Systems and methods for generating data explanations for neural networks and related systems

• US 10,275,705 B2
• Filed: 08/10/2015
• Issued: 04/30/2019
• Est. Priority Date: 08/08/2014
• Status: Active Grant

• Current Assignee: Intrinsic Innovation LLC (Alphabet Inc.)
• Original Assignee: Vicarious FPC, Inc. (Alphabet Inc.)
• Inventors: George, Dileep, Kansky, Kenneth Alan, Laan, Christopher Remmert, Lehrach, Wolfang, Marthi, Bhaskara Mannar, Phoenix, David Scott, Purdy, Eric
• Primary Examiner(s): Misir, Dave
• Application Number:
  US14/822,730
• Publication Number:
  US 20160292567A1
• Time in Patent Office: 1,359 Days
• Field of Search: 706 20
• CPC Class Codes: G06F 18/24137 Distances to cluster centroïds
• CPC Class Codes: G06F 18/24155 Bayesian classification
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/047 Probabilistic or stochastic networks
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects
• CPC Class Codes: G06V 10/82 using neural networks

Abstract

A method for generating data explanations in a recursive cortical network includes receiving a set of evidence data at child feature nodes of a first layer of the recursive cortical network, setting a transformation configuration that directs messaging of evidence data and transformed data between layers of the network, performing a series of transformations on the evidence data according to the transformation configuration, the series including at least one forward transformation and at least one reverse transformation, and outputting the transformed evidence data.

Claims

1. A method for generating descriptions of image data in a bi-directional layer-based network comprising:

receiving a set of image data having detectable features;

wherein the image data is captured by a camera;

setting a transformation configuration that directs messaging of image data and transformed data between layers of the network;

at a first layer of the network, performing a first forward transformation of the set of image data into a first set of transformed data;

at a second layer of the network, performing a second forward transformation of the first set of transformed data into a second set of transformed data;

at the second layer of the network, performing a first reverse transformation of the second set of transformed data into a third set of transformed data;

at the second layer of the network, performing a third forward transformation of the third set of transformed data into a fourth set of transformed data; and

outputting the fourth set of transformed data;

wherein outputting the fourth set of transformed data comprises providing descriptions of the detectable features of the set of image data;

wherein performing the second forward transformation comprises;

receiving the first set of transformed data at one or more inputs of the second layer;

calculating a first posterior probability distribution based on the first set of transformed data and a set of prior probabilities and likelihood relationships encoded in the second layer; and

generating the second set of transformed data from the first posterior probability distribution at an output of the second layer;

wherein performing the first reverse transformation comprises;

receiving the second set of transformed data at an output of the second layer;

calculating a set of updated likelihoods based on the second set of transformed data and the set of prior probabilities and likelihood relationships encoded in the second layer; and

generating the third set of transformed data from the set of updated likelihoods at one or more inputs of the second layer.

2. The method of claim 1, wherein the bi-directional layer-based network is a recursive cortical network.

3. The method of claim 1, wherein performing the third forward transformation comprises:

receiving the third set of transformed data at one or more inputs of the second layer;

calculating a second posterior probability distribution based on the first set of transformed data, the set of updated likelihoods of the third set of transformed data, and the set of prior probabilities and likelihood relationships encoded in the second layer; and

generating the fourth set of transformed data from the second posterior probability distribution at an output of the second layer.

4. The method of claim 3, wherein the bi-directional layer-based network is a recursive cortical network.

5. The method of claim 4, wherein the recursive cortical network is implemented by a distributed computing system.

6. The method of claim 4, wherein the recursive cortical network comprises:

A recursively architected network of sub-networks organized into a plurality of hierarchical layers;

the sub-networks comprising at least a parent feature node, a pool node, a parent-specific child feature (PSCF) node, and a child feature node;

the parent feature node of at least one sub-network configured with a selection function actionable on at least two pool nodes connected to the parent feature node of the at least one sub-network;

the pool node of the at least one sub-network configured with a selection function actionable on at least two PSCF nodes connected to the pool node of the at least one sub-network;

the PSCF node of the at least one sub-network configured to activate a connected child feature node;

the child feature node connectable to at least a parent feature node of a second sub-network at a lower hierarchical layer; and

a constraint node with at least two connections from the at least two PSCF nodes, with a selection function to augment selection by the pool node.

7. The method of claim 1, wherein the image data comprises image data processed by an edge detection filter.

8. The method of claim 3, further comprising performing additional forward and reverse transformations based on a static transformation configuration.

9. The method of claim 3, further comprising performing additional forward and reverse transformations based on a dynamic transform configuration;

the dynamic transform configuration directing messaging based on layer output probability thresholds.

10. The method of claim 3, further comprising performing additional forward and reverse transformations based on a dynamic transform configuration;

the dynamic transform configuration directing messaging based on recursion level thresholds.

11. The method of claim 3, further comprising performing forward transformations based on lateral constraints encoded in the network.

12. The method of claim 11, wherein performing forward transformations based on lateral constraints comprises enforcing an activation constraint between at least two nodes of the network.

13. The method of claim 3, wherein outputting the fourth set of transformed data comprises outputting a set of data classifiers.