Simple and performance neural network .NET library
Its use native C++ dll for AVX acceleration!
Supported x64 and x86 platforms.
In this test 5620 examples in epoch, epoch time is ~66 (±2) seconds.
From this - 66 / 5620 = ~0,0117437 sec = ~11.7 ms for 1 example in training mode.
In image below, results of training in 2400 epochs! Reached error per epoch: ~270.
Even for classical neural networks with direct distribution, the recognition of numbers is not very effective.
But I decided to do a test with partial MNIST dataset. (only 10000 images for train and 1000 for test)
In this test network trained in 41 epochs. Reached test error per epoch: ~16 (error for test dataset!).
Layers cfg: 256 sigmoid (input), 256 tanh, 256 tanh, 256 tanh, 10 sigmoid (output)
Tests coming soon ;)