Alex Krizhevsky's CUDA Neural Network Code

Categories: Coding & Developer Tools, Research | Pricing: Free | Official Website ↗

A collection of CUDA/C++ code for implementing various neural network architectures and related routines.

This page presents a collection of CUDA/C++ code developed by Alex Krizhevsky, primarily focusing on neural network implementations and low-level GPU routines for deep learning. The code includes an abstract convolutional neural network framework with a Python front-end, supporting various layer types like convolutional, pooling, and fully-connected layers. It also features optimized 2D local filter routines and 2D convolution routines for different generations of NVIDIA GPUs. The repository provides implementations for specific neural network models such as Convolutional RBMs with Gaussian visible units and binary hidden units, and a general RBM training code. It also includes utility routines for subsampling, supersampling, and handling non-overlapping sub-squares of images, useful for building multi-layer convolutional networks and local softmax. The code was developed with a focus on performance, demonstrating significant speedups over CPU implementations for tasks like CIFAR-10 classification.

Key Features

Abstract convolutional neural network implementation
Support for various neural network layer types (convolutional, pooling, fully-connected)
Optimized 2D local filter routines for CUDA
Fast 2D convolution routines for CUDA
Implementations of RBMs (Restricted Boltzmann Machines)
Subsampling and supersampling routines
Routines for local softmax and multinomial sampling
Python front-end for neural network framework

Pros

Provides highly optimized CUDA implementations for core deep learning operations
Includes a flexible convolutional neural network framework
Demonstrates significant performance improvements over CPU implementations
Code is freely available for use and study
Includes examples of learned filters and RBMs

Cons

Code is described as 'very outdated'
Descriptions are 'rather skimpy', requiring email for help
Specific code versions are tied to older GPU architectures (GTX 580, GTX 280)
Parts of the older code may not work on current-generation GPUs
Primarily a collection of low-level code, not a user-friendly application

Use Cases

Implementing custom neural network architectures
Benchmarking GPU performance for deep learning tasks
Studying early deep learning code implementations
Developing specialized image processing filters on GPUs

Best For

Deep learning researchers
CUDA developers
Students learning neural network implementations
Engineers optimizing deep learning algorithms on GPUs

Integrations: Python

Platforms: Linux, Windows

Watch demo on YouTube ↗

View full Alex Krizhevsky's CUDA Neural Network Code profile on Tools-Radar | Browse Coding & Developer Tools tools | Alternatives to Alex Krizhevsky's CUDA Neural Network Code

Tools-Radar is a free directory of 10,000+ AI tools — discover, compare, and choose the right AI software for your needs. Visit tools-radar.com