System and method for emulating quantization noise for a neural network

A method for training a classification system's quantized neural network dataset, comprising:

• receiving by at least one hardware processor digital input data comprising a plurality of training input value sets and a plurality of target value sets;
• in each training iteration of a plurality of training iterations:
• for each layer, comprising a plurality of weight values, of one or more layers of a plurality of layers of a neural network:
• computing a set of transformed values by applying to a plurality of layer values, comprising a plurality of previous output values of a previous layer and the layer's plurality of weight values, one or more emulated non-uniformly quantized transformations by adding to each value of the plurality of layer values one or more uniformly distributed random noise values;
• and
• computing a plurality of output values by applying to the set of transformed values one or more arithmetic operations;
• computing a plurality of training output values from a combination of the plurality of output values of a last layer of the plurality of layers;
• and
• updating one or more of the plurality of weight values of the one or more layers to decrease a value of a loss function computed using the plurality of target value sets and plurality of training output values;
• and
• outputting the updated plurality of weight values of the plurality of layers.

A system for detection and classification of findings in digital data, comprising at least one hardware processor configured to:

• accessing a non-uniformly quantized neural network data;
• and
• classifying at least one finding detected in the digital data according to the non-uniformly quantized neural network data set in response to receiving the digital data;
• wherein the non-uniformly quantized neural network data set is generated by:
• receiving digital input data comprising a plurality of training input value sets and a plurality of target value sets;
• in each training iteration of a plurality of training iterations:
• for each layer, comprising a plurality of weight values, of one or more layers of a plurality of layers of a neural network:
• computing a set of transformed values by applying to a plurality of layer values, comprising a plurality of previous output values of a previous layer and the layer's plurality of weight values, one or more emulated non-uniformly quantized transformations by adding to each value of the plurality of layer values one or more uniformly distributed random noise values;
• and
• computing a plurality of output values by applying to the set of transformed values one or more arithmetic operations;
• computing a plurality of training output values from a combination of the plurality of output values of a last layer of the plurality of layers;
• and
• updating one or more of the plurality of weight values of the one or more layers to decrease a value of a loss function computed using the plurality of target value sets and plurality of training output values;
• and
• outputting the updated plurality of weight values of the plurality of layers.

A system for regression of digital data, comprising at least one hardware processor configured to:

• accessing a non-uniformly quantized neural network data;
• and
• predicting at least one output value according to the non-uniformly quantized neural network data set in response to receiving the digital data;
• wherein the non-uniformly quantized neural network data set is generated by:
• receiving digital input data comprising a plurality of training input value sets and a plurality of target value sets;
• in each training iteration of a plurality of training iterations:
• for each layer, comprising a plurality of weight values, of one or more layers of a plurality of layers of a neural network:
• computing a set of transformed values by applying to a plurality of layer values, comprising a plurality of previous output values of a previous layer and the layer's plurality of weight values, one or more emulated non-uniformly quantized transformations by adding to each value of the plurality of layer values one or more uniformly distributed random noise values;
• and
• computing a plurality of output values by applying to the set of transformed values one or more arithmetic operations;
• computing a plurality of training output values from a combination of the plurality of output values of a last layer of the plurality of layers;
• and
• updating one or more of the plurality of weight values of the one or more layers to decrease a value of a loss function computed using the plurality of target value sets and plurality of training output values;
• and
• outputting the updated plurality of weight values of the plurality of layers.