Generative Adversarial Networks (GAN)
GAN[ii] was introduced in 2014 by Ian Goodfellow. This technique combines generative and discriminative data models that are simultaneously trained and pitted against each other - like a two player Min-Max game. The generative model is trained to generate synthetic data from random noise while the discriminative model is trained to classify if the generated data is part of the training data or not i.e. to make the generator rollout better data.
Let us illustrate this with an example. If we want to train a binary classifier on 50 independent variables with say, 1 million annotated data samples, it essentially leaves a gap of 250 - 106 possible occurrences in the real world whose classes cannot be trained in the model. The generator can fill this gap by generating sample data points, which the discriminator will ensure are correct and facilitate accurate overall prediction.
GANs are being applied in many innovative ways like generating images from text description[iii] and cancer drug discovery[iv]. Apart from applications in industrial domains, IT service providers can apply them to solve complex tasks like generating project plans from requirement documents or code from use cases.
Transfer Learning[v] leverages knowledge from one or more related tasks to achieve higher learning performance by:
- Having a higher initial performance in the target task as compared to that of an untrained agent
- Requiring lesser time to learn the target task
- Achieving higher final performance in the target task as compared to the performance achieved without transferred knowledge
There are several popular pre-trained models available in computer vision and natural language processing such as:
- Keras[vi] – It has models like Microsoft ResNet50 and Google Inception-V3 for image recognition. These models are trained on ImageNet[vii], an annotated library of 14 million images
- Google - It offers word2vec[viii], a pre-trained model for natural language processing which is trained in 100 billion words
To conclude, the techniques discussed in this article enable new levers for IT service providers, helping them to develop recombinant data strategies and machine learning models to create innovative offerings in the AI-ML space.