Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
G-Net: A Provably Easy Construction of High-Accuracy Random Binary Neural Networks
Authors: Alireza Aghasi, Nicholas F. Marshall, Saeid Pourmand, Wyatt Whiting
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, our binary models match convolutional neural network accuracies and outperform prior HDC models by large margins, for example, we achieve almost 30% higher accuracy on CIFAR-10 compared to prior HDC models. G-Nets are a theoretically justified bridge between neural networks and randomized binary neural networks, opening a new direction for constructing robust binary/quantized deep learning models. |
| Researcher Affiliation | Academia | Alireza Aghasi Dept. Electrical Engineering and Computer Science Oregon State University EMAIL Nicholas F. Marshall Dept. Mathematics Oregon State University EMAIL Saeid Pourmand Dept. Electrical Engineering and Computer Science Oregon State University EMAIL Wyatt D Whiting Dept. Mathematics Oregon State University EMAIL |
| Pseudocode | No | The paper includes architectural diagrams (Figure 2) but does not contain explicitly labeled pseudocode blocks or algorithms in a structured, step-by-step format. |
| Open Source Code | Yes | Our implementation is available at https://github.com/GNet2025/GNet. |
| Open Datasets | Yes | Here, we selectively present classification results on MNIST, CIFAR-10, and human activity recognition (HAR-WSS [29]). Also, a detailed and extended version of the numerical experiments presented in this section, along with additional accuracy and robustness analysis across other datasets, is included in Section O. ... MNIST, Fashion MNIST, and CIFAR10. We also include two human-activity recognition datasets, HARWSS (walking, sitting, standing) [65] and Epilepsy [70]; an automotive dataset, Ford-A [40]; a natural-language dataset, AG News [72]; and a time-series dataset, Fault Detection-A [64]. |
| Dataset Splits | Yes | The experiments involve fitting a G-Net to the original training data, followed by evaluation in the binary hyperspace by applying the corresponding EHD G-Net to the random sign embedding of test data. ...Table 1 summarizes the datasets used, including the number of input features (n0), output classes (n L), and the corresponding G-Net architectures. |
| Hardware Specification | Yes | The G-Net training is quick, for instance, training ten epochs on MNIST requires roughly 48 seconds on a desktop computer equipped with a Ge Force RTX 4090 GPU. |
| Software Dependencies | No | The paper does not explicitly state specific software dependencies with version numbers. |
| Experiment Setup | Yes | Table 1: Dataset and G-Net architecture specifications. In layer descriptions: CN denotes a convolutional layer (input channels, filters, output channels), EM an embedding layer (vocabulary size, number of outputs), FC a fully connected layer (number of outputs), and CL a classification layer (number of classes)... training ten epochs on MNIST... train a fully connected G-Net with two hidden layers (each of width 256) on Fashion MNIST and then form a binary EHD G-Net using a hyperdimensional embedding of size N = 1000. |