Deep Learning for Imaging. ML at laptop scale vol II

When an aspiring deep machine learning practitioner starts learning about neural nets, especially in the context of image recognition, he/she would typically start with something like MNIST (handwritten digits) dataset. It is small enough to train corresponding network on a laptop PC, yet it allows to see how to use convolutional neural nets and even peek at “what’s happening inside”.

Then there are bigger datasets, like ImageNet. It has more images, more classes, but structurally it’s not very different. The task is still assigning one label for a given image, and images are fairly well centered and properly cropped. On the other hand, there are many other tasks in image processing, like image segmentation, counting objects in the picture, even using some spatial reasoning about image content. Many of these methods are described in corresponding papers and often illustrated with unique datasets which are appropriate for a given task. Sometime these datasets are small, sometimes really big and most often - they are pretty different from each other: from pictures of cats and dogs, to shots from street cameras to satellite imagery.

Targets of this Hackathon

The goal of this episode of Tokyo Machine Learning Gym is to work on a set of simple cases which would illustrate how different features in deep neural net architectures allow to solve structurally complex tasks/datasets. Our focus is twofold:

• learning/improving models allowing to solve various complex tasks by using features like visual attention and adversarial approaches using “small but tricky” datasets problems, selected or created by organizers and improved by participants during the hackathon.

• Creating/ improving abovementioned problem sets, making them small in size yet requiring complex solutions.

This is non-exclusive set of problems we prepare for this hackathon:

• basic image recognition

• image segmentation / object localization

• visual attention

• image generation

• object counting

• judging about spatial relations between objects

Format

Hackathon will run one day from 10am to ~ 5pm and will be organized in two sessions: hands-on mini-tutorial in the morning and free hacking sessions after the lunch break.

In the tutorial, organizers will introduce a set of pre-selected/pre-made small datasets and walk you briefly through different features of neural net architectures with the focus of which problems they allow to solve.

For the free hacking session we expect the participants to form small (3-6 persons) teams to work on selected problems. We would provide a set of pre-selected problems in as well as “starter” codes. We expect several mentors to be on site to help. However, people are very welcome to propose own ideas and lead own groups.

We will conclude by sharing and discussing results.

Participants

Minimal deep learning proficiency is required for this hackathon: be able to create simple convolutional neural net and train it on mnist-like dataset. Starting from there any level participants are welcome. Beginners can experiment with starter codes we provide or follow selected team leaders; seasoned deep learning hackers welcome to propose own ideas and lead other people.

Please keep in mind: this is not pure tutorial where you come and people teach you everything. You are expected to try doing things by yourself, although we will help as much as possible.

Few words about “starter” codes: data handling/generation is implemented in Python and is deep learning framework- agnostic. Sample/baseline solutions are implemented with Chainer and partially with Keras, but you are welcome to use DL framework on your choice if you feel confident enough.

Compute resources

This hackathon is titled “ML on a laptop scale” and indeed we expect all workloads to be small enough to be trainable in reasonable time on laptop CPU / integrated graphics cards. Nevertheless, we will provide several [think one per team] GPU servers for those resource-hungry (And well, it’s more fun this way). In any case don’t forget your laptop's, those servers will be remote, there will be no PCs provided at the venue!

Space constraints

Due to the limited space at the venue we have to restrain the number of attendees to ~16 persons, including organizers and invited mentors.

But please, people who could not get in, don’t despair. We will publish problems and “starter” codes online before the hackathon and also start an online discussion channel in parallel to the offline hackathon, so that you can participate right from your couch.

Photo and video

Video footage and/or photos might be taken during this event, which may or may not include your recognizable image. Photos will be shared with all participants so that we have more fun and fond memories from the event ;) If you do not wish to have us use your image, please notify organizers.