Ambarella International LP

Malhar Palkar

Santosh Chilkunda

Sabhatina Palani Selvam

A method of generating a quantized neural network comprises (i) receiving a neural network model, (ii) modifying the neural network model by quantizing input of at least convolution layers of the neural network model, and (iii) training the modified neural network model using a dataset that is representative of one or more desired inferences.

Quantization aware training by constraining input

An apparatus includes a preprocessing network and a quantized neural network. The preprocessing network may be configured to (a) receive an input, (b) perform a transform operation on the input, and (c) generate an intermediate result in response to the transform operation. The quantized neural network may be configured to (a) receive the intermediate result, (b) perform a plurality of weighted operations on the intermediate result, and (c) generate an output in response to the weighted operations. The quantized neural network is generally adapted to a hardware constraint of an edge device. The weighted operations generally have a quantization loss corresponding to the hardware constraint. The transform operation performed by the preprocessing network is generally configured to provide compensation for the quantization loss.

Modeling quantization loss using a preprocessing network that simulates hardware behavior

A method of generating a quantized neural network comprises (i) receiving a pre-trained neural network model and (ii) modifying the pre-trained neural network model to calculate one or more statistics on an output of one or more layers of the pre-trained neural network model based on a current image and set up an output data format for one or more following layers of the pre-trained neural network model for one or more of the current image and a subsequent image dynamically based on the one or more statistics.

Growing community of inventors

Cupertino, CA, United States of America

Malhar Palkar