Tabular Data Subset Selection for Memory-Limited AI Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Transformer-based in-context learning models like TabPFN face limitations due to their limited memory size, which restricts the amount of input data they can process during a single forward pass, especially when dealing with large and complex tabular datasets.
Innovation Solution
A system that retrieves a subset of data from a table based on context, ensuring it fits within the memory of the in-context learning model and is relevant to the task, thereby enabling accurate predictions in a single pass.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the memory size of the AI model is increased to process larger tabular datasets, then the amount of input data that can be processed during a single forward pass is improved, but the device complexity and computational resource requirements worsen
Solution Approach 1:
The patent segments the large tabular dataset into multiple smaller subsets, each fitting within the limited memory capacity of the AI model. The system identifies and selects relevant subsets based on the target record and task requirements, processes them separately, and aggregates the results. This segmentation allows the model to handle large datasets without requiring increased memory size.
Solution Approach 2:
The patent extracts only the necessary subset of data from the complete tabular dataset that is relevant to the specific prediction task and target record. By filtering and selecting only the essential data points needed for accurate prediction, the system reduces the data volume to fit within the model's memory constraints while maintaining predictive accuracy.
2Device complexity
If the subset of data is reduced to fit within the memory, then the device complexity is reduced, but the measurement precision of predictions may worsen due to loss of information
Solution Approach 1:
The patent performs preliminary actions by identifying and selecting the most relevant data subsets before feeding them to the AI model. The system analyzes the target record and task requirements in advance, pre-selects the necessary data points and features, and organizes them into optimized subsets. This preliminary selection ensures that the most informative data is retained while excluding redundant information, maintaining prediction accuracy within memory constraints.
Solution Approach 2:
The patent applies local quality by making different parts of the data have different levels of inclusion based on their relevance to the specific prediction task. Instead of uniformly reducing all data, the system selectively retains high-quality, task-relevant data points while excluding less important ones. This selective data retention ensures that the subset contains the most valuable information for accurate predictions.
3Productivity
If the AI model processes only a subset of data, then the productivity and processing speed are improved, but the reliability of predictions on the complete dataset may worsen
Solution Approach 1:
The patent implements feedback mechanisms where the system continuously evaluates the performance and reliability of predictions made on data subsets. By monitoring prediction outcomes and comparing them against expected results or validation data, the system adjusts its subset selection strategy to ensure reliable predictions. The feedback loop helps maintain prediction reliability while preserving the processing speed benefits of using subsets.
Solution Approach 2:
The patent performs preliminary validation and selection of data subsets to ensure they are representative and sufficient for reliable predictions. Before processing, the system evaluates potential subsets based on their ability to capture the essential patterns and relationships needed for accurate predictions. This preliminary assessment ensures that the selected subsets maintain prediction reliability while enabling fast processing.
Data Source
AI summary
An example operation may include one or more of identifying dimensional parameters of a memory of an artificial intelligence (AI) model, receiving tabular data for execution by the AI model, determining a subset of data from within the tabular data that fits within the dimensional parameters of the memory, extracting the subset of data from the tabular data and converting the subset of data into at least one vector, and executing the AI model on the subset of data to generate a predictive result.