Hacker News new | comments | ask | show | jobs | submit login
Understanding the state of Artificial Intelligence academic research (2016) (ai-academy.com)
149 points by sytelus on Jan 13, 2018 | hide | past | web | favorite | 31 comments

I wouldn't trust any of the graphs or much of the statistical analysis. I was particularly intrigued by the last graph showing the geographic distribution of research. Thankfully, the author has published their methodology on github [1] so we can see that this data is in fact just papers published in the "Energy" category containing the words "artificial intelligence" somewhere in them. This explains why Iran and Malaysia rank above many European nations.

[1] https://github.com/gianlucahmd/Science_Direct_API_trends_ana...

Where in the code does it indicate this?

In the "Get most active nations" section the URL being queried is:

"http://api.elsevier.com/content/search/scopus?apiKey=" + apikey + "&query=Artificial+Intelligence+" + query + "&subj=ener" + "&start=" + str(start)

And "ener" is the id for the energy subject.

Thanks. That is really odd.

Maybe it is a typo: ENER (Energy) vs ENGI (Engineering)? Is there anyone with a Science Direct API key to check what the actual numbers look like?

The author is obsessed with fitting things with a logistic function and drawing wild conclusions even when it directly contradicts itself.

You can't have the rate of publication be a logistic function and the cumulative publication be a logistic function. The derivative(i.e., rate) of the logistic function should be 0 at both tails which isn't S-shaped anymore.

Yes, instead of slowing down, the curve could just continue growing straight, there is no inflection point in the graph. Also, area under logistic function is also more like a straight line: https://www.google.co.in/search?q=graph+log(1%2Be%5Ex)

If you want to understand the state of AI research, just pull out the origins of the papers from top conferences, that is it. Or for conference like ICLR, there is a rating attached for each paper, even easier for sorting top out influential ones. Anecdotally, US/UK/China are countries frequently popped into my mind, Iran is nowhere to be found.

Research doesn't work by aggregate, in fact most of the papers are useless. The top 1% influential ones are the ones that really matters.

Poor study IMO. I stopped when I read the

> “Computer Science”, “Mathematics” and “Engineering”, obviously the leaders in research, with respectively 963274, 491191 and 352870 publications.

Why "obvious" ?? I barely think so.

I think that the author is trying to point to the contains relation of CS, Maths and Engineering to AI. I think that Engineering is tenuous as AI in Aeronautical Engineering is obviously of interest as an application that would create extrinsic value. But of course AI for software would also create value!

One thing I don't understand about human intelligence:

Why do instances like this one publish their articles on Medium? As a reader, I am annoyed that Medium puts two dick bars on the page so it is hard to read. As a writer, I would not want to have my content controlled by a third party. And not have all these links to the third party on my content. What counterweights this? Are there any benefits of giving your content to Medium?

Eh... extremely much larger exposure? Those playing the game of posting on Medium and Linkedin are gaining the game of reachability. That's the point of publishing out content.

What is that large exposure? How does the article reach more readers because it is on Medium?


* It is linked (vs. a post in a personal blog post which would only be indexed by search engines) * There is a large community of users on Medium, and posts are easy to share. * Medium highlights popular posts.

    It is linked
Linked from where?

    Posts are easy to share
How can sharing be easier then ' read this: https://blog.ai-academy.com/3a79cac4c9c2 '?

    Medium highlights popular posts
But is it worth all the downsides? The post is on the HN frontpage, yet it only has 51 likes on Medium. How much additional readers does such a post get from Mediums free marketing?

>As a reader, I am annoyed that Medium puts two dick bars on the page so it is hard to read.

This, I remember blocking those bars a month ago with uBlockOrigin. It was very, very annoying the first time around now that you reminded me.

The benefits are that Medium is now more of a social network than a writing platform: you can follow people and give them digital applause (claps) if they’re doing some fine thought leadering.

The downside, as you’ve pointed out, is that content is becoming watered down and the UX of Medium is going downhill.

http://aiindex.org/ agrees and its AI Index peaked in 2015.

The AI Index does a reasonable job to understand the state of AI research [0].

> This report aggregates a diverse set of data, makes that data accessible, and includes discussion about what is provided and what is missing. Most importantly, the AI Index 2017 Report is a starting point for the conversation about rigorously measuring activity and progress in AI in the future.

[0] https://aiindex.org

Graphs? All I see is charts ;)

Early on the article made an interesting thought - once we get ai right, we won't need to solve any other problems...

> once we get ai right, we won't need to solve any other problems...

Only if it is superhuman AI, we already have 'human' level AI and as long as that isn't enough there is no reason to believe that 'ordinary' AI will solve anything that we can't solve.

Human BI is constrained by biological concerns. It seems plausible to expect AI to be able to scale a lot better.

> we already have 'human' level AI

And you're able to say that with a straight face?

He/she probably meant the AI that is perfectly capable of understanding human sarcasm.

ps: explanation: AI is at best at simple insect level intelligence and sarcasm is especially challenging for "AI assistants"

There are about 7 billion examples of this so yes.

Getting AI "right" isn't defining what "right" is, only saying it's clever enough to do the job of designing the even more powerful ai for us...

Once we get ai right, we have completely new set of problems.

I agree. Who knows the types of problems we will encounter once AI agents become sufficiently intelligent. We will face unprecedented challenges, even way before AI agents' intelligence gets close to human intelligence.

What does the number of publications have to do with anything? One paper can change an entire field, while another hundred could be just filler non-sense to get a PhD with.

From October, 2016.

Thanks, we've updated the headline.

Applications are open for YC Summer 2019

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact