Hierarchical Temporal Memory for Online Demand Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current demand forecasting methods are insufficient due to inefficiencies and ineffectiveness in predicting future demand, particularly in online environments, where they often require costly supervised learning and are not resistant to noisy data.
Innovation Solution
Implementing an online demand prediction framework using Hierarchical Temporal Memory (HTM) that learns temporal patterns from time-series data, enabling continuous adaptation and resistance to noisy inputs without the need for a distinct training phase, allowing for efficient forecasting of future demand across multiple points in time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional supervised learning methods are used for demand forecasting, then the system can learn from labeled data, but it requires costly training and is not resistant to noisy data
Solution Approach 1:
The patent replaces traditional supervised learning mechanisms with Hierarchical Temporal Memory (HTM) unsupervised learning. Instead of using costly labeled training data and iterative optimization algorithms, the HTM system learns temporal patterns directly from raw time-series data through biologically-inspired neural mechanisms that are inherently robust to noise, eliminating the need for expensive training phases while improving reliability.
Solution Approach 2:
The patent changes the fundamental learning parameters from supervised (requiring labeled data and loss function optimization) to unsupervised HTM learning (using temporal pattern recognition). This parameter change allows the system to achieve noise resistance without complex training, as HTM learns by identifying temporal sequences and predictions directly from the data structure itself.
2Reliability
If traditional demand forecasting methods are used, then the system can provide predictions, but they are insufficient for online environments with noisy data
Solution Approach 1:
The patent converts the harmful effect of noisy data into a benefit by using HTM's temporal pattern recognition capability. Instead of treating noise as something to be filtered out, the HTM system learns to distinguish meaningful temporal sequences from random variations, actually improving forecasting accuracy by focusing on persistent temporal patterns that survive noise contamination.
Solution Approach 2:
The patent employs HTM's ability to make rapid, lightweight predictions without expensive computational training. Each prediction is generated through efficient temporal pattern matching that requires minimal computational resources compared to traditional methods, allowing continuous online forecasting even in noisy environments.
3Productivity
If multi-step demand predictions are made, then future demand at multiple points can be forecasted, but computational complexity increases
Solution Approach 1:
The patent segments the demand forecasting task into hierarchical temporal levels using HTM. Instead of computing all multi-step predictions simultaneously, the system processes temporal patterns at different hierarchical levels, where lower levels capture short-term fluctuations and higher levels capture long-term trends, dividing the computational complexity into manageable segments.
Solution Approach 2:
The patent performs preliminary learning of temporal patterns during an unsupervised training phase, storing learned sequences and predictions in the HTM structure. Once learned, multi-step predictions are generated by matching current data against stored temporal patterns, significantly reducing the computational complexity of generating future predictions compared to re-computing from scratch.
Data Source
AI summary
In general, embodiments of the present invention provide systems, methods and computer readable media to forecast demand by implementing an online demand prediction framework that includes a hierarchical temporal memory network (HTM) configured to learn temporal patterns representing sequences of states of time-series data collected from a set of one or more data sources representing demand and input to the HTM. In some embodiments, the HTM learns the temporal patterns using a Cortical Learning Algorithm.