Parsimonious modeling by non-uniform kernel allocation
AssigneeMicrosoft Corporation
InventorsPeng Liu, Jian-Iai Zhou, Frank Kao-ping Soong
Life of this patent
Estimating the term…
Abstract
A multi-state pattern recognition model with non-uniform kernel allocation is formed by setting a number of states for a multi-state pattern recognition model and assigning different numbers of kernels to different states. The kernels are then trained using training data to form the multi-state pattern recognition model.
Claims16
A method comprising:
- setting a number of states for a multi-state pattern recognition model;
- a processor assigning different numbers of kernels to different states based on likelihoods of training data, wherein assigning different numbers of kernels to different states comprises:
- training an initial multi-state pattern recognition model using a uniform number of kernels for each state;
- training a second initial multi-state pattern recognition model using a second uniform number of kernels for each state;
- determining changes in likelihood caused by reducing the number of kernels for each state by one using the initial multi-state pattern recognition model and the second initial multi-state pattern recognition model;
- and
- reducing the number of kernels in the one state that results in the smallest change in likelihood;
- and
- training the kernels using training data to form the multi-state pattern recognition model.
The method of 1 wherein training the kernels using training data to form the multi-state pattern recognition model comprises forming a multi-state pattern recognition model that has the same number of model parameters as a multi-state pattern recognition model with a uniform number of kernels for each state but that outperforms the multi-state pattern recognition model with a uniform number of kernels for each state.
A method comprising:
- training an initial multi-state pattern recognition model using a uniform number of kernels per state;
- training a second initial multi-state pattern recognition model using a second uniform number of kernels per state;
- a processor reducing the number of kernels for at least one state based on likelihoods for training data;
- and
- using the reduced number of kernels to retrain the multi-state pattern recognition model.
A method comprising:
- assigning kernels to each of a plurality of states for a multi-state recognition model;
- a processor determining a reduction in a likelihood for each of the plurality of states that would occur if the number of kernels for a respective state was reduced, wherein determining a reduction in likelihood comprises training a plurality of initial multi-state recognition models, each initial multi-state recognition model associated with a separate uniform number of kernels per state, and using at least two of the initial multi-state recognition models to determine the reduction in likelihood;
- reducing the number of kernels for the state with the smallest reduction in likelihood;
- and
- training parameters for the kernels in the states to form the multi-state recognition model.
Dependencies are inferred from the claim text.