lars

I basically share your sentiment. However, Greg Yangs work seems to have produced something of direct practical benefit for training large neural nets, based on the NTK literature. µ-Parametrization is apparently very useful in real world practice.

Yes, it was. From their site: "It has been trained on a selection of English language and source code from publicly available sources, including code in public repositories on GitHub."

Note: Ian Linkletter, the one who had spent 100k in april, is not the guy being backed by EFF.

Linkletter has a gofundme here: https://ca.gofundme.com/f/stand-against-proctorio

I've donated, and would recommend others do the same.

I think they would in any case. My impression is that data is siloed internally at Google, and that data sharing between departments would be way more complex than just setting up some (possibly redundant) logging.

Idea: We could stop using cryptocurrencies.

Any downsides to that?

Daniel Lemire's blog is brilliant: https://lemire.me/blog/

Similar to how "god of the gaps" arguments work, there's a "god of complexity" phenomenon in tech, where people will project infinite possibilities onto technology they don't understand.

Blockchain is a brilliant solution to an extremely rare problem. It lets you do distributed consensus among untrusted nodes in a setting where contributing compute power is economically incentivized (e.g. where there's mining). That applies to cryptocurrencies, but basically nothing else.

It's amazing how much bullshit has been pitched under the flag of "blockchain". Millions of dollars have been poured into projects that are technically unsound or that had no use for a blockchain to begin with. This has been going on for years.

Did you respond to right comment? You seem to be arguing against something that I'm not arguing for.

Well, this is the discussion I'd like to see.

You have people born in the 90s hospitalized in Norway now. Many experience long term physical and cognitive effects. And as long as we're assuming everyone goes unvaccinated, the chance of contracting COVID goes towards 1 over time, while the R-number explodes as we open up society to avoid economic ruin. Not a good scenario IMO.

As I said, the relevant comparison in the real world is versus waiting for another vaccine, where it the choice is much less obvious, and dependent on the disease level in the population.

I'm happy this research has been done so quickly, but am somewhat annoyed that the public conversation is so focused on whether or not the vaccine can possibly cause any problems, and not at all focused on the risk management question of what the optimal choice here is.

The chance of getting a blood clot seems so low that it may for all I know be equivalent to the risk of dying in the streets while jaywalking to get to the vaccination center. Compared to the risk of going unvaccinated, it's an obvious choice.

That's not the choice we have though. The correct comparison seems to be the expected risk of taking the vaccine vs expected risk of waiting for a dose of another vaccine. Calculating this on an individual level should be straight forward given the risk of contracting COVID in a given area. Ideally the calculation would also take into account the effect that vaccination has on the virus reproduction number. That's harder to model, but approximate models like this exist.

If you actually put numbers to it and do this computation, you'll get an answer, or at least a distribution of outcomes, where it's straight forward to see what choice is optimal. Then public discussion can be around what the parameters of this model should be, and we'll stand a chance of making the right choice.

This is a potentially very consequential optimization problem, and the public discussion about it is as if no one understands that that is what it is.

Lars from CatchJS here.

Anyone can go on archive.org and verify that they changed their color scheme and header to ones very similar to CatchJS in Feb 2020, after CatchJS had used that look for 2 years. Thank you for also pointing out the pricing page, where the same is true.

I've now received a cease and desist from their attorney stating that the looks are too similar. It's nice that they've paid for an attorney to make my point for me, but they seem to have mixed up who did what here.

A middle ground could be nice: An IDE extension that notifies you when something you're writing will conflict in the future, should you and your coworker both commit and push what you've typed out. It would allow you to sort that out immediately, or at least plan ahead, rather than being surprised by a large merge conflict n days down the road.

It really is a tiny subset of the C++ standard library, but I'm happy to see they're continuing to expand it: https://nvidia.github.io/libcudacxx/api.html

That somebody was George Clooney: https://www.youtube.com/watch?v=Qa_HnP9TP5s

If you talk to Norwegians about Norwegians, you'll discover that we're very good at patting ourselves on the back. You might say that we're one of the best countries in the world when it comes to patting ourselves on the back.

The problem isn't having an MBA though is it, it's (allegedly) not being competent enough in engineering. All the CEOs you mentioned have engineering degrees, while Intel's CEO does not.

That's an excellent example to show why statistical NLP outperforms discrete parsing using logical rules. That A1 and A2 means the same is something that is clear to us from the context. This is something a continuous vector space model can capture and a discrete rules based model cannot.

The difficulty of the problem is that of a 50-way classification. If the only goal was to minimize WER, a simple post-processing step choosing the nearest sentence in the training set could easily bring the WER down further. They've chosen to do it the way they did it presumably to show that it can be done that way, and I don't fault them for it.

They claim that word-by-word decoding implies that the network has learned to identify words. This may well be true, but it isn't possible to claim that from their result. For example, let's say you average all electrode samples over the relevant timespan, transform that representation with a FFW neural net, and feed that into the an RNN decoder. It would still predict word-by-word, on a representation that necessarily does not distinguish between words (because the time dimension has been averaged over). Such a model can still output words in the right order, just from the statistics of the training sentences being baked into the decoder RNN.

