US Patent Application 18202459. DISTRIBUTED MODEL TRAINING WITH COLLABORATION WEIGHTS FOR PRIVATE DATA SETS simplified abstract
Contents
DISTRIBUTED MODEL TRAINING WITH COLLABORATION WEIGHTS FOR PRIVATE DATA SETS
Organization Name
Inventor(s)
Jesse Cole Cresswell of Toronto (CA)
Brendan Leigh Ross of Toronto (CA)
Ka Ho Yenson Lau of TORONTO (CA)
DISTRIBUTED MODEL TRAINING WITH COLLABORATION WEIGHTS FOR PRIVATE DATA SETS - A simplified explanation of the abstract
This abstract first appeared for US patent application 18202459 titled 'DISTRIBUTED MODEL TRAINING WITH COLLABORATION WEIGHTS FOR PRIVATE DATA SETS
Simplified Explanation
The patent application describes a system for training models without sharing private data sets.
- Private data sets learn client weights for computer models during training.
- Inference for a specific data set is determined by a combination of model parameters based on client weights.
- Client weights are updated based on how well sampled models represent the private data set.
- Gradients are calculated for each sampled model and can be weighted according to the client weight for that model.
- This increases the contribution of a private data set to the model parameters that are more relevant to that data set.
Original Abstract Submitted
Model training systems collaborate on model training without revealing respective private data sets. Each private data set learns a set of client weights for a set of computer models that are also learned during training. Inference for a particular private data set is determined as a mixture of the computer model parameters according to the client weights. During training, at each iteration, the client weights are updated in one step based on how well sampled models represent the private data set. In another step, gradients are determined for each sampled model and may be weighed according to the client weight for that model, relatively increasing the gradient contribution of a private data set for model parameters that correspond more highly to that private data set.