Distributed image distillation for private and efficient event prediction in logistics
One example method includes, in an environment having a first near-edge node and a second near-edge node, each of which is operable to communicate with a respective set of edge nodes and with a central node: instantiating, by the central node, a dataset distillation process, wherein the dataset includes data collected by the edge nodes, and the data remains at the near-edge nodes and is not accessed by the central node; performing the dataset distillation process to create a distilled dataset; pre-training a machine learning model using the distilled dataset; comparing the pre-trained machine learning model to one or more other pre-trained machine learning models; and deploying, to the edge nodes, the pre-trained learning model that has been determined, based on the comparing, to provide the best performance as among the pre-trained machine learning models that have been compared.
1 . A method for distributed image distillation for private and efficient event prediction in an environment, comprising:
in an environment that includes a first near-edge node and a second near-edge node, each of which is operable to communicate with a respective set of edge nodes and with a central node, and each of the edge nodes comprises a respective camera operable to gather data about the environment:
instantiating, by the central node, a distributed dataset distillation process, wherein the dataset includes image data collected by the cameras of the edge nodes, wherein each near-edge node maintains image data collected by its respective edge nodes;
performing the dataset distillation process to create a distilled dataset that is based on the image data, wherein the distilled dataset aggregates information derived from image data maintained at multiple near-edge nodes;
pre-training an event detection machine learning (ML) model using the distilled dataset, at the central node and prior to deployment of the event detection ML model to any edge node, and the event detection ML model enables one of the edge nodes to use a feed from the camera at that edge node to detect an event occurring in the environment, including detection of an event represented in image data contributed by other near-edge nodes;
comparing the event detection ML model to one or more other pre-trained event detection ML models that have been trained respective distilled datasets generated by prior executions of the distributed dataset distillation process; and
deploying, to the edge nodes, the event detection ML model that has been determined, based on the comparing, to provide the best performance as among the event detection ML models that have been compared.
2 . The method as recited in claim 1 , wherein each camera is associated with a respective piece of mobile equipment, and the image data includes data regarding operation of the mobile equipment in the environment.
3 . The method as recited in claim 2 , wherein, at each edge node, the camera and the event detection ML model deployed at that edge node cooperate to predict and/or detect occurrence, in the environment, of an event involving the respective piece of mobile equipment.
4 . The method as recited in claim 1 , wherein the image data comprise video data of the environment.
5 . The method as recited in claim 3 , wherein prediction and/or detection of the occurrence of the event involving the respective piece of mobile equipment is performed in real time while the piece of mobile equipment is operating.
6 . The method as recited in claim 1 , wherein the image data collected by the edge nodes resides at the near-edge nodes when the data distillation process is instantiated.
7 . The method as recited in claim 1 , wherein determining the pre-trained learning model that has the best performance comprises:
calculating, by each of the near-edge nodes, respective validation metrics for each of the pre-trained machine learning models; and
computing, by the central node based on the validation metrics, an aggregation function to determine a respective number of each of the pre-trained machine learning models, wherein the pre-trained machine learning model with the best performance has the highest number.
8 . The method as recited in claim 1 , further comprising fine-tuning, at one of the near-edge nodes, the best performing machine learning model.
9 . The method as recited in claim 8 , wherein, after the fine-tuning, the best performing machine learning model is deployed by the near-edge node to the edge nodes with which that near-edge node is operable to communicate.
10 . A non-transitory storage medium having stored therein instructions that are executable by one or more hardware processors to perform a method for distributed image distillation for private and efficient event prediction in an environment, that includes operations comprising:
in an environment that includes a first near-edge node and a second near-edge node, each of which is operable to communicate with a respective set of edge nodes and with a central node, and each of the edge nodes comprises a respective camera operable to gather data about the environment:
instantiating, by the central node, a distributed dataset distillation process, wherein the dataset includes image data collected by the cameras of the edge nodes, wherein each near-edge node maintains image data collected by its respective edge nodes;
performing the dataset distillation process to create a distilled dataset that is based on the image data, wherein the distilled dataset aggregates information derived from image data maintained at multiple near-edge nodes;
pre-training an event detection machine learning (ML) model using the distilled dataset, at the central node and prior to deployment of the event detection ML model to any edge node, and the event detection ML model enables one of the edge nodes to use a feed from the camera at that edge node to detect an event occurring in the environment, including detection of an event represented in image data contributed by other near-edge nodes;
comparing the event detection ML model to one or more other pre-trained event detection ML models that have been trained respective distilled datasets generated by prior executions of the distributed dataset distillation process; and
deploying, to the edge nodes, the event detection ML model that has been determined, based on the comparing, to provide the best performance as among the event detection ML models that have been compared.
11 . The non-transitory storage medium as recited in claim 10 , wherein each camera is associated with a respective piece of mobile equipment, and the image data includes data regarding operation of the mobile equipment in the environment.
12 . The non-transitory storage medium as recited in claim 11 , wherein, at each edge node, the camera and the event detection ML model deployed at that edge node cooperate to predict and/or detect occurrence, in the environment, of an event involving the respective piece of mobile equipment.
13 . The non-transitory storage medium as recited in claim 10 , wherein the image data comprise video data of the environment.
14 . The non-transitory storage medium as recited in claim 12 , wherein prediction and/or detection of the occurrence of the event involving the respective piece of mobile equipment is performed in real time while the piece of mobile equipment is operating.
15 . The non-transitory storage medium as recited in claim 10 , wherein the image data collected by the edge nodes resides at the near-edge nodes when the data distillation process is instantiated.
16 . The non-transitory storage medium as recited in claim 10 , wherein determining the pre-trained learning model that has the best performance comprises:
calculating, by each of the near-edge nodes, respective validation metrics for each of the pre-trained machine learning models; and
computing, by the central node based on the validation metrics, an aggregation function to determine a respective number of each of the pre-trained machine learning models, wherein the pre-trained machine learning model with the best performance has the highest number.
17 . The non-transitory storage medium as recited in claim 10 , wherein the operations further comprise fine-tuning, at one of the near-edge nodes, the best performing machine learning model.
18 . The non-transitory storage medium as recited in claim 17 , wherein, after the fine-tuning, the best performing machine learning model is deployed by the near-edge node to the edge nodes with which that near-edge node is operable to communicate.