Sparse Hyperdimensional Models for Edge AI Memory Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning methods, such as Deep Neural Networks, face challenges in resource-constrained IoT devices due to high computational complexity and memory requirements, and sending data to cloud environments is undesirable due to scalability, real-time response, privacy, and security concerns.
Innovation Solution
A SparseHD framework that exploits sparsity in hyperdimensional computing by iteratively retraining HD models to drop least important class elements, using class-wise and dimension-wise sparsity methods, and implements an FPGA-based accelerator to maximize performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Deep Neural Networks are used for complex image classification tasks, then classification accuracy is improved, but computational complexity and memory requirement increase
Solution Approach 1:
The patent extracts and eliminates redundant or less important elements from the neural network model through pruning techniques. This removes unnecessary computational operations while preserving the essential functionality that maintains classification accuracy, thereby reducing computational complexity and memory requirements.
Solution Approach 2:
The patent applies different levels of pruning to different parts of the neural network based on their importance. Critical layers and filters are preserved with higher quality, while less important components are pruned more aggressively. This localized approach maintains overall accuracy while reducing computational burden.
2Measurement precision
If Deep Neural Networks are used for complex image classification tasks, then classification accuracy is improved, but memory requirement increases
Solution Approach 1:
The patent removes redundant parameters, weights, and activations from the neural network through pruning. This extraction of unnecessary elements directly reduces the memory footprint while maintaining the essential computational structures needed for accurate classification.
Solution Approach 2:
The patent changes the parameter representation by eliminating less important weights and biases. This parameter reduction decreases the memory required to store the model while preserving the critical parameters that drive accurate classification performance.
3Power
If data is sent to cloud computing environment, then computational power is improved, but scalability and real-time response are worsened
Solution Approach 1:
The patent segments the computational workload by pruning the model to retain only essential operations. This segmentation allows the model to run efficiently on edge devices with limited power, eliminating the need to offload computations to the cloud and enabling real-time processing.
4Power
If data is sent to cloud computing environment, then computational power is improved, but energy consumption increases
Solution Approach 1:
The patent extracts and removes unnecessary computational operations from the model. This reduction in computational complexity allows the pruned model to execute efficiently on energy-constrained edge devices, eliminating the need for energy-intensive cloud computing and reducing overall energy consumption.
Data Source
AI summary
A method of providing a trained machine learning model can include providing a trained non-binary hyperdimensional machine learning model that includes a plurality of trained hypervector classes, wherein each of the trained hypervector classes includes N elements, and then, eliminating selected ones of the N elements from the trained non-binary hyperdimensional machine learning model based on whether the selected element has a similarity with other ones of the N elements, to provide a sparsified trained non-binary hyperdimensional machine learning model.