Learning Apparatus for Point Process Prediction via Support Set Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing meta-learning techniques for point processes face challenges in efficiently determining the relationship between previous events with a small amount of calculation, particularly when using models like Deepset, Transformer, or RNN, which result in high computational complexity or inability to ascertain relationships between distant events.
Innovation Solution
A learning device that divides a support set into sections, uses neural networks to extract latent vectors from each section, and derives an intensity function based on these vectors to predict event likelihood, reducing computational load and enabling parallel processing and accurate relationship determination.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If meta-learning techniques are used for point process prediction, then the ability to predict event occurrences is improved, but the computational complexity increases making it difficult to ascertain relationships of previous events with small calculation
Solution Approach 1:
The support set is divided into multiple sections, and each section is processed independently to extract latent vectors. This segmentation reduces the computational complexity of analyzing the entire sequence while preserving the ability to capture relationships between events. The patent applies this by splitting the support set into sections and processing each section separately through neural networks.
Solution Approach 2:
The patent extracts only the essential latent representations from event sequences using neural networks, rather than processing all raw event data. This extraction approach maintains prediction accuracy while reducing computational burden by focusing on the most relevant features.
2Loss of information
If models like Deepset, Transformer, or RNN are used to process event sequences, then the relationship between events can be ascertained, but the computational load becomes excessively high
Solution Approach 1:
By dividing the support set into multiple smaller sections, the patent enables parallel processing of these sections through neural networks. This segmentation maintains the ability to determine event relationships while significantly reducing the computational load compared to processing the entire sequence as a single unit.
Solution Approach 2:
The patent processes only the necessary portions of the event sequence by dividing it into sections and extracting latent vectors from each section. This partial processing approach is sufficient for determining event relationships without the excessive computational cost of processing the complete sequence in detail.
3Reliability
If the entire support set is processed as a single sequence, then comprehensive event relationships are captured, but the sequence length increases leading to higher computational complexity
Solution Approach 1:
The patent divides the support set into multiple sections, processing each section independently to extract latent vectors. This segmentation reduces the effective sequence length for each processing unit while maintaining comprehensive relationship analysis through the aggregation of results from all sections.
Solution Approach 2:
After processing individual sections separately, the patent merges the extracted latent vectors to form a comprehensive representation. This combining step ensures that relationships across the entire support set are captured while avoiding the computational complexity of processing the full sequence as a single unit.
Data Source
AI summary
A learning device for predicting an occurrence of an event includes a memory and a processor configured to divide a support set extracted from a set of previous data for learning into a plurality of sections, output a first latent vector based on each of the plurality of divided sections and output a second latent vector based on each of the output first latent vectors, and output an intensity function indicating a likelihood of the event occurring based on the second latent vector.