In this note, we propose a novel technique to reduce the algorithmic complexity of neural network training by using matrices of tropical real numbers instead of matrices of real numbers. Since the tropical arithmetics replaces multiplication with addition, and addition with max, we theoretically achieve several order of magnitude better constant factors in time complexities in the training phase. The fact that we replace the field of real numbers with the tropical semiring of real numbers and yet achieve the same classification results via neural networks come from deep results in topology and analysis, which we verify in our note. We then explore artificial neural networks in terms of tropical arithmetics and tropical algebraic geometry, and introduce the multi-layered tropical neural networks as universal approximators. After giving a tropical reformulation of the backpropagation algorithm, we verify the algorithmic complexity is substantially lower than the usual backpropagation as the tropical arithmetic is free of the complexity of usual multiplication.