More AI Progress

[lpetrich]

Which machine learning algorithm should I use? - Subconscious Musings It depends on what one wants to do. Look for patterns? (unsupervised) Predict some outputs from some inputs? (supervised)

Dimensional reduction can be a good first step, because if much of the variation is in a few directions, one can project the data onto those directions and get a less bulky dataset for other methods.

Data can be either numeric or categorical, being a member of some category. For example, a pet can be a dog, a cat, a rodent, a bird, a reptile, a fish, a tarantula, ...


I'll write out that article's flowchart as pseudocode.

Dimension Reduction?
Yes:
. Unsupervised Learning: Dimension Reduction
No:
. (do nothing)
End

Have Responses?
Yes:
. Predicting Numeric?
. Yes:
. . Supervised Learning: Regression
. No:
. . Supervised Learning: Classification
. End
No:
. Unsupervised Learning: Clustering
End

Here are the subcategories.

Unsupervised Learning: Dimension Reduction:
Topic Modeling?
Yes:
. Probabilistic?
. Yes:
. . Latent Dirichlet Analysis
. No:
. . Singular Value Decomposition
. End
No:
. Principal Component Analysis
End

Unsupervised Learning: Clustering:
Hierarchical?
Yes:
. Hierarchical
No:
. Need to Specify k?
. Yes:
. . Categorical Variables?
. . Yes:
. . . k-modes
. . No:
. . . Prefer Probability?
. . . Yes:
. . . . Gaussian Mixture Model
. . . No:
. . . . k-means
. . End
. No:
. . DBScan
. End
End

Supervised Learning: Regression:
Speed or Accuracy?
Speed:
. Decision Tree, Linear Regression
Accuracy:
. Random Forest, Neural Network, Gradient Boosting Tree
End

Supervised Learning: Classification:
Speed or Accuracy?
Speed:
. Explainable?
. Yes:
. . Decision Tree, Logistic Regression
. No:
. . Data Is Too Large?
. . Yes:
. . . Naive Bayes
. . No:
. . . Linear SVM, Naive Bayes
. End
Accuracy:
. Kernel SVM, Random Forest, Neural Network, Gradient Boosting Tree
End


As one can see, this is a well-developed field, though one being extended with "deep learning" algorithms for huge datasets like big collections of image files.

[lpetrich]

EnhanceNet Single Image Super-Resolution Through Automated Texture Synthesis noting [1612.07919] EnhanceNet: Single Image Super-Resolution Through Automated Texture Synthesis (paper at arxiv)

It's filling in details of a low-resolution image file to make a plausible-looking high-resolution one. A naive way of doing that is to do a pixel-by-pixel fit, but that produces images that are too blurred-looking. These researchers used a fit that tries to approximate what is most apparent to human perception, like sharp lines. So the software tries to find out where such lines might be.

When two competing neural networks result in photorealistic face From a database of pictures of celebrities, these NN's became able to make fake ones, and even able to morph one to another.

So one of the two has a generative function, like rendering images or trying to solve a problem, but the other is in adversarial mode, challenging the first's results.

...
"The key idea is to grow both the generator and discriminator progressively, starting from low-resolution images, and add new layers that deal with higher resolution details as the training progresses. This greatly stabilizes the training and allows us to produce images of unprecedented quality."

Multi-racial facial recognition system provides more accurate results, study says -- not clear in that article how they include racial variation. Do they include it as a training-set output? But doing so produces improved results, including improved results for simulated aging.

Researchers trained neural networks to be fashion designers (sort of)

I've found this interesting application of artificial vision.

Artificial intelligence algorithm can determine a neighborhood's political leanings by its cars

Some AI researchers trained some AI software on pictures of a large variety of cars, and the software was then used to recognized cars in pictures of streets. They found a noticeable correlation between what cars were present and demographic variables like race, education, and income -- and also political leanings.

Using deep learning and Google Street View to estimate the demographic makeup of neighborhoods across the United States From the abstract: "For instance, if the number of sedans encountered during a drive through a city is higher than the number of pickup trucks, the city is likely to vote for a Democrat during the next presidential election (88% chance); otherwise, it is likely to vote Republican (82%)." A sedan is the ordinary sort of car.

