Compression of deep learning models

A computer-implemented method comprising:

• determining, in respect of an artificial intelligence, AI, system comprising a deep learning model with a plurality of layers defining a weight matrix of the model, groups of two or more contiguous layers in the plurality of layers, each group of two or more contiguous layers sharing a common dimension;
• for each determined group of two or more contiguous layers, applying singular value decomposition, SVD, to weights of one layer of the group;
• for each layer with SVD applied thereto, truncating at least some singular values obtained from the application of the SVD based on at least one truncation criterion;
• for each determined group of two or more contiguous layers, selecting a plurality of output channels of the deep learning model for removal based on at least one weight importance criterion, wherein a ratio of a number of output channels in the selected plurality of output channels divided by a total number of output channels in the respective group is equal to or less than a ratio of a number of truncated singular values in the respective group divided by a number of truncated singular values obtained for the respective group;
• and
• for each determined group of two or more contiguous layers, removing at least one of one or more input channels or one or more output channels based on the selected plurality of output channels such that a shared dimension size in the group is maintained.

A computer-readable non-transitory storage medium comprising instructions which, when executed by at least one processing apparatus or system, cause the at least one processing apparatus or system to carry out at least the following steps:

• determining, in respect of an artificial intelligence, AI, system comprising a deep learning model with a plurality of layers defining a weight matrix of the model, groups of two or more contiguous layers in the plurality of layers, each group of two or more contiguous layers sharing a common dimension;
• for each determined group of two or more contiguous layers, applying singular value decomposition, SVD, to weights of one layer of the group;
• for each layer with SVD applied thereto, truncating at least some singular values obtained from the application of the SVD based on at least one truncation criterion;
• for each determined group of two or more contiguous layers, selecting a plurality of output channels of the deep learning model for removal based on at least one weight importance criterion, wherein a ratio of a number of output channels in the selected plurality of output channels divided by a total number of output channels in the respective group is equal to or less than a ratio of a number of truncated singular values in the respective group divided by a number of truncated singular values obtained for the respective group;
• and
• for each determined group of two or more contiguous layers, removing at least one of one or more input channels or one or more output channels based on the selected plurality of output channels such that a shared dimension size in the group is maintained.