zackchase

That's not quite right. Salary bands tend to be guidelines, not constraints. And in ML there are plenty of people getting paid far out of band.

I assure you, people have pointed it out...

Thanks for sharing! Happy to discuss here if anyone wants to engage in some (pre)debate. Also happy to butt out :).

These are excellent questions. I often ask Google researchers, "How Deep is the Mind?" They don't seem to know. You see, it all comes down to democratizing the blockchain, and the only way to do that is with AI in each and everyone's hands. Together, we can make a more open AI, for everyone! And for you too! The only trust required for CryptoMinds is trust in proof of work, am I right?

Thanks for the interest in JnJ AGI Research! As inter-global head of research I'm happy to answer any questions the HN community may have!

Flag-planting is relevant because the practice of flag-planting (and other abuses of the pre-print) is what has prompted some of this over-reaction.

Author here: Grisha Perelman's proof of the Poincaré conjecture has never been (by him, to my knowledge) submitted to or published in any journal. He decided, as is his right, that he could not care less about the professional community of publishing mathematicians or their protocols. Does not invalidate his achievement.

I should note, since I am not and do not expect to be the level of mathematician that Perelman is, I have not actually read his proof. So I defer to other superior mathematicians for this assessment and come by it as hearsay. :)

It's problematic that the "experts in AGI" are generally self-designated, and have mostly produced little that is concrete, either in the form of theory or engineering. We should be concerned about long-term outcomes of technological progress, but I'm not convinced the current conversation centering on the term "AGI" is productive.

While I don't take exception with the word machine learning (it's reasonably well defined), I agree with most of the points here. Some are grossly exaggerated (e.g. needing petabytes of data to do machine learning). In my research on medical time series, we get strong results with a few thousand examples (megabytes). Same for natural language datasets which are modestly sized. But yes, absolutely agree that function approximation for supervised learning and some kind of artificial consciousness as portrayed in the media are very far apart.

As someone with an economics background (BA), and medical research career (concurrent with ML), I can say (subjectively), that this feels different.

I generally agree with this. I used machine learning almost exclusively in the article. Yet "AI misinformation" seems more appropriate in the title since the misinformation, generally, is in reference to "artificial intelligence".

Some strident Singularitarians joined the comment thread.

Some strident Singularitarians have joined the comment thread.

The performance of multilayer perceptrons and linear models is significant. The paper additionally allows the MLP and linear models to "cheat" by giving them heavily hand-engineered features incorporating extensive prior knowledge. Even with this benefit, the LSTM outperforms the baselines, but the improvement is considerably smaller.

The two statements here may be factually true. 1) You had a hard time. 2) Your female classmates, all else being equal, did better. But there's something unsettling about the implication. Namely, by putting these two assertions together, you suggest that they have something to do with each other, which, given the appallingly low representation of women in all areas of computer science, is obviously not true.

These are strong arguments. Unfortunately, while sometime interpretability addresses a fundamental (not frivolous concern), it's also true that often interpretability is sometimes contrived to serve some irrational bias or political motive.

I think this is expressed in the paper's discussion section:

"We should be careful when giving up predictive power, that the desire for transparency is justified and isn’t simply a concession to institutional biases against new methods."

I have good reason to believe the paper's author agrees with this argument.

Hi smackfu. So one wonderful virtue of this technology is that it can handle messes quite gracefully. Part of the fun is to see that it can learn to stay on topic - to decide it's talking about a raspberry beer and continue to talk about rapberries. As far as fruit vegetable beer being a catch-all category, that's only within fruit beers. They're certainly distinct from IPAs, stouts, porters and lagers. Interstingly, for classification, far the hardest categories to disambiguate are Porters and Stouts (probably because stouts really are "stout porters").

Our goal was to pick a few categories would allow us to evaluate the capabilities of the model. Thus having some that are crystal clear and others heterogeneous was appropriate. Regarding American Adjunct Lagers being low rated, yes the lagers are frequently described as "piss", "watery", and "urine".

