Although PoW has achieved great success over the last decade, its core problem, energy waste, has not been well answered. The purpose of PoL is to use the wasted computing power in another field that is also consuming massive computing power, training large-scale deep learning models. The success of PoW is largely due to that its proof can be verified in constant time. In this paper, we propose a new method for replacing cryptopuzzles with large-scale model training based on the observation of the similarity between the gradient changes during model training and hash functions. Our new consensus mechanism can verify a proof in constant time. Experiments show that the performance of our proposed new consensus mechanism meets expectations.

Since Bitcoin, blockchain has attracted the attention of researchers. The consensus mechanism at the center of blockchain is often criticized for wasting a large amount of computing power for meaningless hashing. At the same time, state-of-the-art models in deep learning require increasing computing power to be trained. Proof of Learning (PoL) is dedicated to using the originally wasted computing power to train neural networks. Most of the previous PoL consensus mechanisms are based on two methods, recomputation or performance metrics. However, in practical scenarios, these methods both do not satisfy all properties necessary to build a large-scale blockchain, such as certainty, constant verification, therefore are still far away from being practical. In this paper, we observe that the opacity of deep learning models is similar to the pre-image resistance of hash functions and can naturally be used to build PoL. Based on our observation, we propose a method called Directed Guiding Gradient. Using this method, our proposed PoL consensus mechanism has a similar structure to the widely used Proof of Work (PoW), allowing us to build practical blockchain on it and train neutral networks simultaneously. In experiments, we build a blockchain on top of our proposed PoL consensus mechanism and results show that our PoL works well.