Isn’t this problem fairly trivially solvable without any AI/ML at all? Especially now that it’s constrained to one axis? With training images falling in small increments across the whole range of rotations, one could simply find the most similar images in the training set for each image in the test set, and interpolate to get an even closer approximation of the rotation angle.
Are we required to use a learning-based approach? Will there be a second modification to make this more difficult again? Looking forward to clarifications 
2 Likes
Hi @Johnowhitaker ,
Thats a great observation, and we are very aware of that.
And infact, you could argue that the imputation based approach will work equally well even if you had to predict the rotation angles across three axes.
In any case, this still requires you to understand the notion of a dense “representation” of each of the data points (images of rubiks cube in this case), to be able to interpolate from the training dataset. And we choose a rubiks cube here for a reason : so that participants who are not experienced in AI/ML space can still intuitively try and come up with good representations based on a set of simple rules. The future versions of the task (for the 2nd AIcrowd Blitz scheduled to be launch a week after this one ends) will have arbitrary models which will make that representation learning task a bit more difficult.
But, if you are really into a research-grade-hard-variation (
) of this problem, you should definitely check the NeurIPS 2019 Disentanglement Challenge that we put together to push the state of art in Representation Learning.
But in any case, the goal of these problems in AIcrowd Blitz are not to come up with the best possible solution which solves the said problem, but actually to introduce a series of problems of varying difficulty levels which encourage a lot more people to engage in this whole space of AI/ML.
While problems like these might seem trivial (after cleverly connecting the initial dots in the problem statement), the simple and key ideas here still continue to be major barriers to entry for hundreds and thousands of potential “research collaborators” who never make the leap to try and understand the basic primitives in our domain, because it always looked so daunting and intimidating. Infact, as soon as the competition ends on May 17th, we have a dedicated 1 week of dissecting all the problems mentioned here as a community - but theres a catch - here the goal would be to try and explain the solutions to all these problems to an absolute beginner, who has no experience in Machine Learning, but happens to be a exceptionally great programmer.
Also, its interesting to highlight the story of this paper : https://arxiv.org/pdf/1505.04467.pdf around Image Captioning. This paper came out a year or so later after the First Image Captioning model made waves, in media, in academic circles, on twitter, everywhere !! This was written by the (almost) same authors, but highlighted a very key idea that being able to impute efficiently on your training set, using ridiculously simple ideas, might help you solve rather (seemingly) complex tasks with just simple ideas !
My personal vision here is to build a community here around AIcrowd Blitz, which acts as a safe space for anyone who is willing to learn about AI/ML while using these problems as encouraging milestones. And I am hoping to gather a LOT MORE experienced researchers to come and patiently mentor the next generation of AI Researchers, irrespective of where they come from, their academic pedigree, their age, or their domain of expertise.
And given your thoughtful inputs and also the participation in this competition, I really think you should definitely come help us make many of these ideas accessible to the many who are still daunted by them
4 Likes