Cross-Platform Neural Network Library

I have been spending some time sharpening my skills on Artificial Intelligence again. I have been around and, although it’s nice, powerful and useful, I still like very much to write my own code and completely understand and dominate the object of study, so that is what I did recently — a personal neural network framework entirely written from scratch. I struggled a little bit with the different back-propagation gradient formulas, but after dominating those details I am satisfied with the current results. The acquired knowledge helps me to better understand bigger frameworks like Tensorflow, for example.

This tiny unnamed Neural Network library of mine is cross-platform, compatible with basically all hardware platforms and Operating Systems, and still small and with no external dependencies at all. It is fully self-contained. I like that, because deployment is easy. It can be integrated in any app, on desktop, mobile and embedded platforms in a matter of minutes.

The following simple video shows basic learning and recognition of digits. I ran it inside Unity3D because of its easiness for visual prototyping, but as said, the NN library itself has no dependencies, so it’s not tied to Unity or any other engines or libraries.

I will be constantly adding features to this personal lib — it’s not just for digits recognition! — and I intend to have it running on an intelligent robot which is going to entertain the family for a long time.

More on this later, thanks for reading.

Neural Networks

I have been creating a Deep Neural Network library on weekends. It is not tied or dependent on any particular engine, so it can be used both on database-related systems and VR/Games.

Neural Networks are awesome, as they try to emulate how real neurons work. The artificial ones also have dentrites (which I resume to input/output ports) and synapses (which I resume to connections between different neurons), and from that simple structure, some really interesting things can be created. As you probably know already, they can be used in many different domains: vision, voice and general cognition are some of them.

For now, my implementation is all CPU based. When it becomes rock solid and fully featured, I will consider moving some or maybe all parts to a GPU implementation, using Compute Shaders. Right now I’m still satisfied with CPU performance, so no urgency on GPU translation yet.

The following screenshot is taken from a small network visualized in Unity3D — so I can quickly test/confirm that the actual topologies and synapses are being created correctly.

The above network is a Feedforward one, showing an extremely simple (but correct and useful) neural net in this case, however the underlying structure can automatically build and interconnect a neural net of any size (only limited by memory), using either Perceptrons or Sigmoid neurons.

More on this later. Thanks for reading.

Dev Log opened

Hi there. So, after a few years, I just decided to reopen a public log. Here I’ll be posting about my personal progress in general, mostly about programming and robotics, and some times about real-life subjects. Welcome and feel free to leave me a comment. Thanks. Work on Etherea² Etherea² for Unity3D can potentially handle…