Method and system for producing variable length context models
View Patent ↗Various embodiments of the present invention provide methods and systems for determining, representing, and using variable-length contexts in a variety of different computational applications. In one embodiment of the present invention, a balanced tree is used to represent all possible contexts of a fixed length, where the depth of the balanced tree is equal to the fixed length of the considered contexts. Then, in the embodiment, a pruning technique is used to sequentially coalesce the children of particular nodes in the tree in order to produce an unbalanced tree representing a set of variable-length contexts. The pruning method is selected, in one embodiment, to coalesce nodes, and, by doing so, to truncate the tree according to statistical considerations in order to produce a representation of a variably sized context model suitable for a particular application.
1. A method, carried out on an electronic computer controlled by instructions of a computer program, for determining a variably sized context model from input signal data having symbols chosen from an alphabet having n symbols {a 1 , a 2 , . . . , a n }, the method comprising:
counting, by the electronic computer, a number of occurrences of metasymbols in the input signal data formed by a context of K symbols and a currently considered symbol;
preparing, by the electronic computer, a context tree, each node having n children, the context tree having leaf nodes, each leaf node associated with a particular context derived from a path of edges and nodes leading to the leaf node from a root node of the context tree;
pruning, by the electronic computer, the context tree by an ordered sequence of merge operations that each merges the children of a merge-candidate node; and
storing, by the electronic computer, the pruned context tree in a computer-readable medium for subsequent use as the variably sized context model in one of a data-compression system, a data-prediction system, a digital-data denoising system, and another such data-processing system that transforms input data to output data using the variably sized context model.
2. The method of claim 1 wherein pruning the context tree by an ordered sequence of merge operations that each merges the children of a merge-candidate node further includes:
while the number of leaf nodes in the context tree is greater than a specified number of leaf nodes,
selecting a merge-candidate having a most favorable merge metric;
merging the merge-candidate node by promoting sums of occurrence counts associated with each child node of the merge-candidate node to the merge-candidate node; and
considering the merged merge-candidate node to be a leaf node of the context tree.
3. The method of claim 2 wherein selecting a merge-candidate having a most favorable merge metric further includes:
when a node of the context tree is not associated with a current merge metric, computing a merge metric for the node; and
selecting a node having a current merge metric most favorably indicative of being next merged.
4. The method of claim 1 wherein the current merge metric for a node is one of:
a smallest occurrence count associated with a child of the node;
a sum of occurrence counts associated with the children of the node; and
a merge metric based on similarities between empirical symbol probability distributions in contexts that are candidates for merging and an empirical symbol probability distribution in a merged context.
5. A variably sized context model, stored in a computer readable medium, prepared by a method for determining a variably sized context model comprising:
counting a number of occurrences of metasymbols in an input signal data formed by a context of K symbols and a currently considered symbol;
preparing a context tree, each node having n children, the context tree having leaf nodes, each leaf node associated with a particular context derived from a path of edges and nodes leading to the leaf node from a root node of the context tree;
pruning the context tree by an ordered sequence of merge operations that each merges the children of a merge-candidate node; and
storing the pruned context tree in the computer-readable medium for subsequent use as the variably sized context model in one of a data-compression system, a data-prediction system, a digital-data denoising system, and another such data-processing system that transforms input data to output data using the variably sized context model.
6. A computer readable medium containing executable instructions which, when executed in a processing system, causes the system to perform a method for determining a variably sized context model by:
counting a number of occurrences of metasymbols in an input signal data formed by a context of K symbols and a currently considered symbol;
preparing a context tree, each node having n children, the context tree having leaf nodes, each leaf node associated with a particular context derived from a path of edges and nodes leading to the leaf node from a root node of the context tree;
pruning the context tree by an ordered sequence of merge operations that each merges the children of a merge-candidate node; and
storing the pruned context tree in a computer-readable medium for subsequent use as the variably sized context model in one of a data-compression system, a data-prediction system, a digital-data denoising system, and another such data-processing system that transforms input data to output data using the variably sized context model.
