With hundreds of variations in publications, Generative Adversarial Networks, or GANs, have been creating a big wave in the AI and machine learning research community since their invention by Ian Goodfellow in 2014. Yann Lecun, vice president and chief AI scientist for Facebook, and founding father of convolution neural networks (CNN), noted, “The (GAN) variations that are now being proposed is the most interesting idea in the last 10 years in machine learning.”
How can GANs help us develop better products and bring value to our customers? At Climate LLC, our machine learning Data Scientists are exploring the potential of GANs in our disease identification model.
Applications for a Disease Identification Model
1. An accurate, robust and generalized mobile phone camera-based disease identification model is necessary to deliver innovative new value to farmers who use our digital farming platform, FieldViewTM; however, several challenges stand in the way. Different data sources have different fields of view and lighting conditions, the availability of different disease types is imbalanced, and disease syndromes vary across environments, hybrids and stages.
Collecting enough images for each of these variations is nearly impossible, but a lack of data variations leads to undesirable model generalization. In order to improve model performance, many variations of data augmentation techniques have been adopted, including random cropping, horizontal flipping, and shearing. They can boost the model generalization, but the resulting augmented images are highly correlated with the original sample. This is where we can take advantage of GANs to generate practical disease images based on limited real disease information in our hands. GANs provide a good estimation of the real data distribution by randomly generating samples according to the true distribution. By doing this we are enriching our dataset and saving expensive efforts in manual imagery collection.
Training vanilla GANs is a fairly heuristic process with obstacles such as unstable training, mode collapse, and gradient vanishing. Unstable training and mode collapse are the consequences of the flawed cost function which tries to optimize towards two opposite directions. Gradient vanishing is due to the non-overlapping supports between the true and generator distribution, which makes the distance between the two a constant. Therefore, we used Wasserstein-GAN(WGAN), a famous GANs variation, that enjoys stable training and good image quality. In order to convert the 100-d noise vector to an image size, several upsample steps are needed across different layers.
The WGAN discriminator is very similar to a typical CNN: The model takes the image as the input, feeds into many convolutional layers and outputs the prediction. The cost function of WGAN discriminator is Wasserstein distance rather than KL-divergence and JS-divergence that vanilla GANs use. Wasserstein distance is valid when the supports of two distributions are non-overlapping, which is exactly the scenario that GANs face. The discriminator in WGAN estimates the Wasserstein distance between true and generated distributions, while the generator tries to generate realist images to minimize such distance. Instead of the classification task that vanilla GANs discriminator conduct, WGAN discriminator actually performs a regression task. Such conversion proves to be very stable during the training, leading to a very encouraging generator performance.
As we introduced earlier, GANs can estimate the data distribution by generating data samples, enabling a better understanding of our data structure, and facilitating further mining of our data. Such exploration can lead to other GANs application, including semi-supervised learning.
Even though the unlabelled data couldn’t give a clear boundary of the final classification in a direct way, the underlying structure of the unlabelled data learned by GANs will help the classifier to gain more information when finding the boundaries of different classes. However, relevant literature only showed success in small image sizes, 64x64 pixels, and in disease modeling we target over 200x200 pixels, which can pose a big challenge on training the GANs. The necessity for a larger image is to identify details of disease symptoms and recognize the subtle differences. A larger image size will complicate the GANs model structure and manifest the unstable training, leading to unexpected model performance. An alternative is to use WGAN to support larger image sizes and more complicated architectures. Unfortunately, the discriminator in WGAN doesn’t act as a classifier. Instead, it tries to estimate the distance between the true data distribution and the generated data distribution, which is only suitable for generating realist images. Meaning, the motivation of the whole WGAN model doesn’t fit into the semi-supervised framework. We are actively monitoring GAN communities and are working for a solution.
Through all the detours, we learn that careful and thorough research needs to be carried out before we adopt any cutting-edge AI technology. We are not just AI executors, but rather AI strategists. The Data Scientists on our Geospatial Science Team and throughout Climate LLC are dedicated to delivering new features in the FieldView™ platform that benefit our customers by fusing cutting-edge AI technologies together with domain knowledge.
*Nag, Dev (2017) “Generative Adversarial Networks (GANs) in 50 lines of code (PyTorch)”
* Radford, Alec & Metz, Luke (2016) “Unsupervised Representation Learning With Deep Convolutional Generative Adversarial Networks”
Retrieved from https://arxiv.org/pdf/1511.06434.pdf