While much progress has been achieved over the last decades in neuro-inspired
machine learning, there are still fundamental theoretical problems in
gradient-based learning using combinations of neurons. These problems, such as
saddle points and suboptimal plateaus of the cost function,