7. A method, carried out on an electronic computer controlled by instructions of a computer program, for generating a variably sized context model from a fixed length context model expressed as a context tree of nodes, each node having n children, each child associated with a symbol of an alphabet having n symbols {a 1 , a 2 , . . . , a n }, the context tree having n K leaf nodes, each leaf node associated with a context and an occurrence count, and a root node, the method comprising:
determining, by the electronic computer, a number of desired leaf nodes;
while the number of leaf nodes in the context tree is greater than the desired number of leaf nodes,
selecting, by the electronic computer, a next node to merge; and
merging, by the electronic computer, the next node to merge, the merged node thereafter considered to be a leaf node; and
storing, by the electronic computer, the context tree in a computer-readable medium for subsequent use as the variably sized context model in one of a data-compression system, a data-prediction system, a digital-data denoising system, and another such data-processing system that transforms input data to output data using the variably sized context model.
8. The method of claim 7 wherein selecting a next node to merge further includes:
evaluating each node by obtaining for the node a merge metric; and
selecting as the next node to merge the node with a most favorable merge-metric value.
9. The method of claim 8 wherein most favorable merge-metric value is a lowest merge-metric value and obtaining for the node a merge metric further includes:
obtaining a merge-metric value associated with the node by a previous computation.
10. The method of claim 8 wherein most favorable merge-metric value is a lowest merge-metric value and obtaining for the node a merge metric further includes:
obtaining the merge-metric as the lowest occurrence count associated with any child of the node, when all children are leaf nodes, and otherwise indicating the node to be not ready for merging.
11. The method of claim 8 wherein most favorable merge-metric value is a lowest merge-metric value and obtaining for the node a merge metric further includes:
obtaining the merge-metric as the sum of occurrence counts associated with the children of the node, when all children are leaf nodes, and otherwise indicating the node to be not ready for merging.
12. The method of claim 8 wherein most favorable merge-metric value is a lowest merge-metric value and obtaining for the node a merge metric further includes:
obtaining the merge-metric S as
S
a
1
,
a
2
,
…
,
a
n
=
∑
i
=
1
n
C
a
i
D
(
P
a
i
||
P
a
1
,
a
2
,
…
,
a
n
)
where
C a i denotes the occurrence count for symbol a i in each child context,
P a i denotes an empirical symbol probability distribution for symbol a i in each child context,
P a 1 , a 2 , . . . , a n denotes an empirical symbol probability distribution of the merged context, and
D(•∥•) denotes the Kullback-Leibler distance between distributions when all children are leaf nodes, and otherwise indicating the node to be not ready for merging.
13. The method of claim 8 wherein merging the next node to merge further includes associating with the next node to merge an occurrence count equal to the sum of the occurrence counts associated with the children nodes of the next node to merge.
14. The method of claim 7 employed within a discrete denoiser to computer minimal-distortion replacement of symbols within variably sized contexts.
15. A variably sized context model, stored in a computer readable medium, prepared by a method for generating a variably sized context model comprising:
determining a number of desired leaf nodes;
while the number of leaf nodes in a context tree is greater than the desired number of leaf nodes,
selecting a next node to merge; and
merging the next node to merge, the merged node thereafter considered to be a leaf node; and
storing the pruned context tree in the computer-readable medium for subsequent use as the variably sized context model in one of a data-compression system, a data-prediction system, a digital-data denoising system, and another such data-processing system that transforms input data to output data using the variably sized context model.
16. A computer readable medium containing executable instructions which, when executed in a processing system, causes the system to perform a method for generating a variably sized context model comprising:
determining a number of desired leaf nodes;
while the number of leaf nodes in a context tree is greater than the desired number of leaf nodes,
selecting a next node to merge; and
merging the next node to merge, the merged node thereafter considered to be a leaf node; and
storing the pruned context tree in a computer-readable medium for subsequent use as the variably sized context model in one of a data-compression system, a data-prediction system, a digital-data denoising system, and another such data-processing system that transforms input data to output data using the variably sized context model.