Microsoft Technology Licensing, LLC

Daniel Lo

Eric S Chung

Ritchie Zhao

Amar Phanishayee

Yiren Zhao

Blake Douglas Pelton

Douglas Christopher Burger

Bita Darvish Rouhani

Kalin Ovtcharov

Taesik Na

David Earl Heckerman

Ravindra Pandya

Jialiang Zhang

Haishan Zhu

Eric Sen Chung

Jeremy Halden Fowers

Roman Snytsar

Deeksha Dangwal

Jeremy Fowers

Bita Rouhani Darvish

Methods and apparatus are disclosed for providing emulation of quantized precision operations in a neural network. In some examples, the quantized precision operations are performed in a block floating-point format where values of a tensor share a common exponent. Techniques for selecting higher precision or lower precision can be used based on a variety of input metrics. When converting to a quantized tensor, a residual tensor is produced. In one embodiment, an error value associated with converting from a normal-precision floating point number to the quantized tensor is used to determine whether to use the residual tensor in a dot product calculation. Using the residual tensor increases the precision of an output from a node. Selection of whether to use the residual tensor can depend on various input metrics including the error value, the layer number, the exponent value, the layer type, etc.

Residual quantization for neural networks

Apparatus and methods for training a neural network accelerator using quantized precision data formats are disclosed, and in particular for storing activation values from a neural network in a compressed format for use during forward and backward propagation training of the neural network. In certain examples of the disclosed technology, a computing system includes processors, memory, and a compressor in communication with the memory. The computing system is configured to perform forward propagation for a layer of a neural network to produced first activation values in a first block floating-point format. In some examples, activation values generated by forward propagation are converted by the compressor to a second block floating-point format having a narrower numerical precision than the first block floating-point format. The compressed activation values are stored in the memory, where they can be retrieved for use during back propagation.

Neural network activation compression with narrow block floating-point

Methods and systems are provided for decoding variable-length codes in a parallel process. A stream of variable-length code words is divided into fixed length words. A plurality of parallel sets of decoder circuits each receive, in parallel, a current fixed length word and a prior fixed length word. Each decoder circuit has a respective fixed leftover bit-count. Each decoder circuit generates a respective output that may include a decoded symbol and a new leftover bit-count. Each respective output is determined based on the respective current fixed length word, the respective prior fixed length word, and the respective fixed leftover bit-count. A set of selected decoder circuit outputs is generated for each set of the parallel sets of decoder circuits based on a set of first leftover bit-counts. One output from each set of selected decoder circuit outputs is selected as a final output based on a second prior leftover bit-count.

Parallelized decoding of variable-length prefix codes

Apparatus and methods for training a neural network accelerator using quantized precision data formats are disclosed, and in particular for storing activation values from a neural network in a compressed format having lossy or non-uniform mantissas for use during forward and backward propagation training of the neural network. In certain examples of the disclosed technology, a computing system includes processors, memory, and a compressor in communication with the memory. The computing system is configured to perform forward propagation for a layer of a neural network to produced first activation values in a first block floating-point format. In some examples, activation values generated by forward propagation are converted by the compressor to a second block floating-point format having a non-uniform and/or lossy mantissa. The compressed activation values are stored in the memory, where they can be retrieved for use during back propagation.

Neural network activation compression with non-uniform mantissas

Apparatus and methods for training a neural network accelerator using quantized precision data formats are disclosed, and, in particular, for adjusting floating-point formats used to store activation values during training. In certain examples of the disclosed technology, a computing system includes processors, memory, and a floating-point compressor in communication with the memory. The computing system is configured to produce a neural network comprising activation values expressed in a first floating-point format, select a second floating-point format for the neural network based on a performance metric, convert at least one of the activation values to the second floating-point format, and store the compressed activation values in the memory. Aspects of the second floating-point format that can be adjusted include the number of bits used to express mantissas, exponent format, use of non-uniform mantissas, and/or use of outlier values to express some of the mantissas.

Adjusting activation compression for neural network training

Apparatus and methods for training a neural network accelerator using quantized precision data formats having outlier values are disclosed, and in particular for storing activation values from a neural network in a compressed format for use during forward and backward propagation training of the neural network. In certain examples of the disclosed technology, a computing system is configured to perform forward propagation for a layer of a neural network to produced first activation values in a first block floating-point format. In some examples, activation values generated by forward propagation are converted by the compressor to a second block floating-point format having a narrower numerical precision than the first block floating-point format. Outlier values, comprising additional bits of mantissa and/or exponent are stored in ancillary storage for subset of the activation values. The compressed activation values are stored in the memory, where they can be retrieved for use during back propagation.

