Pattern Recognition via Multi-Lag Transition Probability Matrices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing pattern recognition methods, such as Markov chains, fail to accurately model sequences with strong temporal structures and dependencies over time or space, as they assume future states are independent of past states, leading to inaccuracies in modeling and recognition, especially in noisy or distorted conditions.
Innovation Solution
A method that learns transition probability matrices from frequency data of sequence symbol pairs at various distances, normalizing them to derive activation functions that capture statistical regularities across multiple lag values, enabling pattern recognition in sequences with complex dependencies and noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If Markov chain is used for pattern recognition, then the model is simple and computationally efficient, but it cannot accurately model sequences with temporal structures and dependencies over time
Solution Approach 1:
The patent segments the temporal dependency modeling into multiple independent Markov chains, each operating at different lag values. Instead of using a single complex model, the system divides the problem into multiple simpler sub-models that each capture temporal relationships at specific time intervals, then combines their outputs to achieve accurate pattern recognition while maintaining computational efficiency
Solution Approach 2:
The patent introduces a new dimension by incorporating lag values as an additional parameter. Instead of modeling temporal dependencies in a single time dimension, the system creates multiple Markov chains across different lag dimensions (lag=1, lag=2, lag=3, etc.), allowing the model to capture long-range temporal dependencies without requiring a single overly complex model
2Measurement precision
If transition probability matrices are learned for each tag and distance, then pattern recognition accuracy improves, but computational complexity and data processing requirements increase
Solution Approach 1:
The patent applies partial action by selectively processing only the most relevant features and transitions for each tag-distance combination. Instead of computing all possible transition probabilities uniformly, the system focuses computational resources on the most informative transitions and distances, achieving high accuracy while reducing unnecessary computational overhead
Solution Approach 2:
The system performs preliminary normalization of transition frequency matrices into transition probability matrices before pattern recognition. This pre-processing step organizes the data into a standardized format that facilitates more efficient subsequent computations and reduces the complexity of the main recognition algorithm
3Adaptability or versatility
If sequences with noise and distortion are processed, then real-world applicability improves, but recognition accuracy decreases due to high distortion and noise
Solution Approach 1:
The patent incorporates feedback mechanisms where the system continuously monitors recognition confidence levels and adjusts its processing accordingly. When noise and distortion are detected, the system can request re-processing, adjust parameter thresholds, or combine results from multiple lag values to compensate for degraded input quality, thereby maintaining accuracy in real-world noisy conditions
Solution Approach 2:
The system converts the harmful effect of noise and distortion into a benefit by using multiple lag values. The temporal spreading of information across different lags means that noisy observations at one lag can be compensated by cleaner observations at other lags, transforming the presence of noise into an opportunity for temporal averaging and error correction
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
The invention relates to a method for a pattern discovery and recognition, wherein a first sequence comprising first sequence symbols relating to a concept and a tag associated to the first sequence are received, transition probability matrices are obtained from transition frequency matrices representing the frequency data of the occurrences of the transitions between the first sequence symbols at different distances in the first sequence, and the transition probability matrices for each tag and each distance are learnt for obtaining an activation function determining the concept occurring in a second sequence. The invention also relates to a computer program product and an apparatus executing the pattern discovery and recognition method.