It is incomprehensible to you, because you just simply do not understand what your parent is talking about. You are the ignorant one here and indeed quite rude. Doesn't matter that genetics is not natural language. The point is we can train large transformers auto regressively and the representation it learns turns out to be useful for a) all kinds of supervised downstream tasks with minimal fine-tuning and b) interpreting the data by analysing the attention weights. There is a huge amount of literature on this topic and what your parent says is quite sensible.
Agreed, I trained a 3D version of b0-b2 on a classification task I worked on and besides being very slow to train they did not outperform a simple baseline VGG architecture. Interestingly training time was much improved by setting the cudnn benchmark flags in pytorch. Haven’t seen this reduction in training time for any other architecture but for b2 it went from about 12 sec per iteration to about 0.8. I guess there is more margin optimizing NAS models.
Email: nils [dot] eckstein [at] googlemail [dot] com
Machine Learning Researcher in the last stage of my PhD, where I designed terabyte scale computer vision algorithms for Connectomics (Neuroscience). Looking for Machine Learning Research/Engineer positions and interesting problems.
1. Isn't an issue. They make inference on a sample by sample basis. The network has no memory so it won't expect a 50/50 distribution on the test set just because its trained like that. Having a balanced distribution is the exact right thing to do because you do not want the network to be biased to one or the other class for any given sample. If it were unbalanced the network could achieve almost 0 training error by just predicting negative all the time. This is not what you want.
My main concerns with the imbalance are undersampling of the negative class data distribution relative to the positive class, and overestimating performance on the test splits. I can buy that you may want to train on a balanced dataset, but the testing condition should reflect the true case distribution as closely as possible.
I agree that you would not want to use only the class priors for prediction. However, I do not think it is clear that you would want to throw that information out. Also not sure that I agree with the statement that neural network has “no memory” of the prior class distribution. That is a strong claim to make about something as opaque as a neural net model.
They could have used all negative samples for testing (and even training if they would have done it better), yes. But once your test set is large enough, whatever that means, its not that relevant anymore. They are anyway "under sampling" by not recording data from all humans that are negative right now.
And no, it's not a strong claim to make. Of course the network learns the distribution of your training set. That's why you want it balanced. But during successive applications of inference the weights do not change, it has no state. So it cannot, for example, store that it just predicted 90% negative and now it would be time again for some positive prediction.
The fraction of people in AI working on problems that need to consider diversity/fairness/etc. is rather small. Yes, those people, these specific applications, should be designed with care, should be overseen properly and in the best of all worlds should be unbiased. However, the recent discussion around this topic is framed like AI research in general is somehow unethical and should be informed more by minority populations. While this is something to strive for, for the sake of having a more just society, the impact of diversity on an AI system that say classifies mitochondria in yeast is not clear to me. I’d argue the majority of problems in AI right now are of this form and not of the form addressed by the author. If you want to push quotes, increase minority representation for occupations, do so. If you want to address specific biases in very specific applications then do that. Don’t use the latter to make an argument for the former though. These are two different issues and while I recognize the author sees a link here I’m not so certain about that. To be clear I think that both are worthy goals. But it’s somewhat disingenuous to make an inherent political argument, people may disagree with, and justify it by overgeneralizing a specific, niche technological ethics question to an entire field.
I don't think you're making much of an effort to describe the variety of AI that are appearing. The thing about AI is it thrives on data and lots of large institutions have data that they use to make decision about people. Fairness questions come in.
Google randomly, I see:
AI Applications: Top 10 Real World Artificial Intelligence Applications
Marketing, Banking, Finance, Agriculture, Health Care, Gaming, Space Exploration, In Autonomous Vehicles.
About half of those seem like situation where fairness questions enter - banking and finance - plausibly (who gets a loan), marketing - plausibly (who gets sold what), health care - plausible (who gets treated, what groups' data is and isn't used for what, etc). The other not so much.
Well, academia pretty has to study AI applications.
AI is not a field like physics, which can be roughly separated into theoretical tools and applications of those tools.
AI is creating heuristics, approximations to data that "generalize" while keep how that generalization works vague. Essentially it's a very leaky abstraction so researchers need to be concerned how that leaking happens, what it's implications are.
I happen to work in AI research and what you are saying isn’t true. There is theoretical machine learning and applications of it. They are distinct.
The former is largely task agnostic and deals with fundamental issues such as, how to train networks in an unsupervised way, how do you do hard statistical inference for intractable models in a approximate but well enough manner, how is gradient descent behaving exactly, why does it work, are there better ways for optimizing nns, what about pruning? The list goes on.
On the other hand you have applications of AI to other fields that require their own research. In biology for example you may want to segment cells and their compartments or design better point spread functions for you microscope or classify cell types. These are applied problems.
Again, as stated in my original post, striving for more diversity is a good thing and should and is done. Why make it about AI ethics and bias though when large portions of this field have no contact point with it?
In your research, how do you define the process of "generalizes"?
