Systems and methods for accelerating sparse neural network execution

A method for dynamic sparse execution of a neural network, comprising:

• providing, via a buffer, inputs for a neural network to at least one processor;
• executing, via the at least one processor, ternary random projection to reduce at least one dimension of the inputs according to a tunable degree for indicating a degree at which missing values are to be predicted rather than calculated;
• generating, via the at least one processor, a corresponding predictable output neuron map;
• executing, via a plurality of processing elements, one or more first activation functions of the neural network using the reduced inputs to generate first outputs, wherein at least one of the first outputs is expanded using the corresponding predictable output neuron map;
• providing, via the buffer, the first outputs to the at least one processor;
• reducing, via the at least one processor, at least one dimension of the first outputs;
• updating, via the at least one processor, the corresponding predictable output neuron map based on the reduced first outputs;
• and
• executing, via the plurality of processing elements, one or more second activation functions of the neural network using the reduced first outputs to generate second outputs, wherein at least one of the second outputs is expanded by setting values corresponding to one or more second predictable output neurons.

A non-transitory computer-readable storage medium storing a set of instructions that is executable by a computing device to cause the computing device to perform a method for dynamic sparse execution of a neural network, the method comprising:

• providing, via a buffer, inputs for a neural network to at least one processor;
• executing, via the at least one processor, ternary random projection to reduce at least one dimension of the inputs according to a tunable degree for indicating a degree at which missing values are to be predicted rather than calculated;
• generating, via the at least one processor, a corresponding predictable output neuron map;
• executing, via a plurality of processing elements, one or more first activation functions of the neural network using the reduced inputs to generate first outputs, wherein at least one of the first outputs is expanded by setting values corresponding to one or more first predictable output neurons;
• providing, via the buffer, the first outputs to the at least one processor;
• reducing, via the at least one processor, at least one dimension of the first outputs;
• updating, via the at least one processor, the corresponding predictable output neuron map based on the reduced first outputs;
• and
• executing, via the plurality of processing elements, one or more second activation functions of the neural network using the reduced first outputs to generate second outputs, wherein at least one of the second outputs is expanded by setting values corresponding to one or more second predictable output neurons.