Adaptive and evolutionary federated learning system
This application discloses a system and method for federated collaborative machine learning model development using local training datasets that are not shared. An adaptive and evolutionary approach is used to select local training nodes that are most fit from one training round to the next training round to optimize an overall cost and performance function for the federated learning, to cross-over model architecture between local training nodes, and to perform model architecture mutation within local training nodes. The local training nodes are further clustered to account for the inhomogeneity in the local datasets. Such adaptive, evolutionary, and collaborative federated learning thus provides cost-effective and high-performance model development.
1 . A system comprising:
memory circuitry for storing computer instructions;
a network interface circuitry; and
processor circuitry in communication with the network interface circuitry and the memory circuitry, the processor circuitry configured to execute the computer instructions to:
receive sharable data from a plurality of local computation nodes;
cluster the plurality of local computation nodes into a plurality of clusters based on a set of clustering features extracted from the sharable data;
select a subset of local computation nodes from the plurality of local computation nodes as representatives of the plurality of clusters to participate in a collaborative machine learning; and
iteratively provision the collaborative machine learning by the subset of the local computation nodes until a termination condition is met by:
receipt, from the subset of local computation node, sets of model hyper parameters and sets of model metrics associated with machine learning models trained at the subset of local computation nodes using non-sharable datasets of the subset of local computation nodes;
performance of at least one model architectural hyper parameter cross-over of the machine learning models among the subset of local computation nodes to update the sets of model hyper parameters for the subset of local computing nodes, wherein the performance of the at least one model architectural hyper parameter cross-over of the machine learning models among the subset of local computation nodes is limited to intra-cluster cross-over;
elimination of selected local computation nodes of the subset of local computation nodes to obtain a remaining subset of local computation nodes using a multi-dimensional cost/performance function; and
instruction of the remaining subset of local computation nodes to perform a next round of training using the non-sharable datasets based on the updated sets of model hyper parameters.
2 . The system of claim 1 , wherein the non-sharable data from the plurality of local computation nodes comprises a historical time series.
3 . The system of claim 2 , wherein the set of clustering features comprise at least one of a mean, a trough, a variance, a trend, a peak, or a seasonality extracted from the historical time series by the plurality of local computation nodes.
4 . The system of claim 1 , wherein to select the subset of local computation nodes as the representatives of the plurality of clusters to participate in the collaborative machine learning comprises selection, by the processor circuitry, of local computation nodes at or near centroids of the plurality of clusters in a clustering space formed by the set of clustering features.
5 . The system of claim 1 , wherein to iteratively provision the collaborative machine learning further comprises performance, by the processor circuitry, of mutation of the set-sets of model hyper parameters with respect to at least one of the subset of local computation nodes.
6 . The system of claim 5 , wherein the mutation comprises modification, by the processor circuitry, of at least one of the sets of model hyper parameters.
7 . The system of claim 1 , wherein the multi-dimensional cost/performance function comprises at least one of a cost component, a performance component, or a local dataset quality component.
8 . The system of claim 7 , wherein the cost component comprises at least one of a communication cost between the subset of local computing nodes and the system or a computation cost at the subset of local computing nodes.
9 . The system of claim 8 , wherein the computation cost is determined by a complexity of the machine learning models trained at the subset of local computation nodes.
10 . The system of claim 9 , wherein the complexity of the machine learning models is computed based at least one of a number of model layers, a number of model elements in each model layer of the machine learning models trained at the subset of local computation nodes.
11 . The system of claim 7 , wherein the performance component comprises a mean absolute percentage error (MAPE) of the machine learning models.
12 . The system of claim 7 , wherein the local dataset quality component comprises at least one of a time take-taken or a learning rate for training the machine learning models at the subset of local computation nodes.
13 . The system of claim 1 , wherein the sets of model hyper parameters comprise parameters representing architectures of the machine learning models.
14 . The system of claim 1 , wherein to cluster the plurality of local computation nodes into the plurality of clusters, the processor circuitry is configured to:
identify a feature space with orthogonal features;
convert the sharable data into the orthogonal feature space; and
establish clustering dimensions using the orthogonal feature space to delineate the local computation nodes into the plurality of clusters.
15 . A method for adaptive federated machine learning performed by a computer server, comprising:
receiving sharable data from a plurality of local computation nodes;
clustering the plurality of local computation nodes into a plurality of clusters based on a set of clustering features extracted from the sharable data;
selecting a subset of local computation nodes from the plurality of local computation nodes as representatives of the plurality of clusters to participate in a collaborative machine learning; and
iteratively provisioning the collaborative machine learning until a termination condition is met by:
receiving, from the subset of local computation node, sets of model hyper parameters and sets of model metrics associated with machine learning models trained by the subset of local computation nodes using non-sharable datasets of the subset of local computation nodes;
performing at least one model architectural hyper parameter cross-over of the machine learning models among the subset of local computation nodes to update the sets of model hyper parameters for the subset of local computing nodes, wherein performing the at least one model architectural hyper parameter cross-over of the machine learning models among the subset of local computation nodes is limited to intra-cluster cross-over;
performing an elimination of selected local computation nodes of the subset of local computation nodes to obtain a remaining subset of local computation nodes using a multi-dimensional cost/performance function; and
instructing the remaining subset of local computation nodes to perform a next round of training using the non-sharable datasets based on the updated sets of model hyper parameters.
16 . The method of claim 15 , wherein selecting the subset of local computation nodes as the representatives of the plurality of clusters to participate in the collaborative machine learning comprises selecting local computation nodes at or near centroids of the plurality of clusters in a clustering space formed by the set of clustering features.
17 . The method of claim 15 , wherein iteratively provisioning the collaborative machine learning further comprises performing mutation of the sets of model hyper parameters with respect to at least one of the subset of local computation nodes.
18 . The method of claim 15 , wherein the multi-dimensional cost/performance function comprises at least one of a cost component, a performance component, or a local dataset quality component.
19 . The method of claim 18 , wherein:
the cost component comprises at least one of a communication cost between the subset of local computing nodes and a system or a computation cost at the subset of local computing nodes;
the performance component comprises a mean absolute percentage error (MAPE) of the machine learning models; and
the local dataset quality component comprises at least one of a time taken or a learning rate for training the machine learning models at the subset of local computation nodes.
20 . The method of claim 15 , wherein clustering the plurality of local computation nodes into the plurality of clusters comprises:
identifying a feature space with orthogonal features;
converting the sharable data into the orthogonal feature space; and
establishing clustering dimensions using the orthogonal feature space to delineate the local computation nodes into the plurality of clusters.