Recent changes to Home

Home modified by Ken Del Signore

Ken Del Signore — Tue, 01 Mar 2016 03:34:18 -0000

--- v68
+++ v69
@@ -23,7 +23,7 @@
 http://sourceforge.net/projects/sequencetree/files/notes.txt/download
 http://sourceforge.net/projects/sequencetree/files/bica6.cpp/download

-11/2015
+

Home modified by Ken Del Signore

Ken Del Signore — Sat, 27 Feb 2016 01:06:05 -0000

--- v67
+++ v68
@@ -1,5 +1,8 @@
 latest code at github:
 https://github.com/kwd2/graphnet/wiki
+
+latest working draft:
+https://docs.google.com/document/d/1nJzJwgtsViThUcien_4tXDAD-sU_6bfY37Iny8WQc0k/edit?usp=sharing

 working draft:
 https://drive.google.com/open?id=1xw6fSTbktPrf3XoaLd4ufSaQqOMqw-EDNiTAWNFmgJw

Home modified by Ken Del Signore

Ken Del Signore — Fri, 27 Nov 2015 00:32:28 -0000

--- v66
+++ v67
@@ -1,3 +1,10 @@
+latest code at github:
+https://github.com/kwd2/graphnet/wiki
+
+working draft:
+https://drive.google.com/open?id=1xw6fSTbktPrf3XoaLd4ufSaQqOMqw-EDNiTAWNFmgJw
+
+
 /* draft*/

 The graphnet project contains code and results of a data analysis technique that uses a recursive graph structure to analyze large volumes of data.  The technique was initially developed to analyze large volumes of packet data generated by smartphones in the 3G/4G telecom networks.  The method is generally applicable to data sets that consist of sequences of symbols, such as the words in a sentence or the letters in a word.

Home modified by Ken Del Signore

Ken Del Signore — Sat, 07 Nov 2015 16:15:33 -0000

--- v65
+++ v66
@@ -10,12 +10,12 @@
 https://docs.google.com/document/d/1xw6fSTbktPrf3XoaLd4ufSaQqOMqw-EDNiTAWNFmgJw/edit?usp=sharing

 code:
-http://sourceforge.net/projects/sequencetree/files/bica7.cpp/download
+http://sourceforge.net/projects/sequencetree/files/notes.txt/download
 http://sourceforge.net/projects/sequencetree/files/bica6.cpp/download

-....
+11/2015

-updating code as on Sept 2015.
+

Home modified by Ken Del Signore

Ken Del Signore — Tue, 03 Nov 2015 19:49:49 -0000

--- v64
+++ v65
@@ -1,73 +1,21 @@
 /* draft*/

-The graphnet project contains code and results of a data analysis technique that uses a recursive graph structure to analyze large volumes of data.  The technique was initially developed to analyze large volumes of packet data generated by smartphones in the 3G/4G telecom networks.  The method appears to be generally applicable to data sets that consist of sequences of symbols, such as the words in a sentence or the letters in a word.
-
-outline of BICA paper:
+The graphnet project contains code and results of a data analysis technique that uses a recursive graph structure to analyze large volumes of data.  The technique was initially developed to analyze large volumes of packet data generated by smartphones in the 3G/4G telecom networks.  The method is generally applicable to data sets that consist of sequences of symbols, such as the words in a sentence or the letters in a word.

-
-symbols are clusters of neurons that are linked together [Hebb 53]
-
-cluster of neurons can be represented as a single "Node"             class Node{...};
-
-symbols are nodes
-
-nodes belong to parcels
-
-a parcel is analogous to a single NN layer
-
-A node (symbol) from one parcel can grow a connection to any other parcel and form a new node in that parcel.
-
-two types of parcels:
-direct      parcel  - two nodes that fire together are directly linked
-associative parcel  - two nodes that fire together are linked with a new association node
-
-Simple Hebbian learning:  A node's weights to it's child nodes are decreased when that node fires and touches the child nodes.  If a child node fires afterwards, the same weight is enhanced. 
-
-either type of parcel can form connections in three ways: forward, backward, or both.
-
-The example input:
-parcel input = / a / b /
-
-would result in the following connectivity in each parcel:
-
-direct parcel, forw:  a -->  b
-direct parcel, back:  a <--  b
-direct parcel, both:  a --- b  
-
-assoc. parcel, forw:  a --> (ab) --> b 
-assoc. parcel, back:  a <-- (ab) <-- b
-assoc. parcel, both:  a --- (ab) --- b
+The paper and code below are to a graph analysis of the WordNet database:

-symbols(nodes) that fire in one parcel can send their output to other parcels (association areas).  1 or more parcels can provide input to a new parcel.  This is analogous to a neuron growing an axon to another parcel and drafting a free neuron to represent the symbol in that parcel.
+course draft paper:
+https://docs.google.com/document/d/1xw6fSTbktPrf3XoaLd4ufSaQqOMqw-EDNiTAWNFmgJw/edit?usp=sharing

+code:
+http://sourceforge.net/projects/sequencetree/files/bica7.cpp/download
+http://sourceforge.net/projects/sequencetree/files/bica6.cpp/download

+....

