Transfer learning methods and systems for feed-forward visual recognition systems
A method and system for training a neural network of a visual recognition computer system, extracts at least one feature of an image or video frame with a feature extractor; approximates the at least one feature of the image or video frame with an auxiliary output provided in the neural network; and measures a feature difference between the extracted at least one feature of the image or video frame and the approximated at least one feature of the image or video frame with an auxiliary error calculator. A joint learner of the method and system adjusts at least one parameter of the neural network to minimize the measured feature difference.
1. A method for training a neural network of a visual recognition computer system, the method comprising the steps of:
first training the neural network by selecting training image labels of training images stored in a database; and
supplementing the first training with a second training the neural network on the images by extracting features of each input training images stored in the database, including:
predicting at least one class label of the image or video frame with the neural network;
approximating at least one feature of the image or video frame with the neural network;
determining feature recognition errors by measuring a label difference between the selected training image labels and the predicted at least one class label of the image or video frame, and measuring a feature difference between the extracted at least one feature of the image or video frame and the approximated at least one feature of the image or video frame; and
jointly learning the selected training image labels and the feature recognition error values; and
tuning parameters of the neural network to jointly minimize the training image labels and the feature recognition error values.
2. The method of claim 1 , further comprising the step of selecting a class label for the training image or video frame.
3. The method of claim 2 , wherein the class label selecting step is performed manually by a human expert.
4. The method of claim 1 , wherein the class label selecting step is performed manually by a human expert.
5. The method of claim 1 , wherein the neural network comprises a hierarchical feedforward neural network.
6. The method of claim 1 , wherein the extracting step is performed by a feature extractor.
7. The method of claim 1 , wherein the approximating step is performed by an auxiliary output layer of the neural network.
8. A system for training a neural network of a visual recognition computer system, wherein the neural network is trained on, the system comprising:
a labeler for selecting training image labels of training images stored in a database:
a feature extractor for extracting at least one feature of each input training images stored in the database;
an output layer provided in the neural network for predicting at least one class label of the image or video frame;
an auxiliary output layer provided in the neural network for approximating the at least one feature of the image or video frame;
an error calculator for measuring a label difference between the selected training image labels and the predicted at least one class label of the image or video frame;
an auxiliary error calculator for measuring a feature difference between the extracted at least one feature of the image or video frame and the approximated at least one feature of the image or video frame; and
a joint learner operating on training image labels and the feature recognition error values, the joint learner tuning parameters of the neural network to jointly minimize the training image labels and the feature recognition error values of the error calculator and auxiliary error calculator.
9. The system of claim 8 , further comprising an expert labeler for allowing selection of a class label for the training image or video frame.
10. The system of claim 8 , wherein the joint learner also adjusts the at least one parameter of the neural network to minimize the measured label difference.
11. The system of claim 8 , wherein the neural network comprises a hierarchical feed-forward neural network.