SYSTEMS AND METHODS FOR FAULT TOLERANCE RECOVER DURING TRAINING OF A MODEL OF A CLASSIFIER USING A DISTRIBUTED SYSTEM

A distributed system for training a classifier is provided. The system comprises machine learning (ML) workers and a parameter server (PS). The PS is configured for parallel processing to provide the model to each of the ML workers, receive model updates from each of the ML workers, and iteratively update the model using each model update. The PS contains gradient datasets associated with a respective ML worker, for storing a model-update-identification (delta-M-ID) indicative of the computed model update and the respective model update, a global dataset that stores, the delta-M-ID, an identification of the ML worker (ML-worker-ID) that computed the model update, and a model version that marks a new model in PS that is computed from merging the model update with a previous model in PS; and a model download dataset that stores the ML-worker-ID and the model version of each transmitted model.