# US 2014/0279777 A1 · Signal processing systems

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

- Number: US 2014/0279777 A1
- Document ID: US20140279777A1
- Status: Published application
- Kind: A1
- Application number: 13/925,637
- Filed: 2013-06-24
- Published: 2014-09-18
- Earliest claimed: 2013-03-15
- Assignees: Google Inc.
- Applicants: Google Inc.
- Inventors: Julien Robert Michel Cornebise; Danilo Jimenez Rezende; Daniël Pieter Wierstra
- Web page: https://patentgrep.com/patent/US20140279777A1

## CPC

- G06N 3/045: Combinations of networks
- G06N 3/0475: Generative networks
- G06N 3/0495: Quantised networks; Sparse networks; Compressed networks
- G06N 3/0499: Feedforward networks
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/0895: Weakly supervised learning, e.g. semi-supervised or self-supervised learning
- G06N 3/09: Supervised learning
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

We describe a signal processor, the signal processor comprising: a probability vector generation system, wherein said probability vector generation system has an input to receive a category vector for a category of output example and an output to provide a probability vector for said category of output example, wherein said output example comprises a set of data points, and wherein said probability vector defines a probability of each of said set of data points for said category of output example; a memory storing a plurality of said category vectors, one for each of a plurality of said categories of output example; and a stochastic selector to select a said stored category of output example for presentation of the corresponding category vector to said probability vector generation system; wherein said signal processor is configured to output data for an output example corresponding to said selected stored category.

## Claims

### Claim 1

A signal processor, the signal processor comprising:
a probability vector generation system, wherein said probability vector generation system has an input to receive a category vector for a category of output example and an output to provide a probability vector for said category of output example, wherein said output example comprises a set of data points, and wherein said probability vector defines a probability of each of said set of data points for said category of output example;
a memory storing a plurality of said category vectors, one for each of a plurality of said categories of output example; and
a stochastic selector to select a said stored category of output example for presentation of the corresponding category vector to said probability vector generation system;
wherein said signal processor is configured to output data for an output example corresponding to said selected stored category.

### Claim 2 (depends on 1)

A signal processor as claimed in claim 1 further comprising a context vector input to receive a context vector, wherein said context vector defines a relative likelihood of each of said plurality of said categories; wherein said context vector input is coupled to said stochastic selector such that said selection of said category of output example is dependent on said context vector.

### Claim 3 (depends on 2)

A signal processor as claimed in claim 2 wherein said probability vector generation system is coupled to said context vector input, and wherein said probability vector is dependent upon said context vector.

### Claim 4 (depends on 2)

A signal processor as claimed in claim 2 further comprising a mapping unit coupled between said context vector input and said stochastic selector, wherein said context vector has length dc, wherein K is a number of said categories, and wherein said mapping unit is configured to map said context vector of length dc to a category probability vector of length K, and wherein said stochastic selector is configured to select said stored category of output example dependent on said category probability vector.

### Claim 5 (depends on 4)

A signal processor as claimed in claim 4 wherein said mapping unit comprises a deterministic neural network.

### Claim 6 (depends on 1)

A signal processor as claimed in claim 1 wherein said category vector comprises a compressed representation of said probability vector and wherein said probability vector generation system comprises a deterministic neural network.

### Claim 7 (depends on 6)

A signal processor as claimed in claim 6 further comprising a context vector input to receive a context vector, wherein said context vector defines a relative likelihood of each of said plurality of said categories; wherein said context vector input is coupled to said stochastic selector such that said selection of said category of output example is dependent on said context vector; and wherein said deterministic neural network has a first input comprising said category vector and a second input dependent on said context vector, and an output to provide said probability vector.

### Claim 8 (depends on 1)

A signal processing chain comprising a chain of signal processors each as claimed in claim 1, each signal processor after a first signal processor further comprising a context vector input to receive a context vector, wherein said context vector defines a relative likelihood of each of said plurality of said categories; wherein said context vector input is coupled to said stochastic selector such that said selection of said category of output example is dependent on said context vector; and wherein said output data from one said signal processor provides at least part of said context vector for a next said signal processor in the chain.

