Tensor processing using low precision format

A computer-implemented method of performing training for a deep learning network, said method comprising:

• accessing a first matrix, X;
• accessing a second matrix, Y, wherein said matrices X and Y are memory resident and comprise values represented in a float16 floating point format and wherein said float16 format represents values from a low threshold to a high threshold;
• determining a first scaling factor, a, for matrix X;
• determining a second scaling factor, b, for matrix Y, wherein said determinings are performed to avoid overflow and underflow for matrix computations involving a*X and b*Y;
• and
• performing said training by performing computations of a forward propagation layer, a convolutional layer, an inner product layer and a backward propagation layer, wherein said convolutional layer comprises a gradient computation that comprises a matrix computation that takes a*X and b*Y as inputs and that generates a gradient output.

A non-transitory computer readable medium comprising a plurality of programmed instructions stored thereon, wherein, the plurality of programmed instructions are configured to be executed by a processing device to perform training for a deep learning network, said plurality of programmed instructions comprising:

• instructions to access a first matrix, X;
• instructions to access a second matrix, Y, wherein said matrices X and Y are memory resident and comprise values represented in a float16 floating point format and wherein said float16 format represents values from a low threshold to a high threshold;
• instructions to determine a first scaling factor, a, for matrix X;
• instructions to determine a second scaling factor, b, for matrix Y, wherein said instructions to determine said first and said second scaling factors are performed to avoid overflow and underflow for matrix computations involving a*X and b*Y;
• and
• instructions to perform said training by performing computations of a forward propagation layer, a convolutional layer, an inner product layer and a backward propagation layer, wherein said convolutional layer comprises a gradient computation that comprises a matrix computation that takes a*X and b*Y as inputs and that generates a gradient output.