Microsoft Technology Licensing, LLC

Tiyasa Mitra

Marc Tremblay

Andy Wagner

Saurabh Mohan Kulkarni

Sujeeth Subramanya Bharadwaj

Andrew Wagner

Bharadwaj Pudipeddi

Maral Mesmakhosroshahi

Layali Rashid

Mohit Mittal

Gautham Popuri

Embodiments of the present disclosure include systems and methods for packing tokens to train sequence models. In some embodiments, a plurality of datasets for training a sequence model is received. Each dataset in the plurality of datasets includes a sequence of correlated tokens. A set of training data is generated that includes a subset of a sequence of tokens from a first dataset in the plurality of datasets and a subset of a sequence of tokens from a second, different dataset in the plurality of datasets. The sequence model is trained using the set of training data.

Token packing for sequence models

Embodiments of the present disclosure include systems and methods for forcing weights of transformer model layers when training a transformer model. In some embodiments, input data is received at a first layer included in a transformer model. The input data is processed through the first layer of the transformer model to produce a first output data. The first output data is processed through the first layer of the transformer model to produce a second output data. The first output data is processed through a second layer included in the transformer model to produce a third output data. A difference is calculated between the second output data and the third output data. Weights included in the first layer of the transformer model are adjusted based on the calculated difference.

Forcing weights of transformer model layers

Embodiments of the present disclosure include systems and methods for compressing and decompressing data generated by sub-blocks in a neural network. In some embodiment, an input matrix is received at a compression block in the neural network. The compression block compresses the input matrix into a compressed matrix and outputs the compressed matrix. The compressed matrix has a reduced dimensionality relative to a dimensionality of the input matrix. A decompression block retrieves the compressed matrix. The decompression block decompresses compressed matrix into a decompressed matrix and outputs the decompressed matrix. The decompressed matrix has a same dimensionality as the dimensionality of the input matrix. The compression and decompression blocks are optimized based on feedback received from the neural network.

Compressing and decompressing data for language models

Embodiments of the present disclosure include systems and methods for determining position values for training data that is used to train transformer models. In some embodiments, a set of input data for training a transformer model is received. The set of input data comprises a set of tokens. Based on an offset value, a set of successive position values for the set of tokens is determined. Each position value in the set of successive position values represents a position of a token in the set of tokens relative to other tokens in the set of tokens. A set of training data is generated to comprise the set of tokens and the set of successive position values. The transformer model is trained using the set of training data.

Determining position values for transformer models

Embodiments of the present disclosure include systems and methods for training transformer models using position masking. In some embodiments, a set of data for training a transformer model is received. The set of data includes a sequence of tokens and a set of position values. Each position value in the set of position values represents a position of a token in the sequence of tokens relative to other tokens in the sequence of tokens. A subset of the set of position values in the set of data is selected. Each position value in the subset of the set of position values is replaced with a second defined value to form a second set of defined values. The transformer model is trained using the set of data.

Position masking for transformer models

Embodiments of the present disclosure include systems and methods for reducing hardware resource utilization by residual neural networks. In some embodiments, a first matrix is received at a layer included in a neural network. The first matrix is compressed to produce a second matrix. The second matrix has a reduced dimensionality relative to a dimensionality of the first matrix. The second matrix is processed through a network block in the layer included in the neural network. The processed second matrix is expanded to produce a third matrix. The third matrix has a dimensionality that is equal to a dimensionality of the first matrix. The third matrix is added to the first matrix to produce a fourth matrix.

Reducing hardware resource utilization for residual neural networks

Systems and methods for pipelined neural network processing with continuous and asynchronous updates are described. A method for processing a neural network comprising L layers, where L is an integer greater than two, includes partitioning the L layers among a set of computing resources configured to process forward passes and backward passes associated with each of the L layers. The method further includes initiating processing of the forward passes and the backward passes using the set of computing resources. The method further includes upon completion of a first set of forward passes and a first set of backward passes associated with a first layer of the L layers, initiating update of parameters associated with the first layer when gradients are available for updating the parameters associated with the first layer without waiting to calculate gradients associated with any of remaining L layers.

Pipelined neural network processing with continuous and asynchronous updates

Embodiments of the present disclosure include systems and methods for training neural networks. In one embodiment, neural network may receive input data and produce output results in response to the input data and weights of the neural network. An error is determined at an output of the neural network based on the output results. The error is propagated in a reverse direction through the neural network from the output and one or more intermediate outputs to adjust the weights.

Systems and methods for training a neural network

Embodiments of the present disclosure include a method for token-position handling comprising: processing a first sequence of tokens to produce a second sequence of tokens, wherein the second sequence of tokens has a smaller number of tokens than the first sequence of tokens; masking at least some tokens in the second sequence to produce masked tokens; moving the masked tokens to the beginning of the second sequence to produce a third sequence; encoding tokens in the third sequence into a set of numeric vectors in a first array; and processing the first array in a transformer neural network to determine correlations among the third sequence, the processing the first array producing a second array.

Token-position handling for sequence based neural networks

Embodiments of the present disclosure include systems and methods for compressing weights for distributed neural networks. In some embodiments, a first network comprising a first set of weights is trained using a set of training data. A second network comprising a second set of weights is trained using the set of training data. A number of weights in the first set of weights is greater than a number of weights in the second set of weights. The first set of weights are adjusted based on a first loss determined by the first network and a second loss determined by the second network. The second set of weights are adjusted based on the first loss determined by the first network and the second loss determined by the second network. Values of the second set of weights are sent to a computing system.

Compressing weights for distributed neural networks

Methods, systems, apparatuses, and computer program products are described herein that enable execution of a large AI model on a memory-constrained target device that is communicatively connected to a parameter server, which stores a master copy of the AI model. The AI model may be dissected into smaller portions (e.g., layers or sub-layers), and each portion may be executed as efficiently as possible on the target device. After execution of one portion of the AI model is finished, another portion of the AI model may be downloaded and executed at the target device. To improve efficiency, the input samples may be divided into microbatches, and a plurality of microbatches executing in sequential order may form a minibatch. The size of the group of microbatches or minibatch can be manually or automatically adjusted to reduce the communication overhead.

Executing large artificial intelligence models on memory-constrained devices

Techniques for training neural networks are provided. According to one set of embodiments, a first array is processed in a spreading component to produce a second array, where a first dimension of the first array corresponds to at least one sequence of approximately orthogonal numeric vectors representing tokens, and where the spreading component combines values along the first dimension. The second array is processed in a transformer neural network to determine correlations between the sequence, which produces a third array. One or more batches of the third array are processed in a de-spreading component to produce a fourth array.

Spread neural networks

Methods for gradient accumulation with free momentum are performed by systems and devices during neural network model training. An accumulator that includes a processor circuit and a memory element generates free momentum between passes of a neural network model training process. The processor circuit receives a difference weight (gradient) and generates a first input by applying a weighting parameter thereto. The processor circuit obtains a prior weight from the memory element and generates a second input by applying another weighting parameter thereto. The processor circuit generates a filtered input with momentum by filtering the first and second input. The memory element generates a stored next pass weight by accumulating the filtered input with the prior weight. A computing resource then processes the next pass of the neural network model training using the stored next pass weight. The methods, systems, and devices are applicable to pipelined model parallelism training processes.

Growing community of inventors

San Jose, CA, United States of America

Tiyasa Mitra