Using an innocuous encoding of a word is a form of encryption. People who expect to be under surveillance agree on a set of code words to denote illegal things. Though hard, there are multiple ways to semi-automatically break such a linguistic encryption.
Imputation. [1] Remove a word from a sentence then try to predict it from its surrounding context. "when I get home tonight, i vape a ___ then space out". Assign predicted probabilities to imputed word ":leaf emoji:" ["marijuana cigarette", "electronic cigarette", "cigar"].
Active learning. Seed the algorithm with expert knowledge from law enforcement, drug users, and social workers, who know of the encryption keys.
Anomaly detection. Though perhaps easily-confused with other, innocuous usage, street slang is a distinct form of language with its own properties and patterns. Compared to common discourse, it is strange and random. This pattern could be measured.
Doing this rigorously, like building search engines for illegal drugs or human trafficking on the deep web, requires a lot of expert knowledge. [2] Maybe future deep learning can do this end-to-end on arbitrary domains? [3] Let's see.
> At some point the machine fact-checking relies on data input by humans.
But so does nearly every ML model? In the case of a spell checker it is using corpora made by humans. If the majority of humans start spelling words differently, then facts about the correct spellings change with them.
> Or does the machine interpret data directly from cameras on the street to determine, e.g. that suspect A shot victim B with weapon C?
If the military gets their way this will happen sooner than later. It is not technically infeasible to do activity detection from drone footage.
> Does it interpret a historical textbook and assess the veracity of its sources and claims?
Yes. Just like a journalist would when fact checking an article about WWII.
> Or does it build a time machine to go into the past and acquire raw data to verify claimed facts?
Raw data is both an oxymoron and a bad idea. Data is brought into existence by human-made measuring devices.
Resulting models will likely be released as open source, and the techniques published in open journals.
The big money is in productionizing and running it on bigger datasets.
Kaggle competitions are not a data-science-product-for-hire kind of thing, like some kind of logo design contest. It is a sport. Super GM chess competitions see smaller prize pools.
Mediocre people, including me, would compete for free (just to get access to interesting data like this). The really talented people are not driven by money, but by competition and fame.
Increasing the prize money is more of a marketing move and only attracts more low-to-mediocre people trying to get lucky in a lottery: It won't increase the quality of top 10 solutions. And no computer vision PhD/professor is going to drop everything he/she is working on for a small chance to win 400k.
1. Google has been an AI company from the very beginning (information retrieval).
2. Google is investing and doing generally useful applied AI, not an AGI moonshot.
3. Google's AI researchers are not 100% sure that AGI is just a few short years away.
4. Major source of income is advertisements. A lot of non-technical people work on this, allowing others to do more research and improve search.
5. Like said, AI is Google's DNA from the start. They are the biggest AI company in the world, and will die/be dethroned when they let AI research wither.
6. Avoid blanket humility, and lose hunger, innovation, dare. "At Hooli, nothing is ever impossible".
Surely Google follows their own guidelines: You can't find Google Search Engine results indexed on Google itself (or any other search engine with or without ads for that matter). Google Search is more of an application than a content site.
Otherwise, when Google finds itself breaking the "rules", they act:
- Google banned the page for Chrome for buying paid links
- Google banned an acquired company (Beatthatquote) for violating rules.
- Google penalized their Adwords FAQ pages for cloaking.
- Google reduced pagerank for Google Japan for buying links.
- Google removed Adwords support pages for keyword stuffing.
AI hype journalists will find something to write about, regardless of the industry making their research accessible to the wider public.
Markov-chain generators have been around for a while, and have been used to throw off spam detectors. This should not stop research, but instead grow more research into adversarial usage of machine learning models.
Beating the state-of-the-art with one-shot learning is not common. Transfer learning for NLP is also quite unchartered.
Also the technique is quite novel: This is not pre-trained nets on labeled data, it is an unsupervised generative model.
Future research directions are exciting: Unsupervised prediction of the next frame in a video, and then being able to one-shot learn a wide range of visual tasks.
It is possible to leak information, but then you are doing it wrong. Don't use only a single out of time test set to do parameter or model selection, keep an out of time holdout set.
But really, this is the bare basic of forecasting. It is somewhat annoying to have to regurgitate all of this: Like non-leaking forecasting is impossible somehow. It would be a better discussion if everyone just assumes proper forecasting practices. Instead people seem to assume I have no clue what I am doing, discarding my technique, because I did not mention removing duplicates, scaling, proper validation techniques, ... and a 100 other things, which are of no importance to the technique itself.
Is a mistake still a mistake when it is a profitable action?
If fitting to human irrationality increases generalization performance, then it does not matter if the "machines seem destined to repeat our mistakes", it is still a useful signal. If fitting to human irrationality decreases generalization performance, your algorithm is overfit to noise (and you have bigger fish to fry than human irrationality).
Overfitting to noise is perfectly avoidable, not pre-destined when part of your data is noisy (noisy data is the rule not the exception).
This implies the time machine can go only one way. Taking information from the future back in time is worth more than taking information from the current to the future. We do the latter all the time, we do not need a time machine for this, just patience.
Then look at backtesting. The evaluation data set is out of time, meaning the better than random performance is on unseen future data. The algorithm was implemented, not merely a suggestion.
> If it's so easy why not to identify directly which stocks to invest in?
It is not easy. Both to set-up, and the problem itself (you won't get a very high accuracy, but you will get much better than random guessing).
> If you can predict which investors will perform well, then just try to predict which stocks will perform well directly.
This won't work, because you don't have access to all the information that the stock pickers have access to, just their advice, and some features about the stock / the company the stock pickers work for.
It's not a case of exclusive or. Gather a list of a 100 or so human stock pickers. Train a model to predict the effect of following their suggestions (buy, hold, sell) to both identify which experts are most right, and which particular experts do well for particular kinds of stocks. This model will do better than random guessing, and quite likely, better than any individual stock picker.
> I didn’t have any money when I started the company to purchase a domain name so I built an algorithm that iterated phonetic domain names and printed out a list of what was available. My wife and I went through the list and “Kaggle” was the one we picked. It’s algorithmically generated.
> It’s a terrible name because most Americans pronounce it “kagel” [rhymes with “bagel”] which sounds like the pelvic floor exercises. Australians pronounce it “kaggel” [rhymes with “haggle”].
Imputation. [1] Remove a word from a sentence then try to predict it from its surrounding context. "when I get home tonight, i vape a ___ then space out". Assign predicted probabilities to imputed word ":leaf emoji:" ["marijuana cigarette", "electronic cigarette", "cigar"].
Active learning. Seed the algorithm with expert knowledge from law enforcement, drug users, and social workers, who know of the encryption keys.
Anomaly detection. Though perhaps easily-confused with other, innocuous usage, street slang is a distinct form of language with its own properties and patterns. Compared to common discourse, it is strange and random. This pattern could be measured.
Doing this rigorously, like building search engines for illegal drugs or human trafficking on the deep web, requires a lot of expert knowledge. [2] Maybe future deep learning can do this end-to-end on arbitrary domains? [3] Let's see.
[1] https://arxiv.org/abs/1312.3005 "One Billion Word Benchmark for Measuring Progress in Statistical Language Modeling"
[2] http://www.darpa.mil/program/memex
[3] https://universe.openai.com/envs#world_of_bits