Poisson Sampling for Random Forest Memory Access Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional random forest memory access patterns are inefficient, leading to performance degradation and increased power consumption due to random sampling of observation data, which affects hardware components like cache memory and system memory.
Innovation Solution
Implementing a Poisson distribution-based approach for bootstrap aggregation, where each observation is included a specific number of times in decision tree training sets, ensuring sequential memory access and reducing cache misses by fitting data into contiguous memory regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If random sampling is used for bootstrap aggregation in random forest, then each observation has equal probability of inclusion, but memory access patterns become inefficient causing performance degradation and increased power consumption
Solution Approach 1:
The patent changes the sampling method from uniform random sampling to Poisson sampling. Instead of selecting observations with equal probability through random sampling, the system uses Poisson distribution parameters (lambda) to determine the number of times each observation is included in the bootstrap sample. This parameter change maintains the required statistical properties while enabling sequential memory access patterns that improve hardware performance.
Solution Approach 2:
The patent replaces the mechanical random sampling process with a Poisson-based deterministic selection mechanism. Rather than using random number generation to select observations, the system calculates Poisson probabilities and directly determines inclusion counts, substituting the random mechanical process with a more efficient mathematical approach that enables sequential memory access.
2Reliability
If random sampling is used for bootstrap aggregation, then statistical properties are maintained, but cache memory performance deteriorates due to random memory access patterns
Solution Approach 1:
The patent applies Poisson distribution parameters to transform the sampling process. By using Poisson sampling with calculated probabilities, the system maintains the statistical properties required for valid random forest analysis while changing the memory access pattern from random to sequential. This reduces cache misses and improves memory hierarchy utilization, thereby reducing power consumption.
3Adaptability or versatility
If uniform random sampling is used for decision tree training, then each observation is treated equally, but memory access efficiency decreases leading to longer processing time
Solution Approach 1:
The patent replaces the uniform random sampling mechanism with Poisson sampling. This substitution allows the system to maintain flexible observation treatment (each observation can be included 0 or more times based on Poisson probability) while achieving sequential memory access patterns that significantly reduce processing time by improving memory access efficiency.
Data Source
AI summary
Systems, apparatuses and methods may provide for technology that generates inclusion data in accordance with a Poisson distribution, wherein the inclusion data specifies a number of inclusions for each observation in a set of observations. The technology may also train a first decision tree in a random forest based at least in part on the inclusion data.