UNIVERSAL, ONLINE LEARNING IN MULTI-MODAL PERCEPTION-ACTION SEMILATTICES

• US 20150262056A1
• Filed: 08/24/2012
• Published: 09/17/2015
• Est. Priority Date: 12/14/2011
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: MODHA, Dharmendra S.
• Granted Patent: US 9,697,461 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: G06N 3/10 Interfaces, programming languages or software development kits, e.g. for simulating neural networks

Abstract

In one embodiment, the present invention provides a method for interconnecting neurons in a neural network. At least one node among a first set of nodes is interconnected to at least one node among a second set of nodes, and nodes of the first and second set are arranged in a lattice. At least one node of the first set represents a sensory-motor modality of the neural network. At least one node of the second set is a union of at least two nodes of the first set. Each node in the lattice has an acyclic digraph comprising multiple vertices and directed edges. Each vertex represents a neuron population. Each directed edge comprises multiple synaptic connections. Vertices in different acyclic digraphs are interconnected using an acyclic bottom-up digraph. The bottom-up digraph has a corresponding acyclic top-down digraph. Vertices in the bottom-up digraph are interconnected to vertices in the top-down digraph.

Claims

1. A method for interconnecting neurons in a neural network, the method comprising:

interconnecting at least one node among a first set of nodes to at least one node among a second set of nodes; and

arranging nodes of the first and second set in a lattice;

wherein at least one node of the first set represents a sensory-motor modality of the neural network, and at least one node of the second set is a union of at least two nodes of the first set.

2. The method of claim 1, wherein:

each node in the lattice has a corresponding acyclic digraph; and

each acyclic digraph comprises;

multiple vertices; and

multiple directed edges interconnecting the vertices in said acyclic digraph.

3. The method of claim 2, wherein some acyclic digraphs are empty.

4. The method of claim 3, further comprising:

interconnecting vertices in different acyclic digraphs using an acyclic bottom-up digraph, wherein the bottom-up digraph includes the vertices and directed edges of each acyclic digraph.

5. The method of claim 4, wherein the bottom-up digraph further includes additional directed edges that interconnect vertices in different acyclic digraphs, wherein a sending vertex in a first acyclic digraph is interconnected to a receiving vertex in a second acyclic digraph only if the node corresponding to the second acyclic digraph is a superset of the node corresponding to the first acyclic digraph.

6. The method of claim 5, wherein the bottom-up digraph has a corresponding acyclic top-down digraph, said top-down digraph comprising multiple vertices and multiple directed edges interconnecting the vertices in said top-down digraph.

7. The method of claim 6, wherein:

each vertex in the top-down digraph corresponds to a vertex in the bottom-up digraph; and

each directed edge in the top-down digraph corresponds to a directed edge in the bottom-up digraph, such that information flows along said directed edge in the top-down digraph in a first direction, and information flows along said corresponding directed edge in the bottom-up digraph in a direction opposite of the first direction.

8. The method of claim 7, further comprising:

interconnecting vertices in the bottom-up digraph to vertices in the top-down digraph using additional directed edges.

9. The method of claim 8, wherein a sending vertex in the bottom-up digraph is interconnected to a receiving vertex in the top-down digraph if the intersection of the node corresponding to the sending vertex and the node corresponding to the receiving vertex is empty.

10. The method of claim 8, wherein a sending vertex in the bottom-up digraph is interconnected to a receiving vertex in the top-down digraph if the sending vertex and the receiving vertex correspond to one another.

11. The method of claim 8, further comprising:

identifying source vertices in the bottom-up digraph and the top-down digraph, wherein source vertices are vertices that do not have incoming directed edges; and

identifying sink vertices in the bottom-up digraph and the top-down digraph, wherein sink vertices are vertices that do have outgoing directed edges.

12. The method of claim 11, further comprising:

for each source vertex in the bottom-up digraph, configuring said source vertex in the bottom-up digraph to receive sensory input;

for each sink vertex in the bottom-up digraph, interconnecting said sink vertex in the bottom-up graph to a vertex in the top-down digraph using at least one outgoing directed edge from said sink vertex;

for each source vertex in the top-down digraph, interconnecting a vertex in the bottom-up graph to said source vertex in the top-down graph using at least one incoming directed edge to said source vertex; and

for each sink vertex in the top-down digraph, configuring said sink vertex in the top-down graph to generate motor output.

13. The method of claim 8, further comprising:

duplicating the bottom-up digraph into a first duplicated bottom-up digraph and a second duplicated bottom-up digraph;

duplicating the top-down digraph into a first duplicated top-down digraph and a second duplicated top-down digraph;

identifying each vertex in the first duplicated bottom-up digraph and the first duplicated top-down digraph as a good vertex; and

identifying each vertex in the second duplicated bottom-up digraph and the second duplicated top-down digraph as a bad vertex;

wherein each good vertex in the first duplicated bottom-up digraph corresponds to a bad vertex in the second duplicated bottom-up digraph, a good vertex in the first duplicated top-down digraph, and a bad vertex in the second duplicated top-down digraph.

