Prediction Model Smoothing via Segmented Distribution Distance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for smoothing posterior probability distributions in prediction models, particularly for sequential data, face significant computational challenges due to high calculation loads when handling inter-distribution distances, hindering efficient generalization capability and semi-supervised learning.
Innovation Solution
The approach involves updating the parameters of a prediction model to minimize the sum of inter-distribution distances among small-scale categorical distributions, rather than directly minimizing the distances among label sequences, allowing for approximate smoothing with reduced computational costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the inter-distribution distances among posterior probability distributions for sequential data are calculated to smooth the distributions, then the generalization capability is enhanced, but the calculation load increases significantly
Solution Approach 1:
The patent segments the calculation of inter-distribution distances by focusing only on small-scale categorical distributions at each time step rather than calculating distances over entire label sequences. This divides the complex global distance calculation into multiple simpler local calculations, reducing the overall computational burden while maintaining the smoothing effect on posterior probability distributions.
Solution Approach 2:
The patent changes the parameter being optimized from the full posterior probability distribution of label sequences to the parameters of small-scale categorical distributions at individual time steps. By updating these local distribution parameters to minimize local inter-distribution distances, the method achieves smoothing with reduced computational complexity compared to optimizing full sequence distributions.
2Stability of the object's composition
If the parameters are updated to minimize inter-distribution distances among posterior probability distributions, then smoothing is achieved, but it takes a significant amount of time
Solution Approach 1:
The patent segments the smoothing process by applying it independently to small-scale categorical distributions at each time step rather than computing smoothing over entire label sequences. This segmentation reduces the computational complexity from exponential in sequence length to linear, significantly reducing processing time while maintaining distribution smoothing.
Solution Approach 2:
The patent applies partial action by computing inter-distribution distances only for small-scale categorical distributions at each time step rather than for the complete posterior probability distributions. This partial computation is sufficient to achieve the desired smoothing effect without the excessive time cost of full distribution comparisons.
3Power
If the sum of inter-distribution distances among small-scale categorical distributions is minimized, then computational costs are reduced, but direct minimization of distances among label sequences is approximated
Solution Approach 1:
The patent segments the measurement of distribution distance into small-scale categorical distributions at each time step. While this approximation reduces computational cost from exponential to linear complexity, the cumulative effect of minimizing these local distances provides a sufficient approximation of the global smoothing objective, maintaining adequate measurement precision for practical applications.
Solution Approach 2:
The patent changes the measurement focus from full posterior probability distributions to small-scale categorical distribution parameters at individual time steps. This parameter change enables computationally efficient optimization while the aggregated effect across all time steps provides a reasonable approximation of the desired smoothing, balancing precision and computational feasibility.
Data Source
AI summary
An information processing apparatus according to an embodiment includes one or more hardware processors. The hardware processors obtain a first categorical distribution sequence corresponding to first input data and obtain a second categorical distribution sequence corresponding to second input data neighboring the first input data, by using a prediction model outputting a categorical distribution sequence representing a sequence of L categorical distributions for a single input data piece, where, L is a natural number of two or more. The hardware processors calculate, for each i of 1 to L, an inter-distribution distance between i-th categorical distributions in the first and second categorical distribution sequences. The hardware processors calculate a sum of L inter-distribution distances. The hardware processors update the prediction model's parameters to lessen the sum.