mannigfaltig

"Computer vision" typically refers to computations that have an interpretative component (image analysis), i.e. programs that actually "see" something. Interpretation basically refers to the inference of the state of the "world" (or a part thereof) from an image. Image processing is often used in pre-processing for computer vision applications, but it also has many other purposes which do not need image analysis, such as in photo and video editing, as well as in generative applications, e.g. 3D engines and video games (which can in some sense be seen as the inverse of vision). Therefore, image processing is not exactly a subset of computer vision.

Secondly, do you think it is a good basis for building communities if everything is associated with monetary rewards? Sites like Reddit are exactly enjoyable because, most of the time, reward is only imaginary, just like in real-world gatherings of people: It is about having a good time rather than about an exchange of money.

Thirdly, how do you ensure that the things people post on Steem are their own work, especially if the data is hosted elsewhere, e.g. on YouTube. There will probably be massive, irreversible thefts of intellectual property (a.k.a. reposters, but with actual money involved). Conversely, a lot of valuable content comes actually from reposting and "karma theft" (which is what linking e.g. a blog article from someone else actually is).

Fourthly, how do you avoid fraud as for example often found on Kickstarter? Kickstarter actually has a review process and even they are not fully equipped for it. I am not seeing reliable moderation ever scale enough such that it can withstand the likely immense fraudulent interests due to the monetary incentives.

https://twitter.com/kalpenn/status/896013769410740224?ref_sr...

Like forgetting the original keybindings and switching to evil.

Are there works that expose this limitation of MLPs more formally?

>not everything good can be made More Neural.

Neural networks are universal function approximators, so you probably mean not everything good can be made with MLPs trained by gradient descent?

>It's the lack of [...] the ability to perform inferences over discrete spaces.

How would you judge the extent to which AlphaGo has learned to react to single discrete changes in the input. It seems that it learned very well to react very sharply to whether a single stone is placed at a strategically significant position.

Do counterfactuals have something to do with learning from negative examples and simulations? For example, if one shoots a ball and misses the goal to the right, one does not 'mindlessly' penalize the circuits that led to the exact motor decisions that were involved, but instead, one simulates alternative actions and uses e.g. (in this case linear) relationships between e.g. the angle of the foot or the wind speed and the shooting direction. The next time, one hence tries to aim slightly to the left.

Or are you referring to a much more fundamental level and my example might rather be a learning strategy that is more likely acquired by trial & error, reinforcement learning, meta learning ("learning how to learn") and/or via the shared concept space of language and culture?

Is it maybe related to e.g. prototype-based associative recall and a counterfactual is basically an alternative way of interpreting the data? "What error signal would I get, if I had interpreted X as Y?"

Or does it come from the Bayesian approach where you marginalize out all hypotheses, including the factual one that corresponds to the state of the world, but also all counterfactual hypotheses. So, including counterfactuals means going beyond the maximum likelihood point estimate e.g. by communicating confidence intervals or even entire distributions from neurons to neurons or neuron populations to other neuron populations?

I am not Einstein but it is probably not a bad idea to have a whole lot of interesting things around you all the time. Sort of like a cache or the way proteins are synthesized, like swimming in a nutritious soup. "Oh, look here is a piece that fits!".

One basically decreases the chance of forgetting interesting and useful information. "Out of sight, out of mind."

https://news.ycombinator.com/item?id=13566917

https://openreview.net/pdf?id=rJv6ZgHYg

https://www.reddit.com/r/MachineLearning/comments/6ailoh/r_u...

https://www.reddit.com/r/MachineLearning/comments/5kfs23/r_u...

https://www.reddit.com/r/MachineLearning/comments/5cw3lr/r_1...

Actually most organized religions are criminally inefficient at this. As an example the Catholic Church in Germany earns about 6 billion euros per year, completely tax-exempt. 50% of that actually comes from tax money. Now, if you look at social institutions which claim to be funded by the church, it is usually only 5-40% (rather the lower end) that the Catholic church contributes and the rest comes from insurance companies, contributions by the patients/care receivers and, again, from taxes. The church does not contribute to religious education at all and there also is no freedom of strike in the German Catholic church and many similarly miserable exception clauses.

I'd say for people with interest in ML and DL, the main insight from PGMs is intuitions about probability theory, especially conditional probabilities and the concept of (un)conditional independence. State of the art in which PGMs would be relevant is currently mostly brute-force learning via stochastic gradient descent (with momentum) in deep directed models/function approximators, but function approximation likely does not solve all problems as adversarial examples show. There are certainly very important applications of PGMs in all kinds of domains as listed in the link above (perhaps I would have added FastSLAM [1]), and there are also variational autoencoders/Helmholz machines and (Deep) Boltzmann Machines, which are especially interesting for research perspective.

[1] http://robots.stanford.edu/papers/montemerlo.fastslam-tr.pdf

To spell it out: backdoors are of course a helpful tool for a security agency, but once the key for the backdoors falls into the wrong hands the potential for misuse is monstrous because, well, the backdoors are in all kinds of potentially critical systems systems, e.g. even in the cellphones of high ranking officials (or in cellphones of their families) etc.

To get the same intel by different means than a backdoor is of course much more expensive. So one needs to find the sweet spot, where the utility as a function of cost and security is maximized. That sweet spot likely does not involve backdoors.

Here is another paper demonstrating very good results with just 6 bit gradients: https://arxiv.org/abs/1606.06160

http://elephantinthebrain.com/outline.html