Shared Isolated Sliced and Aggregated Training

A method of ML that allows for Machine Unlearning

Training

1. Break up training data into disjoint shards
2. Break up each shard into slices
3. Train individual constituent models $M_i$ on distinct shards, saving parameters as you train on more and more slices
4. Aggregate consistent model output for final output (Mean, most common output, other methods…)

Untraining

When an unlearning request comes in:

1. Remove data from shard
2. Identify last relevant constituent model parameters
3. Re-aggregate model