.

### Claim 13 (depends on 11)

The method of claim 11, wherein the first synaptic value is further computed according to a path preceding the first synapse, wherein the path comprises a plurality of paths and wherein exploring the stochastic neural network comprises exploring at least one less utilized path of the plurality of paths.

### Claim 14 (depends on 11)

The method of claim 11, wherein the first synaptic value is further computed according to a path preceding the first synapse, wherein the path comprises a plurality of paths and wherein exploring the stochastic neural network comprises promoting a connection that has a likely positive relationship with the output.

### Claim 15

A method of operating a stochastic neural network, comprising:
providing a stochastic neural network comprising a plurality of nodes, a plurality of synapses and a plurality of processing elements, wherein each of the plurality of synapses comprises a connection between a respective pair of the plurality of nodes, and wherein each of the plurality of processing elements is embedded in a respective synapse of the plurality of synapses, wherein each of the plurality of processing elements is adapted to receive an input and generate an output based on the input, wherein the output of a first synapse of the one or more of the plurality of synapses feeds into a first subset of the plurality of nodes and comprises a first synaptic value computed according to first output of a first node, wherein the output of a second synapse of the one or more of the plurality of synapses comprises a second synaptic value computed according to a second output of the first node, wherein the second output of the first node is distinct from the first output of the first node, wherein, upon receipt of a first problem input, at least a portion of the stochastic neural network is selectively activated by one or more of the plurality of processing elements embedded in one or more of the plurality of synapses; and
regularizing the stochastic neural network.

### Claim 16 (depends on 15)

The method of claim 15, wherein regularizing the stochastic neural network is based upon co-activation of a node.

### Claim 17 (depends on 15)

The method of claim 15, wherein regularizing the stochastic neural network is based upon activation of a node dependent upon a total number of activated nodes.

### Claim 18 (depends on 15)

The method of claim 15, wherein regularizing the stochastic neural network comprises maximizing entropy at each node or at each synapse in the stochastic neural network with an adjustment for activation bias.

### Claim 19 (depends on 15)

The method of claim 15, wherein the action comprises enhanced regularization of the first subset of the plurality of nodes, wherein the first subset has been recently activated.

### Claim 20 (depends on 15)

The method of claim 15, wherein the action comprises the difference between a number of initiated paths and a number of paths that have activated one or more objective nodes.

### Claim 21 (depends on 15)

The method of claim 15, wherein the action comprises the difference between a total input probability of incoming synapses from the plurality of synapses to a first node of the plurality of nodes and a total output probability of outgoing synapses from the plurality of synapses of the first node.

### Claim 22 (depends on 15)

The method of claim 15, wherein the action comprises a temporal diversity of output synapses of the plurality of synapses viewed as a vector.

### Claim 23

A computing system, comprising:
a stochastic neural network comprising:
a plurality of nodes;
a plurality of synapses, wherein each synapse of the plurality of synapses connects a respective pair of the plurality of nodes, wherein each synapse of the plurality of synapses comprises an activation weight, wherein each of the plurality of synapses is operable to control activation of at least one of the plurality of nodes, and wherein a subset of the plurality of synapses shares a first activation weight, wherein the output of a first synapse of the plurality of synapses feeds into a first subset of the plurality of nodes and comprises a first synaptic value computed according to a-first output of a first node, wherein the output of a second synapse of the plurality of synapses comprises a second synaptic value computed according to a second output of the first node, and wherein the second output of the first node is distinct from the first output of the first node; and
a plurality of processing elements, wherein each of the plurality of processing elements is embedded in a respective synapse of the plurality of synapses, wherein, upon receipt of a first problem input, at least a first portion of the stochastic neural network is selectively activated by one or more of the plurality of processing elements embedded in one or more of the plurality of synapses.

### Claim 24 (depends on 23)

The computing system of claim 23 wherein, upon receipt of a second problem input, at least a second portion of the stochastic neural network is selectively activated by one or more of the plurality of processing elements embedded in one or more of the plurality of synapses, wherein the first portion of the stochastic neural network comprises a first subset of nodes, wherein the second portion of the stochastic neural network comprises a second subset of nodes, and wherein the first subset of nodes is different from the second subset of nodes.

### Claim 25 (depends on 24)

The computing system of claim 24, wherein activation of the first portion of the stochastic neural network is in parallel with the activation of the second portion of the stochastic neural network.
