Look-up table based neural networks

One or more non-transitory computer-readable storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations for processing a network input using a neural network to generate a network output for the network input, wherein the neural network comprises a plurality of neural network layers, wherein each neural network layer has a plurality of weights, and wherein each neural network layer generates a layer output at least in part by performing a plurality of multiplications, each multiplication being a multiplication between a weight of the neural network layer and an input value from a layer input to the neural network layer, and generating a plurality of accumulated values that are each based at least in part on a sum of results of a subset of plurality of multiplications, and the operations comprising:

• receiving the network input;
• and
• generating the network output from the network input using particular computer hardware, comprising, for each of one or more of the neural network layers:
• loading, into memory of the particular computer hardware, a look-up table for the neural network layer that maps each possible combination of a quantized input index and a quantized weight index to a multiplication result, wherein each quantized input index represents a quantized input value from a set of possible quantized input values, each quantized weight index represents a quantized weight value from a set of possible quantized weight values, and the multiplication result is a result of a multiplication between the quantized input value and the quantized weight value;
• receiving data specifying a quantized input to the neural network layer, the quantized input comprising a plurality of quantized input values;
• and
• generating, by the particular computer hardware, a layer output for the neural network layer from the quantized input to the neural network layer, comprising:
• for each multiplication between a quantized weight and a corresponding quantized input value that is required to generate the layer output for the neural network layer:
• determining the quantized input index representing the quantized input value;
• determining the quantized weight index representing the quantized weight value;
• and
• identifying, as the result of the multiplication and by accessing the look-up table for the neural network layer in the memory of the particular computer hardware, the multiplication result mapped to in the look-up table for the neural network layer by the determined quantized input index and the determined quantized weight index;
• and
• generating a plurality of accumulated values, comprising for each accumulated value, summing the identified results of the corresponding subset of the plurality of multiplications.

A system comprising one or more computers and one or more storage devices storing instructions that when executed by one or more computers cause the one or more computers to perform operations for processing a network input using a neural network to generate a network output for the network input, wherein the neural network comprises a plurality of neural network layers, wherein each neural network layer has a plurality of weights, and wherein each neural network layer generates a layer output at least in part by performing a plurality of multiplications, each multiplication being a multiplication between a weight of the neural network layer and an input value from a layer input to the neural network layer, and generating a plurality of accumulated values that are each based at least in part on a sum of results of a subset of plurality of multiplications, and the operations comprising:

• receiving the network input;
• and
• generating the network output from the network input using particular computer hardware, comprising, for each of one or more of the neural network layers:
• loading, into memory of the particular computer hardware, a look-up table for the neural network layer that maps each possible combination of a quantized input index and a quantized weight index to a multiplication result, wherein each quantized input index represents a quantized input value from a set of possible quantized input values, each quantized weight index represents a quantized weight value from a set of possible quantized weight values, and the multiplication result is a result of a multiplication between the quantized input value and the quantized weight value;
• receiving data specifying a quantized input to the neural network layer, the quantized input comprising a plurality of quantized input values;
• and
• generating, by the particular computer hardware, a layer output for the neural network layer from the quantized input to the neural network layer, comprising:
• for each multiplication between a quantized weight and a corresponding quantized input value that is required to generate the layer output for the neural network layer:
• determining the quantized input index representing the quantized input value;
• determining the quantized weight index representing the quantized weight value;
• and
• identifying, as the result of the multiplication and by accessing the look-up table for the neural network layer in the memory of the particular computer hardware, the multiplication result mapped to in the look-up table for the neural network layer by the determined quantized input index and the determined quantized weight index;
• and
• generating a plurality of accumulated values, comprising for each accumulated value, summing the identified results of the corresponding subset of the plurality of multiplications.