Apparatus and method for unsupervised domain adaptation
An apparatus is for unsupervised domain adaptation for allowing a deep learning model with supervised learning on a source domain completed to be subjected to unsupervised domain adaptation to a target domain. The apparatus includes a first learning unit to perform a forward pass by inputting a pair (x s i , y s i ) of first data x s i of the source domain and a label y s i for each of the first data and second data x T j belonging to the target domain, and insert a dropout following a Bernoulli distribution into the deep learning model in performing the forward pass, and a second learning unit to perform a back propagation to minimize uncertainty about the learning parameter of the deep learning model by using a predicted value for each class output through the forward pass and the label y s i , and an uncertainty vector for the second data x T j output through the forward pass as inputs.
1. An apparatus for unsupervised domain adaptation for allowing a deep learning model with supervised learning on a source domain completed to be subjected to unsupervised domain adaptation to a target domain, the apparatus comprising at least one processor and a computer-readable storage medium storing one or more programs including one or more computer-executable instructions executed by the at least one processor:
a first learning unit implemented by the at least one processor and the one or more computer-executable instructions executed by the at least one processor, the first learning unit configured to perform a forward pass by respectively inputting a pair (x s i , y s i ) of a plurality of first data x s i belonging to the source domain and a label y s i for each of the first data and a plurality of second data x T j belonging to the target domain, and insert a dropout following a Bernoulli distribution, which is a trial probability p, into the deep learning model in a process of performing the forward pass; and
a second learning unit implemented by the at least one processor and the one or more computer-executable instructions executed by the at least one processor, the second learning unit configured to perform a back propagation to minimize uncertainty about a learning parameter of the deep learning model by respectively using a predicted value for each class output through the forward pass and the label y s i , and an uncertainty vector for the second data x T j output through the forward pass as inputs,
wherein the first learning unit is further configured to iteratively perform the forward pass T times for one input value;
the predicted value for each class is an average value of T score vectors for a class output when the pair (x s i , y s i ) of the first data x s i and the label y s i is input to the deep learning model; and
the uncertainty vector is a standard deviation of T score vectors for a class output when the second data x T j is input to the deep learning model.
2. The apparatus of claim 1 , wherein the first learning unit is further configured to insert the dropout in a Monte-Carlo sampling scheme.
3. The apparatus of claim 1 , wherein a layer a layer into which the dropout is inserted first comprises a feature generator and a classifier into which an output of the feature generator is inputted, and the dropout is inserted into the classifier prior to performing the forward pass.
4. The apparatus of claim 3 , wherein the second learning unit is further configured to adjust the learning parameter of the deep learning model by performing the back propagation so that a value of a loss function for the predicted value for each class and the label y s i is minimized.
5. The apparatus of claim 3 , wherein the second learning unit is further configured to adjust the learning parameter of the deep learning model by performing the back propagation in a manner of setting the value of the loss function for the uncertainty vector as an uncertainty index and allowing the uncertainty index to be learnt by the feature generator and the classifier, respectively, with Mini-Max.
6. The apparatus of claim 5 , wherein the uncertainty index is a value obtained by taking L1 Norm or L2 Norm of the uncertainty vector.
7. The apparatus of claim 1 , further comprising:
an inference unit implemented by the at least one processor and the one or more computer-executable instructions executed by the at least one processor, the inference unit configured to perform inference through the deep learning model after the dropout is removed after the unsupervised domain adaptation of the deep learning model has been completed.
8. A method for allowing a deep learning model with supervised learning on a source domain completed to be subjected to unsupervised domain adaptation to a target domain, the method comprising:
performing a forward pass by respectively inputting a pair (x s i , y s i ) of a plurality of first data x s i belonging to the source domain and a label y s i for each of the first data and a plurality of second data x T j belonging to the target domain;
inserting a dropout following a Bernoulli distribution, which is a trial probability p, into the deep learning model in a process of performing the forward pass; and
performing a back propagation to minimize uncertainty about a learning parameter of the deep learning model by respectively using a predicted value for each class output through the forward pass and the label y s i , and an uncertainty vector for the second data x T j output through the forward pass as inputs,
wherein the performing of the forward pass comprises iteratively performing the forward pass T times for one input value;
the predicted value for each class is an average value of T score vectors for a class output when the pair (x s i , y s i ) of the first data x s i and the label y s i are input to the deep learning model; and
the uncertainty vector is a standard deviation of T score vectors for a class output when the second data x T j is input to the deep learning model.
9. The method of claim 8 , wherein the inserting of the dropout comprising inserting the dropout in a Monte-Carlo sampling scheme.
10. The method of claim 8 , further comprising:
wherein a layer into which the dropout is inserted first comprises a feature generator and a classifier into which an output of the feature generator is inputted, and the dropout is inserted into the classifier prior to performing the forward pass.
11. The method of claim 10 , wherein the performing of the back propagation comprises adjusting the learning parameter of the deep learning model by performing the back propagation so that a value of a loss function for the predicted value for each class and the label y s i is minimized.
12. The method of claim 10 , wherein the performing of the back propagation comprises adjusting the learning parameter of the deep learning model by performing the back propagation in a manner of setting the value of the loss function for the uncertainty vector as an uncertainty index and allowing the uncertainty index to be learnt by the feature generator and the classifier, respectively, with Mini-Max.
13. The method of claim 12 , wherein the uncertainty index is a value obtained by taking L1 Norm or L2 Norm of the uncertainty vector.
14. The method of claim 8 , further comprising:
performing inference through the deep learning model after the dropout is removed after the unsupervised domain adaptation of the deep learning model has been completed.