14. The method of claim 13, further comprising:

identifying good source vertices and bad source vertices in the first and second duplicated bottom-up digraphs, respectively, wherein source vertices are vertices that do not have incoming directed edges;

identifying good sink vertices and bad sink vertices in the first and second duplicated top-down digraphs, respectively, wherein sink vertices are vertices that do have outgoing directed edges; and

for each good sink vertex in the first duplicated top-down graph;

connecting said good sink vertex to an evaluation module, wherein the evaluation module determines whether an output from said good sink vertex is good or bad;

connecting good output from said good sink vertex to a corresponding good source vertex in the first duplicated bottom-up graph; and

connecting bad output from said good sink vertex to a corresponding bad source vertex in the second duplicated bottom-up graph.

15. The method of claim 8, wherein each vertex represents a neuron population comprising multiple neurons.

16. The method of claim 15, wherein each directed edge interconnecting a sending vertex to a receiving vertex comprises multiple synaptic connections such that a neuron in a neuron population corresponding to the receiving vertex receives synaptic connections from a subset of neurons in a neuron population corresponding to the second vertex.

17. A method for interconnecting neural nodes, the method comprising:

interconnecting a plurality of first nodes in a first set of nodes with a plurality of second nodes in a second set of nodes, the connected nodes arranged in an interconnection lattice;

wherein each node represents at least one population of neurons; and

wherein a connected node in the second set is a union of at least two nodes in the first set, such that said connected node exchanges signals with said at least two nodes via the interconnection lattice.

18. The method of claim 17, wherein:

a node comprises one or more neural vertices interconnected via multiple directed edges arranged in an acyclic digraph, wherein each neural vertex comprises one or more neurons, and each edge comprises a signaling pathway in the interconnection lattice.

19. The method of claim 18, wherein interconnecting nodes comprises:

interconnecting said nodes via bottom-up signaling pathways arranged in an acyclic bottom-up digraph in the interconnection lattice, each bottom-up signaling pathway including one or more vertices and directed edges; and

interconnecting said nodes via top-down signaling pathways arranged in an acyclic top-down digraph in the interconnection lattice, each top-down signaling pathway including one or more vertices and directed edges;

wherein a sending vertex of a first node is interconnected to a receiving vertex of a second node only if the second node is a superset of the first node;

wherein each vertex in the top-down digraph corresponds to a vertex in the bottom-up digraph; and

wherein each directed edge in the top-down digraph corresponds to a directed edge in the bottom-up digraph, such that information flows along said directed edge in the top-down digraph in a first direction, and information flows along said corresponding directed edge in the bottom-up digraph in a direction opposite of the first direction.

20. The method of claim 19, further comprising interconnecting vertices in the bottom-up digraph to vertices in the top-down digraph using additional directed edges.

21. The method of claim 20, wherein a sending vertex in the bottom-up digraph is interconnected to a receiving vertex in the top-down digraph if an intersection of the node comprising the sending vertex and the node comprising the receiving vertex is empty.

22. The method of claim 20, wherein a sending vertex in the bottom-up digraph is interconnected to a receiving vertex in the top-down digraph if the sending vertex and the receiving vertex correspond to one another.

23. The method of claim 20, further comprising:

designating vertices that do not have incoming directed edges as source vertices;

designating vertices that do not have outgoing directed edges as sink vertices;

configuring a source vertex in the bottom-up digraph to receive sensory input;

configuring a sink vertex in the top-down digraph to generate motor output;

for each sink vertex in the bottom-up digraph, interconnecting said sink vertex in the bottom-up digraph to a vertex in the top-down digraph using at least one outgoing directed edge from said sink vertex; and

for each source vertex in the top-down digraph, interconnecting a vertex in the bottom-up digraph to said source vertex in the top-down digraph using at least one incoming directed edge to said source vertex.

24. The method of claim 20, further comprising:

for each vertex, designating a first set and a second set of neurons in said vertex as good neurons and bad neurons, respectively, such that the number of good neurons in said vertex is the same as the number of bad neurons in said vertex;

wherein each good neuron in a vertex in the top-down digraph corresponds to a bad neuron in said vertex, a good neuron in a corresponding vertex in the bottom-down digraph, and a bad neuron in the corresponding vertex in the bottom-down digraph.

25. The method of claim 24, further comprising:

designating vertices that do not have incoming directed edges as source vertices;

designating vertices that do not have outgoing directed edges as sink vertices; and

for each good neuron in a sink vertex in the top-down digraph;

connecting said good neuron to an evaluation module, wherein the evaluation module determines whether output from said good neuron is good or bad, such that good output from said good neuron is connected to a corresponding good neuron in a corresponding source vertex in the bottom-up digraph, and bad output from said good neuron is connected to a corresponding bad neuron in the corresponding source vertex in the bottom-up digraph.

26. The method of claim 18, wherein:

each node has a sensory or motor modality;

each directed edge interconnecting a sending vertex to a receiving vertex comprises multiple synaptic connections such that a neuron in the receiving vertex receives synaptic connections from a subset of neurons in the sending vertex; and

some acyclic digraphs are empty.