Method and system for latency reduction in serialized machine learning systems
The present disclosure provides methods for reducing prediction latency. The method comprises: (a) receiving, at a first computing system, a service request including feature data for performing a prediction; (b) the first computing system making a first intermediary prediction using a first machine learning model based on a first set of data and the feature data; (c) concurrent with (b) a second computing system making a second intermediary prediction using a second machine learning model in response to a prediction request received from the first computing system, and the second intermediary prediction is generated based on a second set of data and feature data included in the prediction request; (d) making an enhanced prediction by the first computing system based on the first intermediary prediction and the second intermediary prediction, and the second set of data and the second machine learning model are inaccessible to the first computing system.
1 . A method for reducing prediction latency and maintaining data privacy in a serialized machine learning system, the method comprising:
(I) during an inference stage, running partial inferences by a first trained machine learning model and a second trained machine learning model concurrently, wherein running the partial inferences comprises:
(a) a first computing system making a first intermediary inference using the first trained machine learning model by taking in an input data comprising (i) a first set of data accessible to the first computing system and (ii) data in a request, wherein an inference time for making the first intermediary inference by the first trained machine learning model is T 1 ;
(b) concurrently with (a) and during the same inference stage, a second computing system making a second intermediary inference using the second trained machine learning model, wherein the second intermediary prediction is generated by taking in an input data comprising (i) a second set of data and (ii) the data in the request, and wherein an inference time for making the second intermediary inference by the second trained machine learning model is T 2 ;
and
(II) during the same inference stage, upon determining T 1 >T 2 , transmitting the second intermediary inference from the second computing system to the first computing system and joining, by an integration point, the first intermediary inference and the second intermediary inference based at least in part on a common feature identifier in the data of the request for the first trained machine learning model to make a final inference, wherein an inference time for making the final inference by the first trained machine learning model is T 3 and wherein the overall inference time is T 1 +T 3 which is shorter than the sum of T 1 and T 2 thereby reducing prediction latency or,
upon determining T 1 <T 2 , transmitting the first intermediary inference from the first computing system to the second computing system and joining, by the integration point, the first intermediary inference and the second intermediary inference based at least in part on the common feature identifier in the data of the request for the second trained machine learning model to make a final inference, and wherein the overall inference time is T 2 +T 3 which is shorter than the sum of T 1 and T 2 thereby reducing prediction latency,
wherein model weights of the first trained machine learning model and the first set of data are secured on the first computing system and isolated from the second computing system and wherein model weights of the second trained machine learning model and the second set of data are secured on the second computing system and isolated from the first computing system to preserve privacy of the first set of data and the second set of data, and the model weights of the first trained machine learning model and the second trained machine learning model respectively.
2 . The method of claim 1 , wherein the request is for performing a prediction and wherein the request is received by the first computing system.
3 . The method of claim 1 , wherein the second computing system makes the second intermediary inference in response to a prediction request transmitted from the first computing system, and wherein the prediction request includes the data in the request.
4 . The method of claim 1 , wherein during a training stage, model weights of the second machine learning model are optimized using supervised learning and wherein a labeled dataset for training the second machine learning model comprises model output data generated by the first trained machine learning model.
5 . The method of claim 4 , wherein the first trained machine learning model is a neural network.
6 . The method of claim 5 , wherein the model output data generate by the first trained machine learning model comprises output of a top layer of the neural network.
7 . The method of claim 5 , wherein the model output data generate by the first trained machine learning model comprises a prediction value.
8 . The method of claim 1 , wherein a model output function of the first trained machine learning model or the second trained machine learning model is a simple arithmetic transformation.
9 . The method of claim 1 , wherein T 3 is shorter than T 1 or T 2 .
10 . A serialized machine learning system comprising:
(a) a first computing system comprising a processor and a first trained machine learning model, wherein the first computing system is configured to, during an inference stage, make a first intermediary inference using the first trained machine learning model based on (i) a first set of data accessible to the first computing system and (ii) data in a request, wherein an inference time for making the first intermediary inference is T 1 ;
(b) a second computing system comprising a processor and a second trained machine learning model, wherein the second computing system is configured to, during the same inference stage, make a second intermediary inference concurrently with the first computing system making the first intermediary inference, and wherein the second intermediary inference is made using the second trained machine learning model based on (i) a second set of data and (ii) the data in the request, and wherein an inference time for making the second intermediary inference is T 2 ; and
(c) an integration component configured to:
during the same inference stage, upon determining T 1 >T 2 , transmit the second intermediary inference from the second computing system to the first computing system and join the first intermediary inference and the second intermediary inference based at least in part on a common feature identifier in the data of the request for the first trained machine learning model to make a final inference, wherein an inference time for making the final inference by the first trained machine learning model is T 3 and wherein the overall inference time is T 1 +T 3 which is shorter than the sum of T 1 and T 2 thereby reducing prediction latency, or
upon determining T 1 <T 2 , transmit the first intermediary inference from the first computing system to the second computing system and join the first intermediary inference and the second intermediary inference based at least in part on a common feature identifier in the data of the request for the second trained machine learning model to make a final inference, and wherein the overall inference time is T 2 +T 3 which is shorter than the sum of T 1 and T 2 thereby reducing prediction latency,
wherein the first trained machine learning model and the first set of data are secured on the first computing system and isolated from the second computing system and wherein the second machine learning model and the second set of data are secured on the second computing system and isolated from the first computing system to preserve privacy of the first set of data and the second set of data respectively.
11 . The serialized machine learning system of claim 10 , wherein the request is for performing a prediction and wherein the request is received by the first computing system.
12 . The serialized machine learning system of claim 10 , wherein the first computing system is configured to further send a prediction request to the second computing system, and wherein the second computing system makes the second intermediary inference in response to receiving the prediction request.
13 . The serialized machine learning system of claim 10 , wherein during a training stage, model weights of the second machine learning model are optimized using supervised learning and wherein a labeled dataset for training the second machine learning model comprises model output data generated by the first trained machine learning model.
14 . The serialized machine learning system of claim 13 , wherein the first trained machine learning model is a neural network.
15 . The serialized machine learning system of claim 14 , wherein the model output data generate by the first trained machine learning model comprises output of a top layer of the neural network.
16 . The serialized machine learning system of claim 14 , wherein the model output data generate by the first trained machine learning model comprises a prediction value.
17 . The serialized machine learning system of claim 10 , wherein a model output function of the first trained machine learning model or the second trained machine learning model is a simple arithmetic transformation.
18 . The serialized machine learning system of claim 10 , wherein T 3 is shorter than T 1 or T 2 .