Hi larsga. Thanks for your interest. To begin, yes the concatenated input RNN is much better. A Markov chain could not do this feat (or make comprehensible text at the character level generally). Occasionally the reviews contradict themselves, but they're actually remarkably consistent regarding the conditioned upon attributes. I'd suggest you pay attention to the "temperature". This is a parameter that determines how stochastic the generation is. With low temperatures the reviews are less varied but "make more sense". With higher temperature they are more entropic.

When you tell it to make an IPA review it stays on topic and talks about hoppy flavor. About stouts is consistently calls them black, with hints of chocolate (not to mention using the word "stout").

Here's an example of a review I just generated for "Fruit/Vegetable beer":

"This brew pours a very clear golden color. The finger head is pretty small and fizzy and has a slightly pink color. The smell is really nice. The taste is fruity and sweet, but not overwhelming. The flavor is a little weak and is a bit sweeter than most beers but still very nice. I could drink this all day, but I would probably prefer the fruit beer to be a bit more pronounced. This beer is actually quite smooth and inviting. It has a strong taste of raspberries but is complimented by a nice tartness that comes through as well. The mouthfeel is smooth and creamy with a dry finish. This is a very drinkable beer and I could see myself enjoying to try this one again."

Clearly the RNN learns to form words like "fruity" and "sweet" and "raspberries" to describe a fruit beer. It also says the flavor is a little "weak" and in the next sentence says it would prefer for the taste to be "more pronounced". Keep in mind, this neural network was given no a priori notion of words. A Markov chain cannot produce even remotely similar conditional text at the character level.

For proof that it learns to differentiate the different types of beer, we demonstrate in the paper that the model can be run as a classifier and classify the type of beer form the review with 90% accuracy (on previously unseen test data). This is almost comparable to state of the art logistic regression tf-idf ngram model, despite the fact that we haven't even tuned the model especially carefully to be a classifier (with regularization or hyper-parameter search , for instance.

Here's another example (for an IPA):

"This is a fine IPA for sure, but not a beer I would love to drink a lot of. This is one of the better IPA's I have ever had. I can see why the beer is unlike any IPA should be and the best beer I've ever had. I could drink this all night, but it is very drinkable. I could easily drink a few of these without a problem. I don't know what the malt base but it is so faint and it is pretty tasty. I can see why the beer is a great hop bomb and the flavors are both subtle and superb. It's a great balance and can be a good supply of the style. I like it, but the hops are a bit off the more I drink them, but the hops are very pronounced. The finish is a little bitter, and the hop flavors are great."

I'm not sure what else to say, if you don't believe that the net has learned to distinguish an IPA. Of course, it does contradict itself on sentiment. But with lower temperatures even this is not so common. It can also be addressed by setting extreme star ratings. (we can actually put in a 0 star rating, or as high as a 10 star rating to induce a review of more extreme sentiment).

A comedic point that must be made here is that the source material (the reviews from BeerAdvocate) are themselves absurd, and occasionally contradictory. They English they contain is ungrammatical, ridiculous, and frequently misspelled. Nevertheless it's a fascinating dataset on account of how well-annotated and dense it is (the 190-core has over 250k reviews).

Hi Ian,

First off, thanks for your interest in our work! The tech is a homegrown recurrent neural network (deepx) which is available via pip. https://pypi.python.org/pypi/deepx/ We use theano for compilation to GPU and an original neural network architecture (a concatenated input generative model) to preserve the signal of auxiliary inputs (like star rating and beer category) across long intervals.

To run the model in reverse, we infer the probability of a category via the likelihood of a review. Because the prior over the categories is uniform (balanced dataset) and because the normalization term (marginal probability of a review) doesn't depend on the category, we can exploit the fact that the probability of a beer category given a review is proportional to the probability of a review given a category. In this way we're able to make a text classifier that takes into account word order.

Regarding Google Analytics, we just noticed this ourselves and are fixing it presently. Concerning color, that's a great point and I apologize for the oversight. We certainly bear no malice towards those with protanopia. In future work (and if we have subsequent versions of this paper) we'll do our best to make the lines dotted, dashed, etc to make them more readable to those unable to distinguish reg-green. Generally, I agree with you that a black and white printing should retain all readability.

Thanks for your interest and helpful observations.