Neural network activation compression with outlier block floating-point

The performance of a text parser implemented with a state machine is improved by reducing a critical dependence path. In one aspect, all possible current states for a given text input are read from a state table circuit, and the correct next state and output are then selected therefrom by an output multiplexer based on the current state, removing dependence on the current state from the table read, and allowing the read(s) to be pipelined. Further, multiple input units are configured to operate on multiple text characters in parallel, with each input unit propagating outputs for its state table circuit to the next downstream input unit. Each downstream input unit is configured to use the propagated states to provide the proper outputs to appropriates multiplexer inputs. The number of possible output states may be dynamically reduced, thereby reducing the size of the output multiplexer needed to select the next state.

High-performance microcoded text parser

A table-based state machine is improved by reducing critical dependence path. In one aspect, all current states for a given input are read from a state table circuit, and the next state and output are then selected therefrom by an output multiplexer based on the current state, removing dependence on the current state from the table read, and allowing the read(s) to be pipelined. In a further aspect, multiple input units are configured to operate on multiple inputs in parallel, with each input unit propagating its state table circuit for its current input to the next downstream input unit. Each downstream input unit is configured to use the propagated state table circuit to provide the state table circuit reads to the proper output multiplexer input. The number of possible output states for a given input may be dynamically reduced, reducing the size of the output multiplexer selecting the next state.

High-performance table-based state machine

Bounding box quantization can reduce the quantity of bits utilized to express numerical values prior to the multiplication of matrices comprised of such numerical values, thereby reducing both memory consumption and processor utilization. Stochastic rounding can provide sufficient precision to enable the storage of weight values in reduced-precision formats without having to separately store weight values in a full-precision format. Alternatively, other rounding mechanisms, such as round to nearest, can be utilized to exchange weight values in reduced-precision formats, while also storing weight values in full-precision formats for subsequent updating. To facilitate conversion, reduced-precision formats such as brain floating-point format can be utilized.

Neural network training with decreased memory consumption and processor utilization

A system for training a neural network receives training data and performing lower precision format training calculations using lower precision format data at one or more training phases. One or more results from the lower precision format training calculations are converted to higher precision format data, and higher precision format training calculations are performed using the higher precision format data at one or more additional training phases. The neural network is modified using the results from the one or more additional training phases. The mixed precision format training calculations train the neural network more efficiently, while maintaining an overall accuracy.

Training neural networks using mixed precision computations

Technology related to training a neural network accelerator using mixed precision data formats is disclosed. In one example of the disclosed technology, a neural network accelerator is configured to accelerate a given layer of a multi-layer neural network. An input tensor for the given layer can be converted from a normal-precision floating-point format to a quantized-precision floating-point format. A tensor operation can be performed using the converted input tensor. A result of the tensor operation can be converted from the block floating-point format to the normal-precision floating-point format. The converted result can be used to generate an output tensor of the layer of the neural network, where the output tensor is in normal-precision floating-point format.

Training neural network accelerators using mixed precision data formats

Methods and apparatus are disclosed supporting a design flow for developing quantized neural networks. In one example of the disclosed technology, a method includes quantizing a normal-precision floating-point neural network model into a quantized format. For example, the quantized format can be a block floating-point format, where two or more elements of tensors in the neural network share a common exponent. A set of test input is applied to a normal-precision flooding point model and the corresponding quantized model and the respective output tensors are compared. Based on this comparison, hyperparameters or other attributes of the neural networks can be adjusted. Further, quantization parameters determining the widths of data and selection of shared exponents for the block floating-point format can be selected. An adjusted, quantized neural network is retrained and programmed into a hardware accelerator.

Flow for quantized neural networks

Machine learning may include training and drawing inference from artificial neural networks, processes which may include performing convolution and matrix multiplication operations. Convolution and matrix multiplication operations are performed using vectors of block floating-point (BFP) values that may include outliers. BFP format stores floating-point values using a plurality of mantissas of a fixed bit width and a shared exponent. Elements are outliers when they are too large to be represented precisely with the fixed bit width mantissa and shared exponent. Outlier values are split into two mantissas. One mantissa is stored in the vector with non-outliers, while the other mantissa is stored outside the vector. Operations, such as a dot product, may be performed on the vectors in part by combining the in-vector mantissa and exponent of an outlier value with the out-of-vector mantissa and exponent.

Growing community of inventors

Bothell, WA, United States of America

Daniel Lo