They conclude "Our results suggest that automated systems for monitoring demographics may effectively complement labor-intensive approaches, with the potential to measure demographics with fine spatial resolution, in close to real time." But one will still need to have some basis of comparison.

[lpetrich]

Artificial Intelligence Finds Trump the Biggest Fable-Teller by AI researcher Łukasz Kuncewicz

Once more, we all live in the age of storytelling. ...

This got me curious — can we extract this “fable” factor? Can we measure it? And if so, can we check if somebody uses this style of communication more often than others? How about taking it for a spin on politics? Can we measure how “fable-ish” their speeches are? How oratorical or rhetorical? Is there any difference between how Democrats and Republicans speak? And if so, who’s the biggest fable-teller among them all?

I decided to build a “fable meter” — a computer program that takes text and classifies how scientific /fable-ish it is.

He collected what he thought were good examples of storytelling:  Andersen’s Fairy Tales, Grimm’s Fairy Tales, The Wizard of Oz, King James Bible, a few songbooks and poem books.

He also collected what he thought were good examples of technical prose, factual but dull: a few books about history, algebra 101, a full university chemistry course etc.

He then trained a Recurrent Neural Network on it. It reached 98% accuracy.

He then tried the inaugural speeches of some recent presidents. He chose those ones because they have the same context -- addressing the nation for the first time as president.

• Donald Trump (R): 35% oratorical
• Barack Obama (D): 22% oratorical
• George Bush II (R): 24% oratorical
• Bill Clinton (D): 18% oratorical
• George Bush I (R): Republican, 28% oratorical
• Ronald Reagan (R): 18% oratorical

You can try it out here with some selection of text: Fablemeter.com - measure your fables!

For my documentation of the Mathematica version of my Semisimple-Lie-Algebra package, it found "This text is 9% oratorical and 91% scientific."

Here is a bit of it:

Speed-up trick : use some variables as functions with several fixed arguments. Construct one and see what it evaluates out to. An unevalulated one will have a Head value equal to the variable name. Uses Mathematica's built-in search algorithm. No need for Position or MemberQ, which are much slower.

For "Tunguska and the Titanic" (first part, after "Here it is:"), "This text is 37% oratorical and 63% scientific." For "First Contact" in "Contact across the Solar System", "This text is 41% oratorical and 59% scientific."

But from chapter "Boston" is one of my favorite bits of that story:

Spring turned into summer, and with it came thunderstorms. On an afternoon when they were both home, there was a big one. Kalna was rather scared of its lightning and thunder but Ilmuth was curious enough to watch it. Kalna noticed Ilmuth's seeming fearlessness and Kalna cautiously stepped forward to join Ilmuth. Ilmuth telepathically sensed Kalna's fearfulness and put out a hand. Kalna grasped it and went to Ilmuth's side. Ilmuth put an arm around Kalna and Kalna did likewise with Ilmuth. The two watched the storm together, sitting on a porch, wearing thrift-shop summer dresses and sandals.

"This text is 72% oratorical and 28% scientific."

Isaac Asimov's famous science-fiction short story "Nightfall" is roughly half-half.

[lpetrich]

GreenScreen - Cut your images with AI discussed in Background removal with deep learning – Towards Data Science

The developers first thought of medical applications, but that has potential problems with legality and regulation, so they decided on something else that is potentially useful: background removal.

Much of the article is on how they did it.

Armed With Don't-Hurt-Humans Sensors, Robots Hit Small Factories - Bloomberg Quint

This Army of AI Robots Will Feed the World - Bloomberg

[lpetrich]

Note: I have merged three of my previous threads on AI advances into this one.

Robots Are Already Replacing Human Workers at an Alarming Rate

Acemoglu and Restrepo estimate that overall, for every 1,000 workers, an additional robot reduces the employment-to-population ratio by 0.18 to 0.34 percentage points, while also reducing wages by 0.25 to 0.5 percent. For context, the stock of industrial robots in the US increased fourfold during the period studied.

