The article is about tech firms and universities stocking up on research centers of AI experts, with the claim in its title that there is a high demand for those AI experts.
There might also be a demand for applied statisticians, but that doesn't make AI experts statisticians. I understand the confusion, as the term AI is often misused, but when you see the names mentioned in the article it's clear they're talking about actual AI researchers.
As mentioned in the article, AI is the broader field that encompasses Machine Learning, and to a large extent also Data science (and Computer vision, NLP, Pattern recognition, etc.). And while data science might utilize a lot of statistical techniques, it is a huge stretch to consider the whole AI field to be 'statistics'.
In general, AI borrows many more techniques from mathematics than it does from statistics. However, the field of AI has been quite established since the 1960's, and many techniques have been developed within that field as AI techniques, it's more about being accurate than about being fashionable as AI simply isn't 'just' statistics.
> If this model can provide an explanation for the small noises impacting NN performance on images (from karpathy.github.io, posted to HN earlier today) then that would be rocking
This paper [0] does a pretty good job of giving an explanation for that.
> A fairly easy way to introduce rotation invariance in DCNNS is to perform random rotations on the inputs during training. Likewise for scale invariance.
It is a bit silly to call these invariances, as different filter/kernel combinations will be activated when a rotated or scaled input is encountered, the individual filters are not rotation or scale invariant. The entire network can only deal with rotations and scales it encountered during training, whilst having to learn 'redundant features' to a certain extent.
It will get the job done for many tasks, but it's a brute force sort of approach that will complicate the learning process (i.e. more scales and rotations require more filters, thus needing a more complex network that is harder to train).
I think there's definitely a lot that can be learnt from (classical) signal processing in order to come up with a much more elegant and efficient solution.
>Some of them are very simple, and DO occur a lot in the world. For example, the alternating yellow and black line pattern would be encountered by a driverless car, and it would think it is seeing a school bus.
While the image shows a yellow and black line pattern to us, are you sure this is also what the CNN "sees"? Couldn't this image just be the same as the adversarial images, i.e. it responds to many small input values rather than the overall pattern?
If it's possible to make the CNN predict an ostrich for an image of a car, then the same can be done of an image of an alternating yellow and black line pattern, no?
There are a few things you should take into account:
1. You determine the 'dominant' colours to be the centroids of your clusters. The centroid is the mean of the points within the cluster, this mean is not necessarily a colour that is in your image. If you, for example, take a picture divided into four different solid coloured squares, and use this to find the 3 dominant colours it will average 2 (or more) colours. (The same might happen for more complex images with a a lot of contrast).
2. When randomly initializing k-means there is a good chance you'll find one of the local optima, so running it more than once will return different colours. In general it is good practice to run it several times and choose the outcome with the lowest cost.
3. K-means can take a long time to converge; limit the amount of iterations it can do.
These things aside, very cool usage of k-means on image data!
I really love the elegance of algorithms like these, a bunch of similar algorithms are discussed in Clever Algorithms[1].
It also discusses genetic algorithms[2], providing a really simple implementation that suffers from the same problem as the article in question; the outcome you're looking for is already known so the fitness function is rather trivial, but nonetheless its a good place to get started.
In the comments the author of the article refers to a paper[1] that describes the system (Ludi) that they're using. They came up with 57 aesthetics criteria which they used on games played by humans and compared the scores on the criteria to the human judgements.
Based on this they picked the 16 best criteria and generated games. The fitness function is a score based on the criteria after n amount of automated playthroughs.
These generated games tested by humans and they found that the aesthetics criteria and human judgements correlated significantly.
Besides these criteria (they do not describe all criteria in depth but I assume some might be leaning towards being subjective in regards to the programmers preferences) they also add a restriction to how long the planning of a move for a game can take (15 seconds) and discarding the game if it surpasses this. Which as you say can be seen as designing the search space, but the results indicate that they've come up with a way to effectively measure the "fun" factor of the games they came up with.
You can read on their website [1] what they did, and seems that it involved a lot of manual labour still.
[1] https://www.nextrembrandt.com/