System and method for training deep learning classification networks
Apparatuses and methods of manufacturing same, systems, and methods for training deep learning machines are described. In one aspect, candidate units, such as detection bounding boxes in images or phones of an input audio feature, are classified using soft labelling, where at least label has a range of possible values between 0 and 1 based, in the case of images, on the overlap of a detection bounding box and one or more ground-truth bounding boxes for one or more classes.
1. A method for deep learning training, comprising:
receiving a candidate unit for classification, the candidate unit including an intersection area between a ground-truth bounding box and a detection box;
classifying the candidate unit by labelling, wherein labelling comprises:
assigning a label of 0 when a value based on the intersection area is below a first threshold value;
assigning a label of 1 when the value based on the intersection area is above a second threshold value; and
assigning a label from a range of values between 0 and 1 when the value based on the intersection area is above the first threshold value and below the second threshold value, the label from the range of values being a probability value that a given feature is observed in the intersection area; and
performing deep learning training using the assigned label of the classified candidate unit.
2. The method of claim 1 , wherein the candidate unit is within an image.
3. The method of claim 1 ,
wherein labelling comprises:
providing a label of a class for the detection box based at least partially on the intersection area for the class.
4. The method of claim 3 , wherein providing a label of a class comprises:
assigning a class label whose value is derived using the intersection area.
5. The method of claim 3 , wherein providing a label of a class comprises:
assigning a class label whose value is derived from a ratio involving the intersection area.
6. The method of claim 5 , wherein assigning a class label comprises:
calculating the ratio of the intersection area to an entire area of the detection box.
7. The method of claim 3 , wherein providing a label of a class for the detection box is also based on one or more confidence levels provided by a detection stage which also provided the detection box.
8. The method of claim 3 , wherein providing a label of a class for a detection box comprises:
providing a label of a first class for the detection box based at least partially on the intersection area for the first class; and
providing a label of a second class for the detection box based at least partially on the intersection area for the second class.
9. The method of claim 3 , wherein the value based on the intersection area for the class is a ratio of the intersection area to an entire area of the detection box.
10. The method of claim 1 ,
wherein labelling comprises:
generating soft labels directly from classification scores from a probability model or neural network.
11. The method of claim 1 ,
wherein labelling comprises:
generating soft labels directly from classification scores from a hidden Markov Model (HMM), a Gaussian mixture model (GMM), or a pretrained neural network.
12. The method of claim 1 ,
wherein labelling comprises:
generating soft labels using maximum likelihood decoding, a distance metric, a soft output decoding algorithm, or a list decoding scheme.
13. An apparatus for deep learning training, comprising:
one or more non-transitory computer-readable media; and
at least one processor which, when executing instructions stored on one or more non-transitory computer readable media, performs the steps of:
receiving a candidate unit for classification, the candidate unit including an intersection area between a ground-truth bounding box and a detection box;
classifying the candidate units by labelling, wherein labelling comprises:
assigning a label of 0 when a value based on the intersection area is below a first threshold value;
assigning a label of 1 when the value based on the intersection area is above a second threshold value; and
assigning a label from a range of values between 0 and 1 when the value based on the intersection area is above the first threshold value and below the second threshold value, the label from the range values being a probability value that a given feature is observed in the intersection area; and
performing deep learning training using the assigned label of the classified candidate unit.
14. The apparatus of claim 13 ,
wherein labelling comprises:
providing a label of a class for a detection bounding box based at least partially on the intersection area.
15. The apparatus of claim 13 ,
wherein labelling comprises:
generating soft labels directly from classification scores from a probability model or a neural network.
16. A method, comprising:
manufacturing a chipset capable of deep learning training comprising:
at least one processor which, when executing instructions stored on one or more non-transitory computer readable media, performs the steps of:
receiving a candidate unit for classification, the candidate unit including an intersection area between a ground-truth bounding box and a detection box;
classifying the candidate units by labelling, wherein labelling comprises:
assigning a label of 0 when a value based on the intersection area is below a first threshold value;
assigning a label of 1 when the value based on the intersection area is above a second threshold value; and
assigning a label from a range of values between 0 and 1 when the value based on the intersection area is above the first threshold value and below the second threshold value, the label from the range of values being a probability value that a given feature is observed in the intersection area; and
performing deep learning training using the assigned label of the classified candidate unit; and
the one or more non-transitory computer-readable media which store the instructions.
17. A method of testing an apparatus, comprising:
testing whether the apparatus has at least one processor which, when executing instructions stored on one or more non-transitory computer readable media, performs deep learning training comprising the steps of:
receiving a candidate unit for classification, the candidate unit including an intersection area between a ground-truth bounding box and a detection box;
classifying the candidate units by labelling, wherein labelling comprises:
assigning a label of 0 when a value based on the intersection area is below a first threshold value;
assigning a label of 1 when the value based on the intersection area is above a second threshold value; and
assigning a label from a range of values between 0 and 1 when the value based on the intersection area is above the first threshold value and below the second threshold value, the label from the range of values being a probability value that a given feature is observed in the intersection area;
performing deep learning training using the assigned label of the classified candidate unit; and
testing whether the apparatus has the one or more non-transitory computer-readable media which store the instructions.