This seems to be an unfair comparison to me. The penalty for not paying the fare is not meant to somehow recuperate the lost cost - instead, the penalty is there to incentivize people to pay the regular fare, rather than skipping it. The real question to ask is: if the penalty were reduced, how much more money would be lost due to the increased number of people not paying the subway fare? I don't see a good way to answer this question, but it seems to be the right comparison to me.
I don't know if I'd say it's irrelevant if it's not used by the current industry. It's always helpful to understand what is possible and how to achieve it. And beyond actual application, formalizations of fairness notions and thinking about their implications for computing gives us another lens to think about fairness more broadly (as in, these kinds of technical discussions can give specificity and clarity to discussions of fairness even beyond machine learning and computing).
To your second point, it's very difficult to attribute a single intention to large companies. There are certainly employees and researchers at large tech companies that work on fairness and privacy preservation because they think it's important and want to make their companies more responsible (I know some!). It's also certainly true that exploring this type of research looks good for the company, and some other employees are surely aware of that and promoting it for those reasons.
That's a really interesting point - it does seem like most people immediately jump to skynet or our robot overlords whenever the topic of AI comes up. I don't have any really solid suggestions for what a good mental framework for thinking about AI would be, but I think giving an alternative to these unrealistic versions would be super helpful.
I agree with the sentiment that the discussion about AI and Machine Learning should not be entirely driven by industry.
At the same time, this article seems to be a bit down on AI itself, and part of their message is that AI doesn't provide "relevant and competent" solutions to problems. It also sounds like they are writing off real work (both in industry and academia) focusing on real ethical concerns with AI and what can be done to address them (e.g., the FAT* conference, a growing number of sessions on fairness, privacy, and other related topics at NeurIPS and ICML, etc).
I think the most important issue is educating the general public about AI and giving them familiarity with what types of things are automatable, what things can be learned from their data, where and how AI is being used in the real world currently, etc. A big part of this is to have the mainstream media be a bit more self-guided.
A final thought: one of the suggestions from the Reuter's article is that we should hear more from scientists and activists in the media. This seems a bit troubling to me, since in ML and AI research there are very strong ties between academia and industry (and often people move fairly freely between the two). I'm not sure we would hear a significantly different narrative if we talked to researchers in academia...
Why would it? Eigen-faces tend to not even look like faces (which makes sense, given that the goal is to represent a collection of real faces as linear combinations of the orthogonal eigen-faces). One possibility would be to generate random faces as random linear combinations of the eigen-face basis, but this type of attack could be thwarted by rate limiting.
A real "master face" would be a face that looks like many people, and it seems like you could try to obtain faces like that by playing a two-player game between a recognizer and a face generator (as is done when training GANs).
I think it depends on what you mean. I don't think that just the number of bytes of data should be used to measure the size of your dataset. For example, if I have 1TB of all 1's, this is a lot of data, but not very interesting.
I think a more nuanced notion of size is the /information content/ of the datasets. I haven't thought about it carefully, but I'm sure you can quantify this more explicitly in terms of information theory or other non-bit-based complexity measures.
From this point of view, literature may not be large (in terms of the number of bytes), but in terms of information content, it is incredibly dense. A large portion of human knowledge is written somewhere. From this point of view, it is in fact big data.
I read further down the page - they use a quasi newton method to find inputs that minimize the (squared?) distance between the desired output and the output of the guessed input.
I guess this can only really be expected to work well when there is a input that perfectly matches the output in a convex neighborhood of the starting point. Pretty cool!
Looks like they also have support for a few other types of data, but it's not described in the documentation. Apparently there is support for the backwards computations on strings, nested objects, and you can declare inverses for functions. There's also something mysterious called "cascading approaches".
I'm super interested to know more about how this works!
I'm curious about how the backwards computations work - how do you detect when the backwards computation can be easily calculated? For example, if we hash some input string and then change the value of the hash, it should be very difficult to do the backwards computation.
Does it literally do gradient descent on the input to try to match the output (as is suggested by the backpropagation terminology?) Can it handle discrete valued outputs?
I agree that there are important differences between human and computer languages and that we should be mindful about the gap, especially when we try to automate cognitive tasks like interpreting contracts (the DAO!) , assessing beauty, deciding who lives and who dies, etc. However, I don't think that the differences between human and computer languages justify very many claims about AI, how AI will develop, or what AI will be capable of. After all, the programming language we use to build an intelligence might have very little to do with its inner workings or internal representations of concepts.
I very strongly disagree with the following quotation: "If we had good enough theories of human semantics we could program such theories into the computer and the computer would then understand like a brain after all. But we don't, in most areas, so we mostly can't program computers to emulate humans." I don't think we should try to automate tasks like understanding language by writing computer programs that work the way we imagine ourselves working. I think a more fruitful approach is to design general /learning/ algorithms that allow the computer to figure out what language is and how to use it by observing how it is used in the wild (i.e., by emulating other language users, maybe with some trial and error).
In general it is a misconception that real data always follows a normal distribution. It is true that if you sum many /independent/ random quantities, then the result is approximately normal (e.g., the central limit theorem and generalizations). But real data tends not to be independent. Many real world quantities follow extremely skewed distributions. E.g, Zipf's law, Korcak's law, Pareto's laws.
For a concrete example, if you look at the distribution of the number of friends users have on social networks, you might expect that 95% of people have the mean number of friends +/- a few standard deviations (since this would be the case for a normal distribution). It would be virtually impossible for someone with a number of friends that is say thousands of standard deviations away to exist, yet there will be many such users in social networks (celebrities, bot networks, etc). In reality, the empirical distribution in this case follows an extremely skewed distribution.
I agree. For both the MAD and STD, we are trying to reduce information about the "spread" of a distribution to a single number. Any such reduction must lose information, so you should pick whichever quantity is suitable for your needs.
E.g., in the article they mention that the Pareto distribution has finite MAD but infinite variance. This is meant to be an argument against using the STD, but actually the infinite variance tells us something really important that the MAD does not: the law of large numbers /does not apply/ for the Pareto distribution!
I think the real message should be to avoid blindly applying techniques and tools (especially formal ones) without thinking about why or what they capture.