This is cool. For those who are not super familiar with language processing, I think it's good to point out the limitations of what's been done here though. They mention that professional speech transcription has word error rate around 5%, and that their method gets a WER of 3%. Sure, but the big distinction is that speech transcription must operate on an infinite number of sentences, even sentences that have never been said before. This method only has to distinguish between 30-50 sentences, and the same sentences must exist at least twice in the training set and once in the test set. Decoding word-by-word is really a roundabout way of doing a 50-way classification here.

It's an invasive technique, so they need electrodes on a human cortex. This means data collection is costly, so their operating in very low data regime compared to most other seq2seq applications. It seems theoretically possible that this could operate on Google translate level accuracy if the sentence dataset was terrabyte sized rather than kilobyte sized. That dataset size seems very unlikely to be collected any time soon, so we'll need massive leaps in data efficiency in machine learning for something like this to reach that level. They explore transfer learning for this, which is nice to see. Subject-independent modelling is almost certainly a requirement to achieve significant leaps in accuracy for methods like this.

They only had 36 samples, and a 3sd event occurs once in ~300 cases: this hints to the fact that that data point might have been erroneous

Or it hints that the distribution of learning rates is not gaussian. When there's an "n-sigma" event, it's usually much more likely that the model is wrong than that the event is that rare.

I suspect "stickiness" is an important part of the equation for Reddit as well. With an app they can send notifications, and thereby program the habits of their users.

I am very aware of these, and I like methods like CLRP. It's great to have these tools to debug neural networks.

However, for many that argue for explainable AI, CLRP falls way short of what they want. In particular, the symbolic AI crowd would scoff at it. This is the crux of the issue in my eyes, that the symbolic AI crowd has taken "explainability" as a way to justify methods that don't work.

I have no issue with methods that allow greater understanding of neural net internals, that's essentially what all neural net researchers spend all their time on (and it's the path towards better performing methods).

an elucidation of the inputs and the procedure by which an output is generated does not an explanation of the system make.

That was exactly my point. Are we talking past each other?

My point here, if you wish to engage with it, is that when we evaluate trust in an AI system, we care about how good it is. And it is the case that quality is very often anti-correlated with explainability.

Suppose your life depends on winning a game of Go. Would you want AlphaZero on your side, or a Go engine that would present you with a list of the options it evaluated, so you can verify its decision? Of course, AlphaZero would beat the latter program every time.

If this desire for explainability is taken seriously, the result is that we'll end up picking methods that perform worse, and this will cause real world harm as AI becomes a larger part of life critical systems.

It may well be that the human visual system is better for most tasks, but then that is what matters, which system is better for the task. The presence of an explanation doesn't matter. Neither system comes with an explanation.

On the other hand, if you place Magnus Carlsen against AlphaZero in a game of chess, I will bet on AlphaZero. If however you reduce the complexity of AlphaZero down to a level that it can produce an explanation I can understand, I would instead bet on Magnus Carlsen.

Of course we should care about the quality of AI systems, but chasing a human understandable explanation is just the wrong way to go about it, since it in many cases necessarily limits quality of the decisions.

Ok. I can give you a 100 million floats, and tell you that if you multiply, add and threshold them in a certain order against a set of pixels, that tells you if the picture is of a cat or a dog.

That is an exact reproducable procedure that tells you how the neural network works. But you are a human being, and have a short term memory capacity of about 7 items. And 100 million parameters is too much for a human to really understand.

The point is that there are strong reasons to believe that no procedure for classifying cat vs dog is small enough that humans can wrap their heads around it. And why is this a problem? The human vision system is exactly the same, complete black box, yet we rely on it every day.

If I show you a picture and ask you if its a dog or a cat, you can answer the question, but you can't tell me why. You can make up a story to justify why it's a dog or a cat, but if someone tries to implement the procedure you give, it won't work. If it did, we could do computer vision by implementing your justification.

You are a black box AI. I can nevertheless trust you to classify dogs vs cats.

He's not wrong though. We know precisely how any given neural network works, it's just not understandable to us.

The point is though, the functions we're trying to learn, like image classifiers, have no responsibility to us to be understandable. In fact, the brightest minds of several generations have worked hard on trying to write down rules to do image classification, and they never came up with anything that works.

There is a huge space of functions that are beyond what a human can understand, where we can't write down rules to express the function. This is precisely where we need to use machine learning to find the functions. Lack of explainability is not a bug, the entire point of a neural nets is to find functions that are beyond what we can understand.

While the sources you mention are all good, I don't think any of them actually have much to say about what the link is about, which is bayesian deep learning.

Friendly reminder that churn rate isn't a meaningful metric, because churn probability is not constant throughout a customer lifetime: https://medium.com/swlh/youre-all-calculating-churn-rates-wr...

The churn rate will almost certainly go down on its own as your user base gets older, and consist more of long term customers. What you're usually looking for is to compute average customer lifetime, which is better done without assuming constant churn probability, see the linked post.

The react license had "we won't sue you for violating our patents, except if you sue us first.", which entails "if you use this, you can't sue us for violating your patents, because by using this you guarantee that we have material for a countersuit." People felt that this asymmetrically benefited Facebook. You want Facebook to be on the other side of such an agreement. If a license was made that came with an agreement that neither party can sue anyone in an offensive patent suit, it would limit the destructive power of the big players that hold all the patents.

If BatchNorm, ConvNets, LSTMs, Dropout, ReLUs or any such essential neural net component came with such a clause, I could see the big players accepting the license, especially since it aligns with their publicly stated claim that they only file patents for defensive purposes.