Abstract:Hyperdimensional Computing affords simple, yet powerful operations to create long Hyperdimensional Vectors (hypervectors) that can efficiently encode information, be used for learning, and are dynamic enough to be modified on the fly. In this paper, we explore the notion of using binary hypervectors to directly encode the final, classifying output signals of neural networks in order to fuse differing networks together at the symbolic level. This allows multiple neural networks to work together to solve a problem, with little additional overhead. Output signals just before classification are encoded as hypervectors and bundled together through consensus summation to train a classification hypervector. This process can be performed iteratively and even on single neural networks by instead making a consensus of multiple classification hypervectors. We find that this outperforms the state of the art, or is on a par with it, while using very little overhead, as hypervector operations are extremely fast and efficient in comparison to the neural networks. This consensus process can learn online and even grow or lose models in real time. Hypervectors act as memories that can be stored, and even further bundled together over time, affording life long learning capabilities. Additionally, this consensus structure inherits the benefits of Hyperdimensional Computing, without sacrificing the performance of modern Machine Learning. This technique can be extrapolated to virtually any neural model, and requires little modification to employ - one simply requires recording the output signals of networks when presented with a testing example.

What problem does this paper attempt to address?

### Problems the paper attempts to solve The paper aims to solve how to fuse different neural networks at the symbolic level to jointly solve problems. Specifically, the paper proposes a method using Hyperdimensional Computing (HDC), by directly encoding the final classification output signals of neural networks into binary hypervectors, thereby achieving the fusion of multiple neural networks. This method allows multiple neural networks to work together to improve overall performance while maintaining low overhead. ### Main contributions 1. **Fusing different neural networks**: The paper proposes a method that can fuse neural networks with different architectures, modalities, and even large performance differences together to form a unified prediction model. 2. **Low overhead**: This method takes advantage of the high - efficiency of hypervector operations and can improve overall performance with almost no additional computational cost. 3. **Online learning and dynamic adjustment**: This method supports online learning and can dynamically add or remove models in a real - time environment to adapt to changing data and requirements. 4. **Long - term learning ability**: Hypervectors can be used as memory storage, continuously accumulated and optimized over time, achieving lifelong learning. ### Method overview 1. **Encoding neural network output signals**: Convert the output signals of the final layer of the neural network into hypervectors, and generate hypervector representations through quantization and binding operations. 2. **Training hypervector models**: For each training sample, generate its corresponding hypervector, and form a hypervector model for each category through consensus summation. 3. **Multi - model fusion**: Fuse the hypervector models of multiple neural networks into a single prediction model through consensus summation to achieve the collaborative work of multiple models. 4. **Online learning and dynamic adjustment**: Support online learning, and can dynamically add or remove models to adapt to new data and task requirements. ### Experimental verification The paper has carried out experimental verification on multiple datasets, including MNIST, CIFAR - 10 and CIFAR - 100, demonstrating the effectiveness and superiority of this method in different scenarios. ### Formula representation - **Hypervector generation**: \[ R = D_1\oplus F_1+D_2\oplus F_2+\cdots +D_n\oplus F_n \] where \(D_i\) is a data item and \(F_i\) is a basic hypervector. - **Consensus summation**: \[ \text{Consensus Sum}=\sum_{i = 1}^n X_i \] where \(X_i\) is a hypervector participating in the consensus. - **Weighted consensus summation**: \[ R = a_1D_1\oplus F_1+a_2D_2\oplus F_2+\cdots +a_nD_n\oplus F_n \] where \(a_i\) is a weight, satisfying \(a_1 + a_2+\cdots +a_n=n\). ### Conclusion The paper proposes a neural network fusion method based on hyperdimensional computing. By encoding the output signals of neural networks into hypervectors, it realizes the collaborative work of multiple models. This method not only improves overall performance but also supports online learning and dynamic adjustment, and has broad application prospects.

Gluing Neural Networks Symbolically Through Hyperdimensional Computing

Generalized Holographic Reduced Representations

Spiking Hyperdimensional Network: Neuromorphic Models Integrated with Memory-Inspired Framework

Hyperdimensional Computing for Node Classification and Link Prediction

Hyperdimensional computing with holographic and adaptive encoder

Hardware Optimizations of Dense Binary Hyperdimensional Computing: Rematerialization of Hypervectors, Binarized Bundling, and Combinational Associative Memory

Hyperdimensional Computing with Spiking-Phasor Neurons

Hyperdimensional Computing vs. Neural Networks: Comparing Architecture and Learning Process

In-memory hyperdimensional computing

Efficient Hyperdimensional Computing With Spiking Phasors

Generalized Learning Vector Quantization for Classification in Randomized Neural Networks and Hyperdimensional Computing

GraphHD: Efficient graph classification using hyperdimensional computing

Hyper-Connections

Multi-Label Classification With Hyperdimensional Representations

Modularizing and Assembling Cognitive Map Learners via Hyperdimensional Computing

HyperNode: an Efficient Node Classification Framework Using HyperDimensional Computing

A Unified View Between Tensor Hypergraph Neural Networks And Signal Denoising

HyperFeel: an Efficient Federated Learning Framework Using Hyperdimensional Computing

Classification using Hyperdimensional Computing: A Review

Efficient Hyperdimensional Computing

Nested Hyperbolic Spaces for Dimensionality Reduction and Hyperbolic NN Design