Consistent distributed edge models via controlled dropout model training
Consistent distributed edge models via controlled dropout model training are described. According to some embodiments, a machine learning model is trained using multiple sensor data streams. During the training process, ones of the sensor data streams are dropped to cause the model to be generated to be robust in that it can tolerate missing input data from sensor data sources yet still maintain high accuracy. The model can be deployed to multiple sensor devices within an environment. The sensor devices generate sensor data and exchange a variety of types of data to ultimately result in a distributed, consistent model result being generated that remains accurate despite communication faults that may occur between ones of the sensor devices.
1. A system comprising:
a first camera device and a plurality of other camera devices operating in an environment, wherein the first camera device comprises a sensor component, one or more processors, and a non-transitory computer-readable storage medium storing a model and having instructions which, when executed by the one or more processors, cause the first camera device to:
generate, based on sensor data generated by the sensor component, a feature map through use of a first part of the model, the feature map representing the sensor data of the first camera device, wherein a second part of the model was trained using a plurality of training data streams in which a subset of each of the plurality of training data streams was excluded to simulate a failure of one or more but not all of the plurality of other camera devices;
transmit, to each of the plurality of other camera devices, the feature map representing the sensor data of the first camera device;
receive, from one or more but not all of the plurality of other camera devices, one or more feature maps generated by the one or more other camera devices, each feature map representing sensor data of the corresponding camera device; and
generate, using the second part of the model and the feature map representing the sensor data of the first camera device and the one or more feature maps generated by the one or more other camera devices, a local hypothesis value that represents a state of the environment.
2. The system of claim 1 , wherein the instructions, when executed by the one or more processors, further cause the first camera device to:
transmit, to each of the plurality of other camera devices, the local hypothesis value;
receive, from each of one or more of the plurality of other camera devices, one or more remote hypothesis values each representing a state of the environment as determined by the corresponding camera device; and
generate an ensemble hypothesis value based on the local hypothesis value and the one or more remote hypothesis values.
3. The system of claim 2 , wherein the instructions, when executed by the one or more processors, further cause the first camera device to:
determine that a condition exists within the environment based on an analysis of the ensemble hypothesis value; and
transmit a notification message to another device.
4. A system comprising:
a first sensor device and a second one or more sensor devices operating in a common environment, wherein the first sensor device comprises a sensor component, one or more processors, and a non-transitory computer-readable storage medium storing a model and having instructions which, when executed by the one or more processors, cause the first sensor device to:
receive, from one or more but not all of the second one or more sensor devices, a first one or more messages, wherein each of the first one or more messages includes a first intermediate representation of sensor data generated by an originating sensor device; and
generate, using the model and the one or more first intermediate representations and a second intermediate representation of sensor data generated by the sensor component, a local hypothesis value representing a state of the common environment, wherein the model includes a first part and a second part, wherein the second part was trained using one or more training data streams in which a subset of each of the one or more training data streams was excluded to simulate a failure of one or more but not all of the second one or more sensor devices.
5. The system of claim 4 , wherein the instructions, when executed by the one or more processors, further cause the first sensor device to:
receive, from one or more of the second one or more sensor devices, one or more remote hypothesis values each representing a state of the common environment; and
generate, based on the local hypothesis value and the one or more remote hypothesis values, an ensemble hypothesis value representing a state of the common environment.
6. The system of claim 5 , wherein the instructions, when executed by the one or more processors, further cause the first sensor device to:
determine that a condition exists within the common environment based on an analysis of the ensemble hypothesis value; and
perform an action responsive to the determination that the condition exists.
7. The system of claim 4 , wherein the instructions, when executed by the one or more processors, further cause the first sensor device to:
generate, based on the sensor data generated by the sensor component and the first part of the model, the second intermediate representation of the sensor data.
8. The system of claim 4 , wherein:
the first intermediate representation of each of the first one or more messages is the sensor data generated by the originating sensor device, and
the second intermediate representation comprises the sensor data generated by the sensor component.
9. The system of claim 4 , wherein the model comprises a multilayer neural network.
10. The system of claim 9 , wherein:
the multilayer neural network comprises a convolutional neural network (CNN) or a recurrent neural network (RNN).
11. The system of claim 4 , wherein the first sensor device is a camera device and the sensor component comprises an image signal processor (ISP).
12. The system of claim 11 , wherein the local hypothesis value representing the state of the common environment identifies one or more objects within the common environment and locations of the one or more objects within the common environment.
13. The system of claim 12 , wherein the instructions, when executed by the one or more processors, further cause the first sensor device to:
generate an ensemble hypothesis value representing a state of the common environment based on the local hypothesis value and one or more remote hypothesis values; and
determine or modify a navigational path through the common environment based on the ensemble hypothesis value.
14. A computer-implemented method comprising:
receiving, by a first sensor device from one or more but not all of a second one or more sensor devices operating in a common environment, a first one or more messages, wherein each of the first one or more messages includes a first intermediate representation of sensor data generated by an originating sensor device, wherein the first sensor device comprises a sensor component; and
generating, using a machine learning model and the one or more first intermediate representations and a second intermediate representation of sensor data generated by the sensor component, a local hypothesis value representing a state of the common environment, wherein the machine learning model includes a first part and a second part, wherein the second part was trained using one or more training data streams in which a subset of each of the one or more training data streams were excluded to simulate a failure of one or more but not all of the second one or more sensor devices.
15. The computer-implemented method of claim 14 , wherein the machine learning model comprises a multilayer neural network.
16. The computer-implemented method of claim 15 , wherein the machine learning model comprises a convolutional neural network.
17. The computer-implemented method of claim 14 , further comprising:
receiving, from one or more of the second one or more sensor devices, one or more remote hypothesis values each representing a state of the common environment; and
generating, based on the local hypothesis value and the one or more remote hypothesis values, an ensemble hypothesis value representing a state of the common environment.
18. The computer-implemented method of claim 14 , wherein the machine learning model comprises a recurrent neural network.
19. The computer-implemented method of claim 17 , further comprising:
determining that a condition exists within the common environment based on an analysis of the ensemble hypothesis value; and
performing an action responsive to the determination that the condition exists.
20. The computer-implemented method of claim 14 , wherein:
the first intermediate representation of each of the first one or more messages is the sensor data generated by an originating sensor device, and
the second intermediate representation comprises the sensor data generated by the sensor component.