### Claim 9 (depends on 8)

A signal processing chain as claimed in claim 8 wherein, for each successive said signal processor after said first signal processor, said output data from the preceding signal processors in the chain combine to provide said context vector input for the successive signal processor.

### Claim 10 (depends on 8)

A signal processing chain as claimed in claim 8 wherein said output data from a said signal processor comprises said category vector selected by the stochastic selector of the signal processor.

### Claim 11 (depends on 8)

A signal processing chain as claimed in claim 8, wherein said output data from a last said signal processor of the chain comprises said probability vector.

### Claim 12 (depends on 11)

A signal processing chain as claimed in claim 11 further comprising an output stochastic selector having an input coupled to receive said probability vector from said last signal processor of said chain and configured to generate and output a said output example comprising a said set of data points having values selected with probabilities defined by said probability vector.

### Claim 13 (depends on 5)

A signal processor as claimed in claim 5, wherein said category vector comprises a compressed representation of said probability vector, wherein said probability vector generation system comprises a deterministic neural network; and further comprising a training module coupled to said memory, to said probability vector generation system, to said context vector input, and to said context vector mapping unit, and having a training data input to receive training examples, wherein said training module is configured to compute a responsibility value for each said category, dependent on said a training example presented at said training data input and a context vector presented at said context vector input, and to adjust said stored category vectors, weights of said neural network of said probability vector generation system and weights of said neural network of said mapping unit responsive said computed responsibility values.

### Claim 14 (depends on 1)

A signal processor as recited in claim 1 wherein said stored category vectors, and said probability vectors from said probability vector generation system which depend on said category vectors, comprise a learnt representation of real-world data; and/or embodied as a data generation classification, completion or search system.

### Claim 15

A signal processing system for generating output examples from categories of a plurality of categories, wherein a distribution of training examples across said plurality of categories has been learnt by said signal processing system, the signal processing system comprising:
a chain of signal processors, wherein each signal processor of the chain has learnt a distribution of said training examples across a limited number of categories less than said plurality of categories;
wherein at least each said signal processor after a first said signal processor in the chain has a context input and is configured to generate an output example from said learnt distribution conditional on said context input;
wherein each successive signal processor in said chain receives the output example from the preceding processor in the chain as said context input;
wherein a first said input processor in said chain is configured to stochastically select a said output example according to its learnt distribution; and
wherein a last said signal processor in said chain is configured to provide one or both of an output example and a probability distribution for stochastically selecting a said output example.

### Claim 16 (depends on 15)

A signal processing system as claimed in claim 15 wherein each successive signal processor in said chain receives the output examples from all the preceding processors in the chain.

### Claim 17 (depends on 15)

A signal processing system as claimed in claim 15 wherein said plurality of categories of said chain is defined by a multiplication product of said limited number of categories of each signal processor of said chain.

### Claim 18 (depends on 15)

A signal processing system as claimed in claim 15, wherein each said signal processor comprises a data compression system to represent said output example in a compressed data format, and wherein each successive signal processor in said chain receives said output example from the preceding processor in said compressed data format.

### Claim 19 (depends on 15)

A computer system programmed to implement the signal processing system of claim 15.

### Claim 20 (depends on 15)

Electronic hardware configured to implement the signal processing system of claim 15.

### Claim 21 (depends on 15)

A non-transitory data carrier carrying processor control code to implement the signal processor of claim 15.

### Claim 22

A method of signal processing to generate data for an output example from a plurality of learnt categories of training examples, the method comprising:
storing a plurality of category vectors each defining a learnt category of training example;
stochastically selecting a stored said category vector;
generating a probability vector, dependent upon said selected category vector; and
outputting data for said output example, wherein said output example comprises a set of data points each having a probability defined by a respective component of said probability vector.

