mini-dnn-cpp is a C++ demo of deep neural networks. It is implemented purely in C++, whose only dependency, Eigen, is header-only.

Download and unzip MNIST dataset in mini-dnn-cpp/data/mnist/.

mkdir build
cd build
cmake ..
make

Run ./demo.

Result: simple neural network with 3 FC layers can obtain 0.97+ accuracy on MNIST testset. LeNet-like CNN can obtain 0.98+ accuracy on MNIST testset.