Lossless compression of a content item using a neural network trained on content item cohorts
View Patent ↗Lossless compression of a content item using a neural network trained on content item cohorts. A computing system includes a neural network that is used to train a plurality of symbol prediction models. Each symbol prediction model is trained based on a corresponding cohort of content items. A particular symbol prediction model of the models trained is selected based on an intrinsic characteristic of a particular content item to be losslessly compressed such as, for example, the type or file extension of the content item. The content item is then losslessly compressed based on a set of symbol predictions fed to an arithmetic coder that are generated using the particular symbol prediction model selected.
1. A computer-implemented method, comprising:
using a neural network to train a plurality of symbol prediction models;
wherein each symbol prediction model of the plurality of symbol prediction models is trained based on a corresponding cohort of content items;
selecting a particular symbol prediction model of the plurality of symbol prediction models based on an intrinsic characteristic of a particular content item to be losslessly compressed;
losslessly compressing the particular content item based on a set of symbol predictions generated using the particular symbol prediction model; and
wherein the method is performed by one or more computing devices.
2. The method of claim 1 , wherein the intrinsic characteristic is a type of the particular content item.
3. The method of claim 1 , wherein the losslessly compressing the particular content item based on symbol predictions generated using the particular symbol prediction model is based on:
compressing the particular content item to a baseline lossless compressed data format;
binarizing the baseline lossless compressed data format to a binarized format; and
arithmetically coding the binarized format based on the set of symbol predictions generated using the particular symbol prediction model.
4. The method of claim 1 , wherein the losslessly compressing the particular content item based on symbol predictions generated using the particular symbol prediction model is based on arithmetically coding symbols based on the set of symbol predictions generated using the particular symbol prediction model.
5. The method of claim 1 , wherein the neural network is a recurrent neural network.
6. One or more non-transitory computer-readable media storing one or more programs for execution by one or more computing devices, the one or more programs having comprising instructions configured for:
using a neural network to train a plurality of symbol prediction models based on a first plurality of content item cohorts, the first plurality of content item cohorts including a plurality of content items;
compressing each content item in the plurality of content items using each symbol prediction model of the plurality of symbol prediction models; and
using a neural network to train a cohort predictor model based on results of the compressing each content item in the plurality of content items.
7. The one or more non-transitory computer-readable media of claim 6 , the instructions further configured for:
after the using a neural network to train the cohort predictor model, using the cohort predictor model to predict a cohort of the first plurality of content item cohorts for each content item in the plurality of content items.
8. The one or more non-transitory computer-readable media of claim 7 , the instructions further configured for:
forming a second plurality of content item cohorts based on results of the using the cohort predictor model, the second plurality of content item cohorts including the plurality of content items; and
using a neural network to update the plurality of symbol prediction models based on the second plurality of content item cohorts resulting in a plurality of updated symbol prediction models.
9. The one or more non-transitory computer-readable media of claim 8 , the instructions further configured for selecting a particular updated symbol prediction model of the plurality of updated symbol prediction models based on an intrinsic characteristic of a particular content item; and losslessly compressing the particular content item using the particular updated symbol prediction model.
10. The one or more non-transitory computer-readable media of claim 8 , the instructions further configured for:
using the cohort predictor model to predict a particular cohort of the second plurality of content item cohorts for a particular content item;
based on results of the using the cohort predictor model to predict the particular cohort for the particular content item, selecting a particular updated symbol prediction model of the plurality of updated symbol prediction models; and
losslessly compressing the particular content item using the particular updated symbol prediction model.
11. The one or more non-transitory computer-readable media of claim 6 , wherein the neural network used to train the plurality of symbol prediction models is a recurrent neural network.
12. The one or more non-transitory computer-readable media of claim 6 , wherein the neural network used to train the cohort predictor model is a recurrent neural network.
13. A computing system comprising:
one or more processors;
storage media;
one or more programs stored in the storage media for executions by the one or more processors, the one or more programs having instructions configured for:
using a neural network to train a plurality of symbol prediction models;
wherein each symbol prediction model of the plurality of symbol prediction models is trained based on a corresponding cohort of content items;
selecting a particular symbol prediction model of the plurality of symbol prediction models based on an intrinsic characteristic of a particular content item to be losslessly compressed;
losslessly compressing the particular content item based on a set of symbol predictions generated using the particular symbol prediction model; and
using a neural network to train a cohort predictor model based at least in part on results of the losslessly compressing the particular content item.
14. The computing system of claim 13 , the instructions further configured for:
after the using a neural network to train the cohort predictor model, using the cohort predictor model to predict a cohort of a first plurality of content item cohorts for each content item in a plurality of content items.
15. The computing system of claim 14 , the instructions further configured for:
forming a second plurality of content item cohorts based on results of the using the cohort predictor model, the second plurality of content item cohorts including the plurality of content items; and
using a neural network to update the plurality of symbol prediction models based on the second plurality of content item cohorts resulting in a plurality of updated symbol prediction models.
16. The computing system of claim 15 , the instructions further configured for selecting a particular updated symbol prediction model of the plurality of updated symbol prediction models based on an intrinsic characteristic of a particular content item; and losslessly compressing the particular content item using the particular updated symbol prediction model.
17. The computing system of claim 15 , the instructions further configured for:
using the cohort predictor model to predict a particular cohort of the second plurality of content item cohorts for a particular content item;
based on results of the using the cohort predictor model to predict the particular cohort for the particular content item, selecting a particular updated symbol prediction model of the plurality of updated symbol prediction models; and
losslessly compressing the particular content item using the particular updated symbol prediction model.
18. The computing system of claim 13 , wherein the intrinsic characteristic is a type of the particular content item.
19. The computing system of claim 13 , wherein the losslessly compressing the particular content item based on symbol predictions generated using the particular symbol prediction model is based on:
compressing the particular content item to a baseline lossless compressed data format;
binarizing the baseline lossless compressed data format to a binarized format; and
arithmetically coding the binarized format based on the set of symbol predictions generated using the particular symbol prediction model.
20. The computing system of claim 13 , wherein the losslessly compressing the particular content item based on symbol predictions generated using the particular symbol prediction model is based on arithmetically coding symbols based on the set of symbol predictions generated using the particular symbol prediction model.