Influence of Digital Fluctuations on Behavior of Neural Networks

This paper deals with effect of digital noise to numerical stability neural networks. Digital arises from the inexactness floating point values operations. Accumulated errors finally lead loss significance. Experiments show that more redundant networks have higher influence. is tested in both model and real world samples. As a result, one should exclude all results beginning fluctuations. Results experiments allow us hypothesize minimal function preserving significance were achieved for size close complexity dataset. So, it reason choose sizes network layers accordance particular datasets not universally an architecture general problem statement without relation data. In case fine tuning this suggests pruning can improve result accuracy reliability prediction due decrease article are based on analysis train than 50000 thousands epochs each network. Almost begin fluctuate.