Storage system of DNN outputs for black box
In an example, an apparatus comprises logic, at least partially including hardware logic, to save one or more outputs of a deep learning neural network in a storage system of an autonomous vehicle and upload the one or more outputs to a remote server. Other embodiments are also disclosed and claimed.
1. A graphics multiprocessor comprising:
an instruction cache to receive a stream of instructions;
an instruction unit to execute the stream of instructions;
a plurality of execution units comprising at least a first type of execution unit having operating at a first speed and a first set of execution resources and a second type of execution unit operating at a second speed, different from the first speed, and having a second set of execution resources, different from the first set of execution resources;
a shared memory communicatively coupled to the plurality of execution units; and
a processor to:
scan a field of view with a low resolution camera of a vehicle;
analyze one or more backpropagated gradient maps from low resolution image data collected by the low resolution camera;
identify, from the low resolution image data and based on the analysis of the one or more backpropagated gradient maps, at least one region of interest in the field of view;
collect high resolution images from the at least one region of interest in the field of view using a high resolution camera of the vehicle;
process the low resolution image data collected by the low resolution camera in the first type of execution unit and the high resolution images collected by the high resolution camera in the second type of execution unit; and
upload at least a portion of the low resolution image data and the high resolution images to a datacenter for inclusion in a neural network model, wherein an approximation of an original image is re-created, based on the neural network model, using a pre-trained Generative Adversarial Network and used for unsupervised adaptation.
2. The graphics multiprocessor of claim 1 , the processor to:
implement a dropout regularization technique to preserve privacy of compute operations.
3. The graphics multiprocessor of claim 1 , the processor to:
save data from an environment surrounding the vehicle.
4. The graphics multiprocessor of claim 3 , the processor to:
save an output of a fully connected layer of a neural network communicatively coupled to at least one of the low resolution camera or the high resolution camera.
5. The graphics multiprocessor of claim 4 , the processor to:
train a neural network to reconstruct one or more original images collected by at least one of the low resolution camera or the high resolution camera based on the output of the fully connected layer of the neural network.
6. An electronic device, comprising:
a display; and
a graphics multiprocessor comprising:
an instruction cache to receive a stream of instructions;
an instruction unit to execute the stream of instructions;
a plurality of execution units comprising at least a first type of execution unit having operating at a first speed and a first set of execution resources and a second type of execution unit operating at a second speed, different from the first speed, and having a second set of execution resources, different from the first set of execution resources;
a shared memory communicatively coupled to the plurality of execution units; and
a processor to:
scan a field of view with a low resolution camera of a vehicle;
analyze one or more backpropagated gradient maps from low resolution image data collected by the low resolution camera;
identify, from the low resolution image data and based on the analysis of the one or more backpropagated gradient maps, at least one region of interest in the field of view;
collect high resolution images from the at least one region of interest in the field of view
using a high resolution camera of the vehicle; and
process the low resolution image data collected by the low resolution camera in the first type of execution unit and the high resolution images collected by the high resolution camera in the second type of execution unit; and
upload at least a portion of the low resolution image data and the high resolution images to a datacenter for inclusion in a neural network model, wherein an approximation of an original image is re-created, based on the neural network model, using a pre-trained Generative Adversarial Network and used for unsupervised adaptation.
7. The electronic device of claim 6 , the processor to:
implement a dropout regularization technique to preserve privacy of compute operations.
8. The electronic device of claim 6 , the processor to:
save data from an environment surrounding the vehicle.
9. The electronic device of claim 8 , the processor to:
save an output of a fully connected layer of a neural network communicatively coupled to at least one of the low resolution camera or the high resolution camera.
10. The electronic device of claim 9 , the processor to:
train a neural network to reconstruct one or more original images collected by at least one of the low resolution camera or the high resolution camera based on the output of the fully connected layer of the neural network.