Communication efficient federated learning
The present disclosure provides efficient communication techniques for transmission of model updates within a machine learning framework, such as, for example, a federated learning framework in which a high-quality centralized model is trained on training data distributed overt a large number of clients each with unreliable network connections and low computational power. In an example federated learning setting, in each of a plurality of rounds, each client independently updates the model based on its local data and communicates the updated model back to the server, where all the client-side updates are used to update a global model. The present disclosure provides systems and methods that reduce communication costs. In particular, the present disclosure provides at least: structured update approaches in which the model update is restricted to be small and sketched update approaches in which the model update is compressed before sending to the server.
1 . A computer-implemented method for communication efficient machine learning, the method comprising:
obtaining, by a client computing device, global values for a set of parameters of a machine-learned model, wherein the machine-learned model is stored on the client computing device;
training, by the client computing device, the machine-learned model stored on the client computing device based at least in part on a local dataset stored on the client computing device to obtain an update matrix that is descriptive of updated values for the set of parameters of the machine-learned model, wherein the local dataset comprises data descriptive of an input received by the client computing device from a user of the client computing device, wherein training, by the client computing device, the machine-learned model based at least in part on the local dataset comprises:
training, by the client computing device, the machine-learned model based at least in part on the local dataset such that updated values are determined only for a pre-selected portion of the set of parameters, and wherein the update matrix is descriptive of only the updated values for the pre-selected portion of the set of parameters; and
communicating, by the client computing device, information descriptive of the update matrix to a server computing device.
2 . The computer-implemented method of claim 1 , wherein the update matrix is restricted to be a low-rank matrix.
3 . The computer-implemented method of claim 1 , wherein:
training, by the client computing device, the machine-learned model based at least in part on the local dataset to obtain the update matrix comprises:
defining, by the client computing device, the update matrix as a product of a first matrix and a second matrix, wherein the first matrix comprises fixed values and the second matrix comprises optimizable variables; and
training, by the client computing device, machine-learned model based at least in part on the local dataset to obtain the second matrix; and
wherein communicating, by the client computing device, information descriptive of the update matrix to the server computing device comprises communicating, by the client computing device, information descriptive of the second matrix to the server computing device.
4 . The computer-implemented method of claim 3 , further comprising, prior to training, by the client computing device, the machine-learned model:
generating, by the client computing device, the first matrix based at least in part on a seed and a pseudo-random number generator, wherein both the client computing device and the server computing device have knowledge of the seed such that the first matrix is reproducible by the server computing device.
5 . The computer-implemented method of claim 1 , wherein the update matrix is restricted to be a sparse matrix.
6 . The computer-implemented method of claim 1 , further comprising, prior to training, by the client computing device, the machine-learned model:
masking, by the client computing device, the pre-selected portion of the set of parameters from the set of parameters via a parameter mask.
7 . The computer-implemented method of claim 6 , wherein masking, by the client computing device, the pre-selected portion of the set of parameters is based at least in part on a seed and a pseudo-random number generator, wherein both the client computing device and the server computing device have knowledge of the seed such that the parameter mask is reproducible by the server computing device.
8 . The computer-implemented method of claim 1 , wherein the update matrix describes the updated values for the set of parameters or respective differences between the updated values and the global values.
9 . A client computing device, comprising:
at least one processor; and
at least one non-transitory computer-readable medium that stores instructions that, when executed by the at least one processor, cause the client computing device to perform operations, the operations comprising:
obtaining global values for a set of parameters of a machine-learned model, wherein the machine-learned model is stored on the client computing device;
training the machine-learned model stored on the client computing device based at least in part on a local dataset stored on the client computing device to obtain an update matrix that is descriptive of updated values for the set of parameters of the machine-learned model, wherein the local dataset comprises data descriptive of an input received by the client computing device from a user of the client computing device, wherein training the machine-learned model based at least in part on the local dataset comprises:
training the machine-learned model based at least in part on the local dataset such that updated values are determined only for a pre-selected portion of the set of parameters, and wherein the update matrix is descriptive of only the updated values for the pre-selected portion of the set of parameters; and
communicating information descriptive of the update matrix to a server computing device.
10 . The client computing device of claim 9 , wherein communicating information descriptive of the update matrix to a server computing device comprises subsampling the update matrix to obtain the information descriptive of the update matrix.
11 . The client computing device of claim 10 , wherein subsampling the update matrix comprises:
generating a parameter mask that specifies a portion of the set of parameters to be sampled; and
subsampling the update matrix according to the parameter mask.
12 . The client computing device of claim 11 , wherein generating the parameter mask comprises generating the parameter mask based at least in part on a seed and a pseudo-random number generator, wherein both the client computing device and the server computing device have knowledge of the seed such that the parameter mask is reproducible by the server computing device.
13 . The client computing device of claim 9 , wherein communicatin information descriptive of the update matrix to a server computing device comprises probabilistically quantizing one or more values included in the update matrix.
14 . The client computing device of claim 9 , wherein communicating information descriptive of the update matrix to a server computing device comprises performing probabilistic binary quantization for one or more values included in the update matrix to change each of the one or more values to a maximum value included in the update matrix or a minimum value included in the update matrix.
15 . The client computing device of claim 9 , wherein communicating information descriptive of the update matrix to a server computing device comprises:
defining a plurality of intervals between a maximum value included in the update matrix and a minimum value included in the update matrix; and
probabilistically changing each of one or more values included in the update matrix to a local interval maximum or a local interval maximum.
16 . The client computing device of claim 9 , wherein communicating information descriptive of the update matrix to a server computing device comprises multiplying a vector of the update matrix by a rotation matrix to obtain a rotated update.
17 . The client computing device of claim 16 , wherein communicating information descriptive of the update matrix to a server computing device further comprises:
probabilistically quantizing one or more values included in the rotated update.
18 . The client computing device of claim 17 , wherein the rotation matrix is a structured rotation matrix that does not require complete generation of the rotation matrix by the client computing device.
19 . At least one non-transitory computer-readable medium that stores instructions that, when executed by a client computing device, cause the client computing device to perform operations, the operations comprising:
obtaining global values for a set of parameters of a machine-learned model, wherein the machine-learned model is stored on the client computing device;
training the machine-learned model stored on the client computing device based at least in part on a local dataset stored on the client computing device to obtain an update matrix that is descriptive of updated values for the set of parameters of the machine-learned model, wherein the local dataset comprises data descriptive of an input received by the client computing device from a user of the client computing device, wherein training the machine-learned model based at least in part on the local dataset comprises:
training the machine-learned model based at least in part on the local dataset such that updated values are determined only for a pre-selected portion of the set of parameters, and wherein the update matrix is descriptive of only the updated values for the pre-selected portion of the set of parameters; and
communicating information descriptive of the update matrix to a server computing device.