Hyperdimensional Neural Network for Edge AI Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional deep neural network (DNN) architectures are power-intensive, have large memory footprints, and are not reconfigurable, making them unsuitable for low-power devices such as battery-operated devices and edge computing applications.
Innovation Solution
A reconfigurable hyperdimensional neural network architecture that uses non-Multiply and Accumulate (MAC) operations, specifically exclusive OR (XOR) and shift accumulate (SACC) operations, to reduce power consumption and memory usage, enabling field reconfiguration and efficient processing of data using hyperdimensional vectors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional deep neural network architectures are used, then recognition and classification accuracy is improved, but power consumption increases substantially
Solution Approach 1:
The patent transforms the computational parameters from traditional floating-point multiply-accumulate operations to integer-based operations with bit-shift weighting. This parameter change maintains classification accuracy while dramatically reducing power consumption by eliminating complex multiplication hardware and using simpler integer arithmetic operations that are energy-efficient.
Solution Approach 2:
The patent replaces the mechanical computation system of traditional MAC operations with a hyperdimensional computing system using binding and mixing operations. This substitution eliminates the need for energy-intensive multiplication circuits while maintaining the ability to perform recognition and classification tasks through vector-space operations.
2Measurement precision
If traditional deep neural network architectures are used, then recognition and classification accuracy is improved, but memory footprint increases substantially
Solution Approach 1:
The patent transitions from traditional low-dimensional weight matrices to high-dimensional hyperdimensional vectors for parameter representation. This dimensional change allows the network to maintain high classification accuracy with significantly reduced memory requirements by exploiting the properties of high-dimensional vector spaces where information can be more compactly encoded.
Solution Approach 2:
The patent changes the parameter representation from dense floating-point weights to sparse integer-based hyperdimensional vectors. This parameter transformation reduces memory footprint by enabling more efficient storage formats while maintaining the computational accuracy needed for recognition tasks.
3Productivity
If traditional deep neural network architectures are used, then processing capability is improved, but reconfigurability in the field is lost
Solution Approach 1:
The patent implements dynamic reconfigurability by allowing the hyperdimensional codebooks and binding operations to be updated in the field. The system can dynamically adapt its computational parameters and vector representations without requiring full retraining, enabling both strong processing capability and field reconfigurability through incremental learning and codebook updates.
4Measurement precision
If traditional deep neural network architectures are used, then training accuracy is improved, but training time increases
Solution Approach 1:
The patent performs preliminary encoding of training data into hyperdimensional vectors before the main training process. This preliminary action transforms the input data into a format that accelerates subsequent training convergence, reducing overall training time while maintaining the ability to achieve high training accuracy through the properties of hyperdimensional representations.
Data Source
AI summary
Method and apparatus for processing data using a reconfigurable, hyperdimensional neural network architecture comprising a feature extractor and a classifier. The feature extractor comprises a neural network for encoding input information into hyperdimensional (HD) vectors and extracting at least one particular HD vector representing at least one feature within the input information, wherein the neural network comprises no more than one multiply and accumulate operator. The classifier is coupled to the feature extractor for classifying the at least one particular HD vector to produce an indicium of classification for the at least one particular HD vector and wherein the classifier does not comprise any multiply and accumulate operators.