• Distance Correlation
• Hebbian
• Bioinspired
• Deep Learning
• Evolutionary
• Sparse Neural Network

In this thesis automatic classification is applied to human histological images of colorectal cancer. Sparse neural networks and a recently proposed algorithm, namely Sparse Evolutionary Training (SET), are used to challenge the state of the art on a multi-class dataset composed of 5000 patches displaying 8 different tissue types.

Our sparse network reaches an accuracy of 93.2, surpassing the state of the art of 92.6 while having 70 times less parameters and needing 16 times less floating point operations.

Then, a biologically inspired algorithm similar to SET - namely, Distance Correlation connection learning with Sparse Evolutionary Training DCSET – is proposed.
SET prunes every epoch the smallest weights from a sparse network and adds new random connections. DCSET uses distance correlation instead to create new connections between the ones with the most correlated activations - following from Hebbian Learning principle that neurons that fire together wire together. Distance correlation is able to detect dependence among random variables, potentially also when the relation is non-linear.

We compare our algorithm to SET on the colorectal cancer and benchmark datasets and analyse why DCSET fails to bring performance improvements.