Abstract
In the recent decade, artificial intelligence and machine learning has become increasingly popular for solving complex real-world problems. In particular problems which was believed to be very hard or in some cases impossible by computers have seen a surge in interest from both academia but also the indus- try. A recent example is the defeat of the worlds best Go player by Googles AlphaGo using deep neural networks. At the core of neural networks is the a part of the neurons called an activation function. This function is of mayor significance in how the network operates, but is often overlooked. One most often picks one of the commonly chosen non-adaptive functions as activation function. There have been research into using adaptive sigmoid or ReLU functions, but these have the drawbacks that the adaptations caused by data from one local region would effect the global domain. We therefore propose to use adaptive spline functions with free knots. Research into the field of spline networks have been limited in sophistication, with interpolating cubic splines being the most researched. The implementation in this thesis is using splines with B-splines as a basis, and is therefore free to use any polynomial degree.