Object recognition with reduced neural network weight precision
A client device configured with a neural network includes a processor, a memory, a user interface, a communications interface, a power supply and an input device, wherein the memory includes a trained neural network received from a server system that has trained and configured the neural network for the client device. A server system and a method of training a neural network are disclosed.
1. A device, comprising:
a processor; and
a memory,
the processor and the memory being configured as a neural network comprising:
at least one layer comprising an input and an output, the layer configured to receive an input feature map at the input and output an output feature map at the output, at least one of the input feature map and the output feature map comprising at least one first weight kernel value that has been quantized by a unitary quantizing operation to reduce a number of bits of the first weight kernel value from a first predetermined number of bits to a second predetermined number of bits that is less than the first predetermined number of bits without changing a dimension of the feature map corresponding to the quantized first weight kernel value.
2. The device of claim 1 , wherein the neural network further comprises at least one interim feature map between the input and the output of the layer, the interim feature map comprising at least one second weight kernel value that has been quantized by the unitary quantizing operation to reduce a number of bits of the second weight kernel value from a third predetermined number of bits to a fourth predetermined number of bits that is less than the third predetermined number of bits without changing a dimension of the interim feature map corresponding to the quantized second weight kernel value.
3. The device of claim 2 , wherein the second weight kernel value is further quantized by one of the unitary quantizing operation or a supervised iterative quantization operation.
4. The device of claim 1 , wherein the first weight kernel value is further quantized by one of the unitary quantizing operation or a supervised iterative quantization operation.
5. The device of claim 1 , wherein the input feature map comprises an input feature map of an image.
6. The device of claim 1 , wherein the neural network comprises a convolutional neural network.
7. The device of claim 1 , wherein the device comprises a smartphone, a tablet computer, a portable electronic device, a computer or a server.
8. The device of claim 1 , wherein the neural network is configured to perform object recognition.
9. A device, comprising:
a processor; and
a memory,
the processor and the memory being configured as a neural network comprising:
at least one layer comprising an input and an output, the layer configured to receive an input feature map at the input and output an output feature map at the output, at least one of the input feature map and the output feature map comprising at least one first weight kernel value that has been quantized by a supervised iterative quantization operation to reduce a number of bits of the first weight kernel value from a first predetermined number of bits to a second predetermined number of bits that is less than the first predetermined number of bits without changing a dimension of the feature map corresponding to the quantized first weight kernel value.
10. The device of claim 9 , wherein the neural network further comprises at least one interim feature map between the input and the output of the layer, the interim feature map comprising at least one second weight kernel value that has been quantized by the supervised iterative quantization operation to reduce a number of bits of the second weight kernel value from a third predetermined number of bits to a fourth predetermined number of bits that is less than the third predetermined number of bits without changing a dimension of the interim feature map corresponding to the quantized second weight kernel value.
11. The device of claim 10 , wherein the second weight kernel value is further quantized by one of a unitary quantizing operation or the supervised iterative quantization operation.
12. The device of claim 9 , wherein the first weight kernel value is further quantized by one of a unitary quantizing operation or the supervised iterative quantization operation.
13. The device of claim 9 , wherein the input feature map comprises an input feature map of an image.
14. The device of claim 9 , wherein the neural network comprises a convolutional neural network.
15. The device of claim 9 , wherein the neural network is configured to perform object recognition.
16. The device of claim 9 , wherein the device comprises a smartphone, a tablet computer, a portable electronic device, a computer or a server.
17. A device, comprising:
a processor; and
a memory,
the processor and the memory being configured as a neural network comprising:
at least one first layer comprising an input and an output, the first layer configured to receive an input feature map at the input and output an output feature map at the output, the neural network further comprising at least one interim feature map between the input and the output of the first layer, the interim feature map comprising at least one first weight kernel value that has been quantized by a unitary quantizing operation or a supervised iterative quantization operation to reduce a number of bits of the first weight kernel value from a first predetermined number of bits to a second predetermined number of bits that is less than the first predetermined number of bits without changing a dimension of the interim feature map corresponding to the quantized first weight kernel value.
18. The device of claim 17 , wherein the neural network further comprises at least one second layer comprising an input and an output, the second layer configured to receive an input feature map at the input and output an output feature map at the output, at least one of the input feature map and the output feature map comprising at least one second weight kernel value that has been quantized by the unitary quantizing operation or the supervised iterative quantization operation to reduce a number of bits of the second weight kernel value from a third predetermined number of bits to a fourth predetermined number of bits that is less than the third predetermined number of bits without changing a dimension of the feature map corresponding to the quantized second weight kernel value.
19. The device of claim 17 , wherein the input feature map comprises an input feature map of an image.
20. The device of claim 17 , wherein the device comprises a smartphone, a tablet computer, a portable electronic device, a computer or a server.