Cheers,

Zack

Hi, thanks for posting our work! The system generates reviews at the character level based on both category and star rating.

I don't know anyone in computer science that writes academic documents in Microsoft Word. The standard is LaTeX. It seems the author may be well behind. All markup (in any markup language) is simply text (as opposed to WSIWYG), which is the main thing the author seems to find alluring. However, .txt files by themselves give you no way to manage special characters, references, footnotes, or any other typesetting functionality.

This post scares me in the way that most bitcoin journalism does. There is no shortage of press talking about the start ups themselves (as opposed to the state of the capital markets). For many people, the state of the markets really does matter. If if you're looking to raise capital, it helps to know that these are historically favorable times. If you're looking to invest, it's important to consider whether things may be overheated. This article advocates silence (I'm tired of hearing about [insert topic]). This head in the sand talk is alarming in the same way that it's scary to hear people say "I'm tired of people saying that bitcoin is a pump and dump scam". Generally, it's a bad sign when a viewpoint is suppressed. Having lived through the dot-com bust (in which my father's "bubble" talk fell upon deaf ears) and the subsequent housing bubble in which people were similarly dismissive of "negative" voices, I'm wary of this rhetorical style.

Broadly speaking, I think there is a fundamental misunderstanding in much of this thread of what "critical thinking" means.

On the topic of negative comments, this article prompts me to leave what the author assuredly would consider one. Frankly, if our society fails in education it's that students make it through without acquiring critical thinking skills.

The idea that we are critical because "it’s so much easier to challenge other people’s thinking than our own" is hard to take. If anything, this speaks to a failure to be self-critical more than a surplus of outward negativity.

More generally, I'm frustrated by what I see as an increasing tendency of people to parse the world into "positive" and "negative". This by itself speaks to a lack of truly critical thinking.

Further, I believe that the idea that critically taking apart other ideas is somehow a fundamentally distinct activity from developing one's own ideas is incorrect. In mathematics you develop ideas precisely by searching for what might be unsatisfying in the arguments and theories of others. And most successful start-ups seem to have some roots in looking at how what others are doing is sub-optimal.

I think the intuition behind max-pooling is that it says "something is in this local region", without overly specifically saying where it is. Intuitively, a human may detects an edge or intense bright light, but not care so much about precisely where it is in the field of vision. After many successive layers of max-pooling, however (if the pools are not over-lapping), even somewhat course information about locality is lost.

I believe Hinton objects to this gross loss of spatial information for two reasons: 1) Humans don't lose so much spatial information, and Hinton would like his models to ultimately capture a neurologically plausible computation. 2) It may not be necessary for object detection (Imagenet), but it would likely be important for more sophisticated tasks.

These arguments were introduced by Szegedy et. al. earlier this year in this paper: http://cs.nyu.edu/~zaremba/docs/understanding.pdf. Geoff Hinton addressed this matter in his Reddit AMA last month.

The results are not specific to neural networks (similar techniques could be used to fool logistic regression). The problem is that ultimately a trained network relies heavily on certain activation pathways which can be precisely targeted (given full knowledge of the network) to fool networks into misclassification on data points which might to a human seem imperceptibly changed from those which are correctly classified. It is important to understand adversarial cases, but unreasonable to get carried away with sweeping pronouncements about what this does or doesn't about all neural networks, let alone intelligence generally, or the entire enterprise of AI research, as seems to happen after a splashy headline.

I took a stab at trying to interpret the topics output by this run of LDA. Green is one the clearest: generally convolutional deep nets, image classification, empirical work.

Brown seems to have picked up on linear algebra. "Vector", "matrix", "tensor" and "decomposition" all get consistently labeled brown, as do "eigenvalues", "orthogonal" and "sparse".

The rest are not as useful. Black almost always has "number", "set", "tree" and "random", but little else. Purple at times seems to signify topic modeling, but also contains "neural" and "feedforward". Blue seems to be the stats topic, containing "Bayes", "regression", "gaussian", and markov processes. But it also contains random words like "university" and "international".

Overall, very interesting. I wonder if these topics would be even better defined with a higher setting of k.