Blockchain-Enhanced Machine Learning, animated

The ledger records the process, not the data. ML has three problems a distributed ledger is unusually well suited to: training data poisoned by contributors who still collect a reward, models that leak the data they were trained on, and finished weights that carry no record of what happened to them. Recording every transaction tied to the training process makes tampering evident. Note what stays off chain: raw data remains with its owner, while parameters, transactions and validation results are shared. That is what makes the arrangement privacy-preserving rather than merely public.

Consensus can produce something instead of discarding it. Proof of Work spends electricity on a puzzle whose answer is worthless the moment it is found. Proof of Learning makes the training run itself the work that earns consensus; Proof of Deep Learning extends this to the integrity and authenticity of the resulting model; Proof of Training Quality asks the network to agree not merely that work happened but that the contribution was good. The same electricity now buys a trained model, and inherits one hard question in exchange: how do you verify the work was really done?

Pay for the improvement, not the volume. In the Sharing Updatable Model framework a smart contract, CollaborativeTrainer, accepts data, runs an incentive mechanism and updates the model in one place. The prediction-market reward moves a contributor's balance by b_t = b_t−1 + L(h_t−1, D) − L(h_t, D): the loss on a held-out set before your update minus the loss after it. Submit noise and the term is negative. A companion mechanism (deposit, refund, take) requires a stake when submitting, refunds it after a waiting period if the data holds up, and forfeits it when the model disagrees with your label. In simulation the honest and dishonest populations separate cleanly while accuracy is preserved.

Then read the prototypes honestly. DeepChain was actually built: a Corda prototype on MNIST, with parties uploading local gradients and workers paid through a processing contract. Accuracy improves as more parties join, exactly as the argument predicts; but throughput degrades as the gradient count grows, and total training time climbs with every party added. LearningChain shows the companion trade-off: differential-privacy noise buys real privacy and costs real test accuracy, while its l-nearest aggregation blunts Byzantine workers rather than removing them.

Which is why the survey closes on challenges. Scalability, energy cost, and the need for consensus mechanisms designed for machine learning rather than inherited from cryptocurrency. A review that only sells the idea is worth nothing; the contribution is knowing precisely where the seams are. One of those seams, verifying that a claimed training run really happened, became the subject of the author's Proof-of-Learning research and dissertation.