Methods and systems for product quantization-based compression of a matrix

1. A computing system comprising:

a memory; and

a processing device in communication with the memory, the processing device configured to execute instructions to cause the computing system to compress a matrix including:

partition the matrix, having a plurality of rows formed by a respective plurality of vectors, into a plurality of submatrices, each submatrix containing sub-vectors from a respective group of one or more contiguous columns of the matrix;

for each given submatrix, cluster the sub-vectors of the given submatrix into a plurality of clusters;

for each given cluster, compute and store, in the memory, a centroid and a variance based on the sub-vectors belonging to the given cluster; and

store, in the memory, a mapping relating each vector to a respective cluster in each submatrix;

wherein the stored centroids, stored variances and stored mapping form a set of compressed data for reconstruction of the matrix, the set of compressed data requiring fewer bits to store in the memory than the matrix prior to compression.

2. The computing system of claim 1 , wherein the matrix is an embedding matrix, and the plurality of vectors is a plurality of embedding vectors, each embedding vector being a vector representation of a respective token in a vocabulary of a corpus.

3. The computing system of claim 1 , wherein the matrix is a layer of a neural network, and the plurality of vectors are trainable parameters of the neural network.

4. The computing system of claim 1 , wherein the processing device is configured to execute the instructions to cause the computing system to cluster the sub-vectors using K-means clustering.

5. The computing system of claim 1 , wherein the processing device is configured to execute the instructions to cause the computing system to cluster the sub-vectors using weighted K-means clustering, and wherein the centroid of each given cluster is computed using element weights to weight each sub-vector of the given cluster.

6. The computing system of claim 5 , wherein each vector of the matrix represents a respective token in a source dataset, and the element weight for each respective sub-vector of the given cluster is defined based on frequency of the corresponding token in the source dataset.

7. The computing system of claim 5 , wherein the element weight for each respective sub-vector of the given cluster is defined based on a Euclidian norm of a corresponding vector in the matrix.

8. The computing system of claim 1 , wherein the processing device is configured to execute the instructions to cause the computing system to:

communicate the set of compressed data to another device, the set of compressed data to be used as input for performing a machine learning task.

9. A computing system comprising:

a memory; and

a processing device in communication with the memory, the processing device configured to execute instructions to cause the computing system to:

obtain a set of compressed data for reconstruction of a matrix, the set of compressed data including a set of centroids and a set of variances, each centroid being associated with a respective cluster, and each variance being associated with a respective cluster, the set of compressed data further including a mapping relating each sub-vector of the matrix to a respective cluster;

for each given cluster, generate a reconstructed centroid by sampling from a distribution, the distribution being generated using the centroid and the variance associated with the given cluster;

reconstruct each given sub-vector using the reconstructed centroid generated for a relevant cluster, the relevant cluster for the given sub-vector being identified using the mapping;

reconstruct the matrix by concatenating the reconstructed sub-vectors; and

provide the reconstructed matrix as input to a neural network for performing a machine learning task.

10. The computing system of claim 9 , wherein the distribution is a multivariate Gaussian distribution.

11. The computing system of claim 9 , wherein the set of compressed data is obtained in a communication from another device.

12. The computing system of claim 9 , wherein the matrix is an embedding matrix.

13. A method for compressing a matrix, the method comprising:

partitioning the matrix, having a plurality of rows formed by a respective plurality of vectors, into a plurality of submatrices, each submatrix containing sub-vectors from a respective group of one or more contiguous columns of the matrix;

for each given submatrix, clustering the sub-vectors of the given submatrix into a plurality of clusters;

for each given cluster, computing and storing, in a memory of a computing system, a centroid and a variance based on the sub-vectors belonging to the given cluster; and

storing, in the memory, a mapping relating each vector to a respective cluster in each submatrix;

wherein the stored centroids, stored variances and stored mapping form a set of compressed data for reconstruction of the matrix, the set of compressed data requiring fewer bits to store in the memory than the matrix prior to compression.

14. The method of claim 13 , wherein the matrix is an embedding matrix, and the plurality of vectors is a plurality of embedding vectors, each embedding vector being a vector representation of a respective token in a vocabulary of a corpus.

15. The method of claim 13 , wherein the matrix is a layer of a neural network, and the plurality of vectors are trainable parameters of the neural network.

16. The method of claim 13 , wherein the sub-vectors are clustered using K-means clustering.

17. The method of claim 13 , wherein the sub-vectors are clustered using weighted K-means clustering, and wherein the centroid of each given cluster is computed using element weights to weight each sub-vector of the given cluster.

18. The method of claim 17 , wherein each vector of the matrix represents a respective token in a source dataset, and the element weight for each respective sub-vector of the given cluster is defined based on frequency of the corresponding token in the source dataset.

19. The method of claim 17 , wherein the element weight for each respective sub-vector of the given cluster is defined based on a Euclidian norm of a corresponding vector in the matrix.

20. The method of claim 13 , further comprising:

communicating the set of compressed data to another device, the set of compressed data to be used as input for performing a machine learning task.