...
Experts predict that the stock of robots in the US will quadruple by 2025, jumping to 5.25 more robots per thousand workers (there are currently about 1.75 industrial robots in the US per 1,000 workers). This will serve to reduce the employment-to-population ratio by 0.94 to 1.76 percentage points, while resulting in 1.3 to 2.6 percent lower wage growth between 2015 and 2025. The addition of 5.25 more robots per thousand workers could also amount to a staggering 1.9 to 3.4 million additional job losses in just a decade. All of this, of course, is assuming the economies of the future behave the same way as our economy does today in response to additional automatons in the workforce

That may not seem like much, but over this time, I'm sure that AI will get better and better, more and more able to supplant many people's jobs.


What jobs will still be around in 20 years? Read this to prepare your future | US news | The Guardian

Martin Ford, futurist and author of Rise of the Robots: Technology and the Threat of a Jobless Future, explains the jobs that are most at risk are those which “are on some level routine, repetitive and predictable”.

Telemarketing, tax preparation, paralegals, legal assistants, fast-food cooks, ...

Surviving jobs will likely have at least one of these three features:

The first is jobs that involve “genuine creativity, such as being an artist, being a scientist, developing a new business strategy”. ...

The second area is occupations that involve building complex relationships with people: nurses, for example, or a business role that requires you to build close relationships with clients.

The third area is jobs that are highly unpredictable – for example, if you’re a plumber who is called out to emergencies in different locations.

In effect, creative intelligence and social intelligence.

[lpetrich]

AI, Robotics, and the Future of Jobs | Pew Research Center

ome 1,896 experts responded to the following question:

The economic impact of robotic advances and AI—Self-driving cars, intelligent digital agents that can act for you, and robots are advancing rapidly. Will networked, automated, artificial intelligence (AI) applications and robotic devices have displaced more jobs than they have created by 2025?

Opinions are about half-half.

Key themes: reasons to be hopeful

1. Advances in technology may displace certain types of work, but historically they have been a net creator of jobs.
2. We will adapt to these changes by inventing entirely new types of work, and by taking advantage of uniquely human capabilities.
3. Technology will free us from day-to-day drudgery, and allow us to define our relationship with “work” in a more positive and socially beneficial way.
4. Ultimately, we as a society control our own destiny through the choices we make.

Key themes: reasons to be concerned

1. Impacts from automation have thus far impacted mostly blue-collar employment; the coming wave of innovation threatens to upend white-collar work as well.
2. Certain highly-skilled workers will succeed wildly in this new environment—but far more may be displaced into lower paying service industry jobs at best, or permanent unemployment at worst.
3. Our educational system is not adequately preparing us for work of the future, and our political and economic institutions are poorly equipped to handle these hard choices.

[lpetrich]

Adam Lee of Daylight Atheism: The Post-Work Society -- mentioning various solutions to that conundrum. A permenantly unemployable underclass? A Universal Basic Income? Reduction in how much work everybody does?

He continues: The Robot Revolution Is Coming noting a lot of articles on the progress of artificial intelligence.

When Robots Take Over Most Jobs, What Will Be the Purpose of Humans? | HuffPost
Notes Is Job Polarization Holding Back the Labor Market? Liberty Street Economics by four economists at the New York Federal Reserve.

An occupation is routine if its main tasks require following explicit instructions and obeying well-defined rules. These tend to be middle-skilled jobs. If the job involves flexibility, problem solving or creativity, it’s considered nonroutine. Job polarization occurs when employment moves to nonroutine occupations, a category that contains the highest- and lowest-skilled jobs.

From 1978 to 2012, routine jobs declined from 60% to 40% and nonroutine ones rose in the same proportion, approximately linear over time.

The economists make this classification:

• Cognitive, Routine: office assistants, sales agents
• Cognitive, Nonroutine: managers, creatives, engineers, health care
• Manual, Routine: construction, mechanics, assembly line
• Manual, Nonroutine: waiters, security

From 1975 to 2015:

• Routine Manual: 35% - 25%
• Routine Cognitive: 25% - 20%
• Nonroutine Manual: 15% - 15%
• Nonroutine Cognitive: 25% - 40%

In conclusion,

The end state is uncertain, but we are headed toward a disruption of our society on the same order of magnitude as the rise of agriculture and industrialism, but in a much more compressed time frame: decades, not generations or centuries. And that question — what are people for? — will taunt us because it’s unclear if there is an answer or whether it is just an irresolvable dilemma.

