Background
For the last paper, I have only few datasets, one AP in CUICAR suggested me to use GAN to generate more images to improve the accuracy of model.
Now, I have time to read some articles about GANS,also I want to use time signal with deep models.
URLs:https://medium.com/@alexrachnog/gans-beyond-generation-7-alternative-use-cases-725c60ba95e8
7 alternative use cases
1. Data augmentation
two main strategies: we can train our model on “fake” data and check how well it performs on real samples. And the opposite: we train our model on real data to do some classification task and only after check how well it performs on generated data (GAN fans can recognize inception score here)
If it works well in both cases — you can feel free to add samples from generative model to your real data and retrain it again
Example:https://arxiv.org/pdf/1803.01229.pdf NVIDIA used GANs to augment dataset of medical brain CT images( data augmentation) to improve sensitivity .
2. Privacy preserving
This way we won’t share any exact confidential data, just something what looks exactly like it.
3. Anomaly detection
(see https://arxiv.org/pdf/1802.06222.pdf)
Yes, Dr. Zhao(CU ICAR) use Mobilenet with Lidar data to detect abnormal information. And his idea is generated from NIVIDA's paper which used images of three camera to predict steer angle.
Also Dr. Zhao used Particle filter to recognize error signal from AVs, which had bee simulated in Prescan software.
4. Discriminative modeling
First one, already researched, is leveraging discriminator to perform classification of a generated sample into different classes alongside with just telling if it’s real or fake.
Second is unfairly forgotten approach to classification with Bayes’ theorem, where we model p(c_k|x) based on p(x|c_k) (exactly what conditional GAN does!) and priors p(x), p(c_k).
(https://openreview.net/forum?id=BJehNfW0-)
5. Domain adaptation
6. Data manipulation
see gif : https://houxianxu.github.io/assets/project/dfcvae
7. Adversarial training
simply makes input perturbations with some epsilon
make a model more stable to adversarial attacks and improves its performance due to some kind of smart data augmentation and regularization.
I just read and cited some important sentences from his blogs.
I may use the following properites to improve my detection model.
1. Data augmentation,
5. Domain adaptation
7. Adversarial training