Coordinated multiple worker node casual inference framework

A system including:

• network interface circuitry configured to:
• establish multiple communication links with multiple worker nodes, the multiple worker nodes each configured to store a corresponding dataset partition of a global dataset;
• for each model iteration:
• send a respective batch indicator to a respective group of the multiple worker nodes, corresponding dataset partitions of each worker node of the respective group including a portion of a respective batch selected from the global dataset;
• receive a corresponding respective model convergence indicator from each worker node of the respective group;
• and
• send, to each of the multiple worker nodes, a respective model update parameter for the model iteration;
• and
• coordination circuitry in data communication with the network interface circuitry, the coordination circuitry configured to:
• for each model iteration:
• select the respective batch from the global dataset;
• generate the respective batch indicator based on the respective batch selected from the global dataset;
• and
• based on one or more respective model convergence indicators received from the respective group, generate the respective model update parameter.

A method including:

• at network interface circuitry:
• establishing multiple communication links with multiple worker nodes, the multiple worker nodes each configured to store a corresponding dataset partition of a global dataset;
• for each model iteration:
• sending a respective batch indicator to a respective group of the multiple worker nodes, corresponding dataset partitions of each worker node of the respective group including a portion of a respective batch selected from the global dataset;
• receiving a corresponding respective model convergence indicator from each worker node of the respective group;
• and
• sending, to each of the multiple worker nodes, a respective model update parameter for the model iteration;
• and
• at coordination circuitry in data communication with the network interface circuitry:
• for each model iteration:
• selecting the respective batch from the global dataset;
• generating the respective batch indicator based on the respective batch selected from the global dataset;
• and
• based on one or more respective model convergence indicators received from the respective group, generating the respective model update parameter.