Weather conditions are usually spoken in terms of sunny or rainy. Simplification works fine for us, humans. But, if you task an autonomous system like a car to make changes in accordance with the weather, inputs such as sunny and rainy won’t do any good. For instance, a computer vision system is assigned to take the inputs, then it can be fooled when it is presented with an image of the sun amid clouds. The performance model in the real world is bounded by our imagination that decides the constraints and classes. There is a significant difference between weather conditions such as overcast cloudy, the transition state of rainy to non-rainy, light snow versus light rain, etc.
Another example can be the datasets on digits recognition (ex: SVHN, MNIST), which mainly differ from each other in terms of backgrounds and text fonts. But, that doesn’t mean that these datasets should be considered as distinct domains.
The traditional domain adaptation methods fall short when they come across a feature engineering nightmare such as weather.
Therefore, there is an immense need to rethink the way we imbibe domain adaptation into machine learning systems. There have been many few works such as domain generalization and latent domain adaptation, which have been used to tackle complex target domains. However, these methods work on the assumption there is a known clear distinction between domains.
Now, the researchers at UC Berkeley introduced a continuous learning protocol under domain adaptation scenario.
Open Compound Domain Adaptation (OCDA)
The researchers claim that the Open Compound Domain Adaptation (OCDA) is the right candidate for domain adaptation in a realistic setting. The objective here is to train a model from labelled source domain data and adapt it to unlabeled compound target domain data. This can then be used to draw insights by measuring how much the results have deviated from the source domain on various factors.
As shown above, the compound target domain pools the data together. And, at the inference stage, OCDA tests the model both in the compound target domain and open domains, which probably have been unseen during training.
Here, the target domain can be considered as the combination of multiple unlabeled traditionally homogeneous domains where each is distinctive on one or two major factors.
The researchers propose a novel approach to enable OCDA, which works as follows: –
- Separate the characteristics specific to domains from those discriminative between classes
- The separated domain feature is then used to construct a curriculum for domain-robust learning
- Enhance the network with a memory module that facilitates knowledge transfer from the source domain to target domain instances, so that the network can leverage the inputs and memory even in the case of previously unseen domains
The authors state that the experiments conducted on the digit classification, facial expression recognition, semantic segmentation, and reinforcement learning demonstrated the effectiveness of this novel approach.
Although supervised learning has been proven to perform well on computer vision tasks, the real world poses new challenges, which cannot be tackled by traditional methods.
The whole aim of a domain adaptation approach pivots around the notion of making the ML model learned on the training data adapt to the test data of a different distribution. The challenge here is the distributional gap between discrete concepts of well-defined data domains; images obtained in sunny conditions compared to those in rainy weather. The OCDA approach is designed to tackle all these challenges and provide a realistic touchstone for evaluating domain adaptation and transfer learning systems.