### Claim 23 (depends on 22)

A method as claimed in claim 22 further comprising providing said selected category vector to a probability vector generation system to generate said probability vector, wherein said probability vector generation system comprises a data decompression system; and
decompressing said selected category vector using said probability vector generation system to generate said probability vector.

### Claim 24 (depends on 23)

A method as claimed in claim 23 further comprising inputting a context vector, wherein said context vector defines a relative likelihood of each of said plurality of said categories; and wherein said selecting of said stored category vector is dependent upon said likelihood of said categories defined by said context vector.

### Claim 25 (depends on 24)

A method as claimed in claim 24 wherein said context vector has length dc, and wherein K is a number of said categories, where dc is different to K, the method further comprising mapping said context vector of length dc to a category probability vector of length K.

### Claim 26 (depends on 25)

A method as claimed in claim 25 wherein said mapping is performed by a first neural network and said decompressing is performed by a second neural network.

### Claim 27 (depends on 26)

A method as claimed in claim 26 further comprising inputting said context vector into said second neural network.

### Claim 28 (depends on 24)

A method as claimed in claim 24, further comprising chaining said signal processing, wherein said claiming comprises repeating said selecting of said stored category vector in a succession of signal processing stages, the method further comprising using said selected category vector from one of said signal processing stages in the context vector input to the next said signal processing stage.

### Claim 29 (depends on 15)

A method of training a signal processing system as recited in claim 15, the method comprising:
presenting training examples to the signal processing system, wherein a said training example comprises a set of data points corresponding to data points of a said output example;
computing from a said training example a set of responsibility values, one for each said category, wherein a said responsibility value comprises a probability of the training example belonging to the category, each category having a respective stored category vector;
computing a gradient vector for a set of parameters of the signal processor, system or method from said set of responsibility values, wherein said set of parameters includes said stored category vectors and defines a shared mapping between said stored category vectors and a corresponding set of said probability vectors defining probability values for a said set of data points of a said output example; and
updating said set of parameters using said computed gradient vector.

### Claim 30 (depends on 29)

A method as claimed in claim 29, further comprising presenting a said context vector with said training example, and wherein a said responsibility value is further conditional on said context vector.

### Claim 31 (depends on 30)

A method as claimed in claim 30 used for training a chained said signal processor, the method further comprising: stochastically selecting a said stored category vector dependent on said set of responsibility values in one said chained signal processor, and using said stochastically selected category vector as at least part of said context vector presented to a next said chained signal processor.

### Claim 32 (depends on 22)

A non-transitory data carrier carrying processor control code to implement the method of claim 22.

### Claim 33

A signal processor to generate data for an output example from a plurality of learnt categories of training examples, the signal processor comprising:
a memory for storing a plurality of category vectors each defining a learnt category of training example;
a system to stochastically select a stored said category vector;
a system to generate a probability vector dependent on said selected category vector; and
an output to provide an output example which comprises a set of data prints each having a probability defined by a respective component of said probability vector.

### Claim 34 (depends on 33)

A signal processor as claimed in claim 33 wherein said data comprises one or more of: image data, sound data, signal data, sensor data, actuator data, spatial data, and text data.

### Claim 35

A neural network architecture, the architecture comprising:
a first, input layer of stochastic nodes;
a second, output layer of stochastic nodes; and
a deterministic neural network connected between said input and output layer nodes.

### Claim 36 (depends on 35)

A neural network architecture as claimed in claim 35 wherein a said stochastic node has one or more inputs and one or more outputs, and wherein an output value from a said output has a value dependent on a probability distribution determined by input values on said one or more inputs.

### Claim 37 (depends on 35)

A neural network architecture as claimed in claim 35 wherein said deterministic second network comprises a multilayer perceptron having at least one layer with a greater number of nodes than a number of stochastic nodes in said input layer.

### Claim 38 (depends on 35)

A modified Helmholtz machine comprising the neural network architecture of claim 35.

### Claim 39 (depends on 35)

The neural network architecture of claim 35 embodied in: software on a non-transitory data courier, a programmed computer system, programmed memory, or electronic hardware/circuitry.