I understand theoretical research exists but I think it's problematic that theoretical researchers imagine that a kind of "generic" problem exists, even when a variety of test sets exist to
I mean, is SOTA on imagenet or whatever data a theoretical or an applied question? What theoretical research in AI is so theoretical that the question of data sets doesn't appear?
Let me be very concrete: There is currently a lot of research being done in so called “contrastive learning”. This is an unsupervised technique in which you train a network to build good representations of its input data. Not by explicitly telling the network, “this is a dog” but by telling it “this image is different from that image and similar to this other one”.
How this works, why this works, coming up with the technique itself, are all data agnostic. All you did so far is write down a function f(x) -> y and a loss L(x,y), with x the input and y the output, specifying your model.
Of course you use a specific dataset to train your model in the end and see if it works. But the model and the technique itself are not grounded in any specific dataset and thus nothing in this model perpetuates bias.
Now usually the next step for you as a researcher is to evaluate the performance of your model on the test set. Lets take image net. Now there are 3 situations.
A) Your Train set is biased & your test set is biased in the same way.
B) Your train set is biased & your test set is not.
C) Your train set is unbiased & your test set is biased.
With biased I mean any kind of data issue such as discussed in this article, e.g. no women in the class “doctor”.
In situations B) & C) your model wont work well so you actually have an incentive to fix your data. This will happen to you in production if you train your tracker only on white people say.
Situation A) is likely what’s happening with imagenet and other benchmark datasets. In this case your model learns an incomplete representation of the class “doctor” and learns the spurious correlation that all doctors are men. This will work on the test set because it’s equally biased.
You go, get good test results and publish a paper about it, unaware of the inherent dataset biases. (You could have done all this also on MNIST or a thousand other datasets that do not have any issues with societal biases because they are from a totally different domain, but that’s another point).
In this entire process of coming up with the model, training and evaluating it, there is no point at which the researcher has any interest, or benefits from, working with biased datasets. Furthermore, besides potentially overestimating the accuracy of your model, there is nothing in here that would hurt society or further perpetuate biases. That is because models are generally not designed to work on a specific dataset.
Again, this is a different story when you use your model in production. In this case you are in situation B) or C) and here now lies the crux. If you can make money from this bias or maybe it perpetuates your own biases well you might keep it like that. This should be fixed. Here now is a real argument for why there should be diverse populations working on AI systems that are used in industry.
Of course having diverse populations in science is also our goal. But not to fix our datasets but to do better research.
This is a very one-sided view of what this article is talking about. That there is systemic racism in the US is not up for debate. It's a fact. Even if your back of the envelope calculation there were correct, you have to ask yourself why there are more black people committing crimes. The reason isn't skin colour, but socio-economic factors which themselves are a result of a long history of oppression.
It’s an interesting description of the space of computation and I quite enjoyed this article even though, as many, I have issues with the way wolfram participates in science. Certainly what he is describing can be a way to view the universe but that is somewhat trivial. It is not surprising that you get something like the universe when you start considering universal computation. The fact that we can write down e.g. equations of motion means there is a Turing machine that computes it. The real test is whether this perspective is useful in the sense that we can predict properties of the universe with it. That test is still out and until it’s addressed there is no way to evaluate or criticize this model. It’s too flexible to be useful at this point.
Turing completeness isn’t necessarily an interesting thing to have in common. Many (very simple) models of computation are Turing complete but have vastly different properties. Take for example a cellular automata, a Turing machine, Wang tiles, (cyclic) Tag systems, Fractrans, Register machines, string rewriting systems. All of these are Turing complete. Yet they are miles apart in how they carry out computation. In order to understand and do what the brain is doing we have to figure out the brains model of computation. It will also be Turing complete but it will look very different than a Turing machine.
Why exactly shouldn’t scientists be politicians? It’s not like most of the people who are politicians have a formal education in political science. In an ideal world the ministers should come from exactly the background they are tasked with. This seems to fit in Sweden. I don’t know why you believe an economist or lawyer would make better decisions here.
I definitely think scientists should be politicians. Angela Merkel has a doctorate in quantum chemistry, and she is one of the most successful and respected politicians of our time.
But I do think it is dangerous to try to play the two roles at the same time. Science demands an honesty that few politicians can afford, which the current discussion about Anders Tegnell so well demonstrates.
> Science demands an honesty that few politicians can afford, which the current discussion about Anders Tegnell so well demonstrates.
This varies across cultures though. Scandinavian countries have a bit of a different approach to many thing, the most recent example I stumbled across was Danish zoos honestly and explicitly saying that they can't keep all the animals, that they will kill some, and that they will feed them to e.g. the lions. They even make it a public thing where somebody will chop up the animal and explain stuff.
Those things generate outcries - in other countries. We're very used to hide "how the meat is made" from the general population, but that's not necessarily the same in all countries. I find it very refreshing for politicians to not have "a public and a private opinion", and I wish it was more common.
Of course they can be politicians. But a politician is elected and sits on a mandate from the people, and this comes with a clear accountability. A scientist that is not a politician is a scientist. It is a big discussion in Sweden, who really is in charge and who is ultimately responsible for the failures.