[lpetrich]

https://www.theverge.com/2015/1/29/7939 ... -reporting

Minutes after Apple released its record-breaking quarterly earnings this week, the Associated Press published (by way of CNBC, Yahoo, and others) "Apple tops Street 1Q forecasts." It's a story without a byline, or rather, without a human byline — a financial story written and published by an automated system well-versed in the AP Style Guide. The AP implemented the system six months ago and now publishes 3,000 such stories every quarter — and that number is poised to grow.

Robot Radiologists Will Soon Analyze Your X-Rays | WIRED

Autonomous Robot Surgeon Bests Humans in World First - IEEE Spectrum

Robots Are Already Replacing Human Workers at an Alarming Rate noting Robots and Jobs: Evidence from US Labor Markets
From that paper's abstract: "According to our estimates, one more robot per thousand workers reduces the employment to population ratio by about 0.18-0.34 percentage points and wages by 0.25-0.5 percent."

So 400 industrial robots will do the work of 1000 human workers.

Back to Gizmodo:

Experts predict that the stock of robots in the US will quadruple by 2025, jumping to 5.25 more robots per thousand workers (there are currently about 1.75 industrial robots in the US per 1,000 workers). This will serve to reduce the employment-to-population ratio by 0.94 to 1.76 percentage points, while resulting in 1.3 to 2.6 percent lower wage growth between 2015 and 2025. The addition of 5.25 more robots per thousand workers could also amount to a staggering 1.9 to 3.4 million additional job losses in just a decade. All of this, of course, is assuming the economies of the future behave the same way as our economy does today in response to additional automatons in the workforce.

noting - Brynjolfsson_McAfee_Race_Against_the_Machine.pdf

More of a slow growth rather than a big burst. Like other technological revolutions that I've seen: the Internet, computers in general, ...

[lpetrich]

What jobs will still be around in 20 years? Read this to prepare your future | US news | The Guardian

Regular reports warn us that an automation apocalypse is nigh. In January, a McKinsey & Company study found that about 30% of tasks in 60% of occupations could be computerized and last year, the Bank of England’s chief economist said that 80m US and 15m UK jobs might be taken over by robots.

Given the track record so far, routine jobs, and jobs that are low on creative intelligence and social intelligence. But many of the remaining jobs may eventually fall.

Post-work: the radical idea of a world without jobs | News | The Guardian
subtitled

Work has ruled our lives for centuries, and it does so today more than ever. But a new generation of thinkers insists there is an alternative. By Andy Beckett

A nice long essay.

By the end of the 70s, it was possible to believe that the relatively recent supremacy of work might be coming to an end in the more comfortable parts of the west. Labour-saving computer technologies were becoming widely available for the first time. Frequent strikes provided highly public examples of work routines being interrupted and challenged. And crucially, wages were high enough, for most people, to make working less a practical possibility.

Instead, the work ideology was reimposed. During the 80s, the aggressively pro-business governments of Margaret Thatcher and Ronald Reagan strengthened the power of employers, and used welfare cuts and moralistic rhetoric to create a much harsher environment for people without jobs. David Graeber, who is an anarchist as well as an anthropologist, argues that these policies were motivated by a desire for social control. After the political turbulence of the 60s and 70s, he says, “Conservatives freaked out at the prospect of everyone becoming hippies and abandoning work. They thought: ‘What will become of the social order?’”

It sounds like a conspiracy theory, but Hunnicutt, who has studied the ebb and flow of work in the west for almost 50 years, says Graeber has a point: “I do think there is a fear of freedom – a fear among the powerful that people might find something better to do than create profits for capitalism.”

So it's work for the sake of work, it seems. Anyone who believes that ought to be sentenced to digging holes and filling them up again.

Whether you look at a screen all day, or sell other underpaid people goods they can’t afford, more and more work feels pointless or even socially damaging – what the American anthropologist David Graeber called “bullshit jobs” in a famous 2013 article. Among others, Graeber condemned “private equity CEOs, lobbyists, PR researchers … telemarketers, bailiffs”, and the “ancillary industries (dog-washers, all-night pizza delivery) that only exist because everyone is spending so much of their time working”.

Seems like a case of Parkinson's Law -- work expanding to fill the time available.