-1->1 allows some set of developed symbols to re-develop from scratch in an isolated parcel.  Allows new topologies to form and be exploited.
-
-2->1 allows associations to develop
-
-
-
-Many specialized parcels
-
-
-grow parcels manually.
-
-interconnect parcels any way possible (in main() ) to achieve desired input/output.
-
-
-STM parcel forms short term associations between any two or more symbols
-
-
-filter parcels
-
-
-
-
-
-
+updating code as on Sept 2015.

Home modified by Ken Del Signore

Ken Del Signore — Tue, 18 Nov 2014 14:09:41 -0000

--- v63
+++ v64
@@ -67,6 +67,10 @@

+
+
+
+
 /* draft*/
 The current direction of the project is to build graph structures from large volumes of symbol sequences.  Sequences can be of the general form [a,b,c,d], or [man, eat, meat].  Nodes of the graph correspond to each symbol, and connections

Home modified by Ken Del Signore

Ken Del Signore — Mon, 17 Nov 2014 20:47:24 -0000

--- v62
+++ v63
@@ -22,9 +22,9 @@
 direct      parcel  - two nodes that fire together are directly linked
 associative parcel  - two nodes that fire together are linked with a new association node

-Simple Hebbian learning:  A node's weights to it's child nodes are decreased when that node fires and touches the child nodes.  If the child node fires afterwards, the same weight is enhanced. 
+Simple Hebbian learning:  A node's weights to it's child nodes are decreased when that node fires and touches the child nodes.  If a child node fires afterwards, the same weight is enhanced. 

-either type of parcel can adjust weights in three ways: forward, backward, or both.
+either type of parcel can form connections in three ways: forward, backward, or both.

 The example input:
 parcel input = / a / b /
@@ -40,7 +40,9 @@
 assoc. parcel, both:  a --- (ab) --- b

-parcels can send output to other parcels (association areas).  1 or more parcels can provide input to a new parcel.  This is analogous to a neuron growing an axon to another parcel and causing a free neuron to fire there.
+symbols(nodes) that fire in one parcel can send their output to other parcels (association areas).  1 or more parcels can provide input to a new parcel.  This is analogous to a neuron growing an axon to another parcel and drafting a free neuron to represent the symbol in that parcel.
+
+

 1->1 allows some set of developed symbols to re-develop from scratch in an isolated parcel.  Allows new topologies to form and be exploited.

Home modified by Ken Del Signore

Ken Del Signore — Mon, 17 Nov 2014 19:01:34 -0000

--- v61
+++ v62
@@ -40,7 +40,7 @@
 assoc. parcel, both:  a --- (ab) --- b

-parcels can send output to other parcels (association areas).  1 or more parcels can provide input to a new parcel.  
+parcels can send output to other parcels (association areas).  1 or more parcels can provide input to a new parcel.  This is analogous to a neuron growing an axon to another parcel and causing a free neuron to fire there.

 1->1 allows some set of developed symbols to re-develop from scratch in an isolated parcel.  Allows new topologies to form and be exploited.

Home modified by Ken Del Signore

Ken Del Signore — Mon, 17 Nov 2014 18:58:04 -0000

--- v60
+++ v61
@@ -6,7 +6,7 @@

-symbols are clusters of neurons that are linked together
+symbols are clusters of neurons that are linked together [Hebb 53]

 cluster of neurons can be represented as a single "Node"             class Node{...};

Home modified by Ken Del Signore

Ken Del Signore — Mon, 17 Nov 2014 18:56:16 -0000

--- v59
+++ v60
@@ -4,18 +4,25 @@

 outline of BICA paper:

-code in Files section bica3.cpp, bica2.cpp, bica1.cpp

+
+symbols are clusters of neurons that are linked together
+
+cluster of neurons can be represented as a single "Node"             class Node{...};

 symbols are nodes

 nodes belong to parcels

+a parcel is analogous to a single NN layer
+
+A node (symbol) from one parcel can grow a connection to any other parcel and form a new node in that parcel.
+
 two types of parcels:
 direct      parcel  - two nodes that fire together are directly linked
 associative parcel  - two nodes that fire together are linked with a new association node

-A node's weights to it's child nodes are decreased when that node fires and touches the child nodes.  If the child node fires afterwards, the same weight is enhanced. 
+Simple Hebbian learning:  A node's weights to it's child nodes are decreased when that node fires and touches the child nodes.  If the child node fires afterwards, the same weight is enhanced. 

 either type of parcel can adjust weights in three ways: forward, backward, or both.

@@ -41,11 +48,18 @@

+Many specialized parcels

+grow parcels manually.
+
+interconnect parcels any way possible (in main() ) to achieve desired input/output.

+STM parcel forms short term associations between any two or more symbols

+
+filter parcels