Since deep learning became one of the most promising technologies of 21th century, researchers relentlessly try to solve new challenges, and expand network abilities: train better, faster, more general, more accurate models for ever-growing set of tasks.
In this setting, the main bottleneck seems to be the data, and more specifically, the annotations. Indeed, ImageNet is one of a kind, and most data-sets do not exceed 10k images in case of classification tasks, and much lower when it comes to detection and segmentation etc. Which are more challenging to annotate.
This talk will discuss techniques that try to overcome this bottleneck in very creative ways, and specifically the self supervised approach, which tries to extract signal from data that was not annotated.
Some of the ideas we will cover:
* Learning from visual context
* Video as context
* GANs as self supervised technique