System and method for using a deep learning network over time
The present approach relates to a system capable of life-long learning in a deep learning context. The system includes a deep learning network configured to process an input dataset and perform one or more tasks from among a first set of tasks. As an example, the deep learning network may be part of an imaging system, such as a medical imaging system, or may be used in industrial applications. The system further includes a learning unit communicatively coupled to the deep learning network 102 and configured to modify the deep learning network so as to enable it to perform one or more tasks in a second task list without losing the ability to perform the tasks from the first list.
1. A method for updating a deep learning network over time, comprising the steps of:
receiving a first set of parameters from a deep learning network, wherein the deep learning network is trained using a first training dataset to perform a first set of tasks, wherein the first set of parameters specify both a first feature extractor and a first classifier used to perform the first set of tasks;
receiving a first feature set corresponding to the first training dataset;
receiving an input comprising a second set of tasks and a second training dataset;
generating a second set of parameters specifying both a second feature extractor and a second classifier for use by the deep learning network, wherein the second set of parameters are generated using the first set of parameters, the input, and the first feature set, and the first training dataset is not used in generating the second set of parameters; and
modifying the deep learning network to use the second set of parameters so that the deep learning network is trained to perform tasks from the first set of tasks and the second set of tasks without degradation.
2. The method of claim 1 , wherein the deep learning network comprises a memory augmented neural network.
3. The method of claim 2 , wherein the first feature set is stored in a memory of the memory augmented neural network.
4. The method of claim 1 , wherein the deep learning network is trained to process one or more of medical images or industrial images.
5. The method of claim 1 , wherein the first set of parameters are generated by training the deep learning network to perform the first set of tasks using the first training dataset.
6. The method of claim 1 , wherein the second set of parameters enables the deep learning network to perform the tasks from the first task list at a same level as the deep learning network performed the tasks from the first task list prior to generating the second set of parameters.
7. The method of claim 1 , wherein the steps of the method are performed on a learning unit associated with the deep learning network and implemented using one or more processor units and at least one memory unit of the learning unit.
8. The method of claim 7 , wherein the learning unit comprises a data set generator configured to receive at least the first set of parameters, the first feature set, and the second training dataset and to generate an intermediate feature set based on the first feature extractor and the second training dataset.
9. The method of claim 7 , wherein the learning unit comprises a feature transformer unit configured to train a feature transformer based at least on the second training dataset.
10. The method of claim 9 , wherein the feature transformer is trained by minimizing a model loss cost function.
11. The method of claim 7 , wherein the learning unit comprises a deep learning network parameter generator configured to generate the second set of parameters based at least on a feature transformer.
12. A system, comprising:
a deep learning network initially trained using a first training dataset to perform a first set of tasks;
a learning unit in communication with the deep learning network, the learning unit comprising:
one or more memory components storing data and computer logic;
one or more processors configured to execute computer logic stored on the one or more memory components so as to cause acts to be performed comprising:
receiving a first set of parameters from the deep learning network, wherein the first set of parameters specify both a first feature extractor and a first classifier used to perform the first set of tasks;
receiving a first feature set corresponding to the first training dataset;
receiving an input comprising a second set of tasks and a second training dataset;
generating a second set of parameters specifying both a second feature extractor and a second classifier for use by the deep learning network, wherein the second set of parameters are generated using the first set of parameters, the input, and the first feature set, and the first training dataset is not used in generating the second set of parameters; and
modifying the deep learning network to use the second set of parameters so that the deep learning network is trained to perform tasks from the first set of tasks and the second set of tasks without degradation.
13. The system of claim 12 , wherein the one or more memory components and one or more processors facilitate the operation of or implement a dataset generator configured to receive at least the first set of parameters, the first feature set, and the second training dataset and to generate an intermediate feature set based on the first feature extractor and the second training dataset.
14. The system of claim 12 , wherein the one or more memory components and one or more processors facilitate the operation of or implement feature transformer logic configured to train a feature transformer based at least on the second training dataset.
15. The system of claim 14 , wherein the feature transformer is trained by minimizing a model loss cost function.
16. The system of claim 12 , wherein the one or more memory components and one or more processors facilitate the operation of or implement a deep learning network parameter generator configured to generate the second set of parameters based at least on a feature transformer.
17. The system of claim 12 , wherein the deep learning network comprises a memory augmented neural network.
18. The system of claim 12 , wherein the deep learning network is trained to process one or more of medical images or industrial images.
19. The system of claim 12 , wherein the first set of parameters are generated by training the deep learning network to perform the first set of tasks using the first training dataset.
20. The system of claim 12 , wherein the second set of parameters enables the deep learning network to perform the tasks from the first task list at a same level as the deep learning network performed the tasks from the first task list prior to generating the second set of parameters.