How collective learning works
A Colearn experiment begins when a group of entities, referred to as learners, decide on a model architecture and begin learning. Together they will train a single global model. The goal is to train a model that performs better than any of the learners can produce by training on their private data set.
How Training Works
Training occurs in rounds; during each round the learners attempt to improve the performance of the global shared
To do so each round an update of the global model (for example new set of weights in a neural network) is proposed.
The learners then validate the update and decide if the new model is better than the current global model.
If enough learners approve the update then the global model is updated. After an update is approved or rejected a new round begins.
The detailed steps of a round updating a global model M are as follows:
- One of the learners is selected and proposes a new updated model M'
- The rest of the learners validate M'
- If M' has better performance than M against their private data set then the learner votes to approve
- If not, the learner votes to reject
- The total votes are tallied
- If more than some threshold (typically 50%) of learners approve then M' becomes the new global model. If not, M continues to be the global model
- A new round begins.
By using a decentralized ledger (a blockchain) this learning process can be run in a completely decentralized, secure and auditable way. Further security can be provided by using differential privacy to avoid exposing your private data set when generating an update.
Learning algorithms that work for collective learning
Collective learning is not just for neural networks; any learning algorithm that can be trained on subsets of the data and which can use the results of previous training rounds as the basis for subsequent rounds can be used. Neural networks fit both these constraints: training can be done on mini-batches of data and each training step uses the weights of the previous training step as its starting point. More generally, any model that is trained using mini-batch stochastic gradient descent is fine. Other algorithms can be made to work with collective learning as well. For example, a random forest can be trained iteratively by having each learner add new trees (see example in mli_random_forest_iris.py). For more discussion, see here.
The driver implements the voting protocol, so it handles selecting a learner to train, sending the update out for voting, calculating the vote and accepting or declining the update. Here we have a very minimal driver that doesn't use networking or a blockchain. Eventually the driver will be a smart contract. This is the code that implements one round of voting:
def run_one_round(round_index: int, learners: Sequence[MachineLearningInterface], vote_threshold=0.5): proposer = round_index % len(learners) new_weights = learners[proposer].mli_propose_weights() prop_weights_list = [ln.mli_test_weights(new_weights) for ln in learners] approves = sum(1 if v.vote else 0 for v in prop_weights_list) vote = False if approves >= len(learners) * vote_threshold: vote = True for j, learner in enumerate(learners): learner.mli_accept_weights(prop_weights_list[j]) return prop_weights_list, vote
The Machine Learning Interface
propose_weightscauses the model to do some training and then return a new set of weights that are proposed to the other learners. This method shouldn't change the current weights of the model - that only happens when
test_weights- the models takes some new weights and returns a vote on whether the new weights are an improvement. As with propose_weights, this shouldn't change the current weights of the model - that only happens when
accept_weights- the models accepts some weights that have been voted on and approved by the set of learners. The old weighs of the model are discarded and replaced by the new weights.
current_weightsshould return the current weights of the model.
For more details about directly implementing the machine learning interface see the tutorial here