Systems and methods for utilizing a machine learning model combining episodic and semantic information to process a new class of data without loss of semantic knowledge

1. A method comprising:

receiving, by a device, an image for training a neural network model;

utilizing, by the device and based on using the neural network model, an episodic memory to determine a first classification associated with the image;

utilizing, by the device and based on using the neural network model, a semantic memory to determine a second classification associated with the image;

determining, by the device, an accuracy associated with utilizing the semantic memory to determine the second classification;

determining, by the device, that the image is associated with the first classification when the accuracy fails to satisfy a threshold accuracy; and

determining, by the device, that the image is associated with the second classification when the accuracy satisfies the threshold accuracy.

2. The method of claim 1 , wherein utilizing the episodic memory to determine the first classification comprises:

processing the image with an episodic deep neural network model to determine the first classification associated with the image.

3. The method of claim 1 , wherein the episodic memory includes an episodic deep neural network model that includes a memory layer storing one or more samples associated with the first classification, and

wherein utilizing the episodic memory to determine the first classification comprises:

processing the image with the episodic deep neural network model to determine a similarity between the image and the one or more samples; and

determining that the image is associated with the first classification based on the similarity between the image and the one or more samples.

4. The method of claim 1 , wherein utilizing the semantic memory to determine the second classification comprises:

processing the image with a semantic deep neural network model to determine the second classification associated with the image.

5. The method of claim 1 , further comprising:

processing the image with a convolutional neural network model to determine output data;

wherein utilizing the episodic memory to determine the first classification comprises:

processing the output data with an episodic deep neural network model to determine the first classification associated with the image; and

wherein utilizing the semantic memory to determine the second classification comprises:

processing the output data with a semantic deep neural network model to determine the second classification associated with the image.

6. The method of claim 5 , wherein processing the output data with the episodic deep neural network model to determine the first classification comprises:

determining features from the output data as query vectors;

determining features from a portion of the image as key vectors;

performing a dot product operation with the query vectors and the key vectors to determine similarity scores between the query vectors and the key vectors;

performing a softmax operation on the similarity scores to generate a result; and

determining the first classification associated with the image based on the result.

7. The method of claim 1 , wherein determining the accuracy associated with utilizing the semantic memory to determine the second classification comprises:

processing an output generated based on utilizing the semantic memory to determine the second classification associated with the image with a semantic confidence model to generate a confidence score indicating the accuracy associated with utilizing the semantic memory to determine the second classification.

8. A device, comprising:

one or more processors configured to:

receive input data for training a neural network model that includes a semantic deep neural network (DNN) model and an episodic DNN model;

process the input data, with a convolutional neural network model, to determine output data from the input data;

process the output data, with the semantic DNN model, to generate semantic output data;

process the output data, with the episodic DNN model, to generate episodic output data;

process the semantic output data, with a semantic confidence model, to generate confidence scores for the semantic output data;

apply weights to the semantic output data and the episodic output data, based on the confidence scores, to generate weighted output data;

combine the weighted output data to generate final output data;

generate a trained neural network model based on the final output data; and

perform one or more actions based on the trained neural network model.

9. The device of claim 8 , wherein the convolutional neural network model includes a residual networks model or a GoogLeNet model.

10. The device of claim 8 , wherein the one or more processors are further configured to:

receive additional input data;

retrain the neural network model based on the additional input data and to generate a retrained neural network model; and

cause the retrained neural network model to be implemented.

11. The device of claim 8 , wherein the one or more processors, when performing the one or more actions, are configured to one or more of:

implement the trained neural network model; or

utilize the trained neural network model to perform one or more tasks.

12. The device of claim 8 , wherein the one or more processors, when processing the output data, with the episodic DNN model, to generate the episodic output data, are configured to:

select a portion of the input data;

store the portion of the input data in a memory layer associated with the episodic DNN model; and

process the output data with the episodic DNN model based on the portion of the input data stored in the memory layer associated with the episodic DNN model.

13. The device of claim 12 , wherein the one or more processors are further configured to:

identify misclassified data in the final output data; and

replace the portion of the input data stored in the memory layer associated with the episodic DNN model with the misclassified data.

14. The device of claim 8 , wherein the one or more processors, when processing the output data, with the episodic DNN model, to generate the episodic output data, are configured to:

determine features from the output data as query vectors;

determine features from a set of the input data as key vectors;

perform a dot product operation with the query vectors and the key vectors to determine similarity scores between the query vectors and the key vectors; and

perform a softmax operation on the similarity scores to generate the episodic output data.

15. A non-transitory computer-readable medium storing instructions, the instructions comprising:

one or more instructions that, when executed by one or more processors, cause the one or more processors to:

receive input data to be processed by a neural network model that includes a semantic deep neural network (DNN) model and an episodic DNN model,

wherein the input data includes a plurality of images, and

wherein the neural network model processes the input data to detect one or more objects included in the plurality of images;

process the input data, with a convolutional neural network model, to determine output data from the input data;

process the output data, with the semantic DNN model, to generate semantic output data;

process the output data, with the episodic DNN model, to generate episodic output data;

process the semantic output data, with a semantic confidence model, to generate confidence scores for the semantic output data;

combine the semantic output data and the episodic output data, based on the confidence scores, to generate final output data; and

perform one or more actions based on the final output data.

16. The non-transitory computer-readable medium of claim 15 , wherein the one or more instructions, that cause the one or more processors to perform the one or more actions, cause the one or more processors to one or more of:

provide the final output data for display;

generate and implement a trained neural network model based on the final output data;

generate and utilize a trained neural network model to perform one or more tasks based on the final output data;

receive additional input data to further train the neural network model based on the final output data;

modify and retrain the neural network model based on the final output data; or

retrain one or more of the convolutional neural network model, the semantic DNN model, the episodic DNN model, or the semantic confidence model based on the final output data.

17. The non-transitory computer-readable medium of claim 15 , wherein the one or more instructions, that cause the one or more processors to process the output data with the episodic DNN model, cause the one or more processors to:

select a set of the input data;

store the set of the input data in a memory layer associated with the episodic DNN model; and

process the output data with the episodic DNN model based on the set of the input data.

18. The non-transitory computer-readable medium of claim 17 , wherein the one or more instructions, when executed by the one or more processors, further cause the one or more processors to:

identify misclassified data in the final output data;

select a set of the misclassified data; and

replace the set of the input data in the memory layer with the set of the misclassified data,

wherein the set of the misclassified data is utilized by the episodic DNN model to process additional output data determined by the convolutional neural network model based on the convolutional neural network model processing additional input data.

19. The non-transitory computer-readable medium of claim 15 , wherein the one or more instructions, that cause the one or more processors to process the output data, with the episodic DNN model, to generate the episodic output data, cause the one or more processors to:

determine features from the output data as query vectors;

determine features from a set of input data as key vectors;

perform a dot product operation with the query vectors and the key vectors to determine similarity scores between the query vectors and the key vectors; and

perform a softmax operation on the similarity scores to generate the episodic output data.

20. The non-transitory computer-readable medium of claim 15 , wherein the one or more instructions, when executed by the one or more processors, further cause the one or more processors to:

receive additional input data;

retrain the neural network model based on the additional input data and to generate a trained neural network model; and

cause the trained neural network model to be implemented.