NEURAL NETWORK APPLICATIONS IN RESOURCE CONSTRAINED ENVIRONMENTS
Systems and methods are disclosed for applying neural networks in resource-constrained environments. A system may include a sensor located in a resource-constrained environment configured to generate first sensor data and second sensor data of the resource-constrained environment. The system may also include a first computing device not located in the resource-constrained environment configured to produce a neural network structure based on the first sensor data. The system may also include a second computing device configured to determine a state of the resource-constrained environment based on input of the second sensor data to the neural network structure. The system may also include a controller located in the resource-constrained environment configured to control a device in the resource-constrained environment based on the state of the resource-constrained environment determined by the second computing device. The second computing device may be further configured to calculate an activation area for the neural network structure.
1 . A system comprising:
one or more sensors located in an environment configured to generate sensor data of the environment;
one or more computing devices configured to receive a neural network structure, wherein the one or more computing devices are configured to receive the sensor data, wherein the one or more computing devices are configured to determine a state of the environment based on input of the sensor data to the neural network structure; and
a controller configured to control a device in the environment based on a result of the determination of the state of the environment by the one or more computing devices,
wherein the one or more computing devices are configured to calculate an activation area for the neural network structure,
wherein the one or more computing devices are configured to calculate the activation area for the neural network structure, at least in part, by providing second sensor data as input to the neural network structure,
wherein the second sensor data is formed by placing a first mask at a first location in the sensor data,
wherein the one or more computing devices are configured to calculate the activation area for the neural network structure, at least in part, by providing third sensor data as input to the neural network structure,
wherein the third sensor data is formed by placing a second mask at a second location in the sensor data,
wherein the one or more computing devices are configured to calculate the activation area for the neural network structure, at least in part, by providing fourth sensor data as input to the neural network structure,
wherein the fourth sensor data is formed by placing a third mask at a third location in the sensor data,
wherein the one or more computing devices are configured to calculate the activation area for the neural network structure, at least in part, by comparing a first result of providing the sensor data as input to the neural network structure to a second result of providing the second sensor data as input to the neural network structure,
wherein the one or more computing devices are configured to calculate the activation area for the neural network structure, at least in part, by comparing the first result of providing the sensor data as input to the neural network structure to a third result of providing the third sensor data as input to the neural network structure, and
wherein the one or more computing devices are configured to calculate the activation area for the neural network structure, at least in part, by comparing the first result of providing the sensor data as input to the neural network structure to a fourth result of providing the fourth sensor data as input to the neural network structure.
2 . The system of claim 1 , wherein the one or more computing devices comprise:
a local computing device located in the environment; and
a remote computing device not located in the environment.
3 . The system of claim 2 ,
wherein the remote computing device is configured to generate the neural network structure based on second sensor data of the environment, and
wherein the local computing device is configured to determine the state of the environment.
4 . The system of claim 3 , wherein the local computing device is configured to calculate the activation area for the neural network structure.
5 . The system of claim 3 , wherein the remote computing device is configured to calculate the activation area for the neural network structure.
6 . The system of claim 1 ,
wherein the environment is an automobile, and
wherein the sensor data comprises images of an interior of the automobile.
7 . The system of claim 6 , wherein the controller configured to control the device in the environment in order to control operation of the automobile in a self-driving mode.
8 . The system of claim 1 ,
wherein the one or more computing devices are configured to calculate the activation area for a first prediction class of the neural network structure, and
wherein the one or more computing devices are configured to calculate a second activation area for a second prediction class of the neural network structure.
9 . The system of claim 8 ,
wherein the one or more computing devices are configured to compare the activation area to an expected activation area for the first prediction class; and
wherein the one or more computing devices are configured to compare the second activation area to an expected activation area for the second prediction class.
10 . The system of claim 9 , wherein the one or more computing devices are configured, based at least in part on a result of comparing the activation area to the expected activation area for the first prediction class, to instruct the one or more sensors to generate additional sensor data for the first prediction class.
11 . The system of claim 1 , wherein the one or more computing devices are configured to compare the first result to the second result, at least in part, by determining if a prediction class provided as part of the second result is different from a prediction class provided as part of the first result.
12 . The system of claim 1 , wherein the one or more computing devices are configured to compare the first result to the second result, at least in part, by determining if a second prediction confidence level provided as part of the second result is different from a first prediction confidence level provided as part of the first result.
13 . The system of claim 12 , wherein the one or more computing devices are configured to determine if the second prediction confidence level is different from the first prediction confidence level, at least in part, by determining if the difference between the second prediction confidence level and the first prediction confidence level is greater than a predefined threshold value.
14 . The system of claim 1 ,
wherein the one or more computing devices are configured to determine an effectiveness value for the neural network structure based at least in part on the calculated activation area, and
wherein the one or more computing devices are configured, based at least in part on the determined effectiveness value, to generate a second neural network structure.
15 . A method comprising:
generating sensor data of an environment;
receiving a neural network structure;
determining a state of the environment based on input of the sensor data to the neural network structure;
controlling, based on a result of the determining the state of the environment, a device in the environment; and
calculating an activation area for the neural network structure,
wherein calculating the activation area for the neural network structure comprises providing second sensor data as input to the neural network structure,
wherein the second sensor data is formed by placing a first mask at a first location in the sensor data,
wherein calculating the activation area for the neural network structure comprises providing third sensor data as input to the neural network structure,
wherein the third sensor data is formed by placing a second mask at a second location in the sensor data,
wherein calculating the activation area for the neural network structure comprises providing fourth sensor data as input to the neural network structure,
wherein the fourth sensor data is formed by placing a third mask at a third location in the sensor data,
wherein calculating the activation area for the neural network structure comprises comparing a first result of providing the sensor data as input to the neural network structure to a second result of providing the second sensor data as input to the neural network structure,
wherein calculating the activation area for the neural network structure comprises comparing the first result of providing the sensor data as input to the neural network structure to a third result of providing the third sensor data as input to the neural network structure, and
wherein calculating the activation area for the neural network structure comprises comparing the first result of providing the sensor data as input to the neural network structure to a fourth result of providing the fourth sensor data as input to the neural network structure.
16 . The method of claim 15 , wherein calculating the activation area for the neural network structure comprises calculating the activation area for a first prediction class of the neural network structure, the method further comprising:
calculating a second activation area for a second prediction class of the neural network structure.
17 . The method of claim 16 , further comprising:
comparing the activation area to an expected activation area for the first prediction class; and
comparing the second activation area to an expected activation area for the second prediction class.
18 . The method of claim 17 , further comprising:
generating, based at least in part on a result of the comparing the activation area to the expected activation area for the first prediction class, additional sensor data for the first prediction class.
19 . The method of claim 15 , wherein comparing the first result to the second result comprises determining if a prediction class provided as part of the second result is different from a prediction class provided as part of the first result.
20 . The method of claim 15 , wherein comparing the first result to the second result comprises determining if a second prediction confidence level provided as part of the second result is different from a first prediction confidence level provided as part of the first result.