[syndicated profile] lewis_and_quark_feed

Neural networks are a kind of machine learning computer program that are very, very good at making their own internal rules about a dataset. If you give them a list of ANYTHING, they’ll figure out how to generate more stuff like it.

This makes neural networks very good at naming things - they can figure out what letter combinations and sounds make a word sound like a name for a kitten, or a paint color, or a Pokemon, or a D&D spell, or a guinea pig, or a metal band, or even a scary guinea pig.

There are almost 3,500 exoplanets confirmed to date and thousands more candidates waiting for confirmation, all with names like “Kepler-452b” and “55 Cnc-c” and “2MASS J01225093-2439505b”. Clearly, these names are not going to work if we have to one day shout them to the ship’s engineer during a raging ion storm.

Fortunately for the future of space travel, blog reader Chris Jones has sent me a list of almost 700 Star Wars planets. And in short order, the neural network was producing this:

Bartan
Cenron
Nalarar
Bondal
Ballor
Van-Karal
Valtane
Vantos
Malalas
Mateot
Tiris
Kanan
Montaan
Tardor
Nananon
Moridia
Tatloor

This is unedited output directly from the neural network, and I had to check to make sure that none of these names were in the original list - they’re just that plausible.

At a higher creativity (more random) setting, the names are a bit less pronounceable on average, but still could pass for Star Wars planets:

La Vok
Slorru G
Sakani
Vaszalu
Varkena
Baro IV
Toran
Hatnlant
Uluunna
Baroa
Tina
Duperda
Bantak
Barkaan
Ban Beraou
Baxuor
Rrarar

That’s not to say there weren’t strange results. 

Birdanan
Boldura
Balmara
Barmen
Garden
Carton
Loner
Robes
Sara
Loon
Laser
Bunlalavor
Bal Panda

This stems partly from the fact that the neural network has no idea what English words are - all it knows is Star Wars planet names, and so sometimes it ventures unknowingly into territory where the English language has gone before. There aren’t nearly as many of these accidental words as in other datasets, perhaps partly because these planets were designed to NOT sound like English.

I wanted to learn more about some of these planets, so I turned to two procedural programs. Unlike neural networks, in procedural generators human programmers make the rules rather than the neural network inventing its own. The only outputs you’ll get are those that the human programmer thought of building in, but this can still produce a huge range of interesting results.

I used the Twitter bot @I_find_planets by Charles Bergquist to generate descriptions of the planets, and the procedural planet generator by wwwtyro to generate the pictures.

image
image
image
image

Will we ever run out of neural network planet names? Sure, because there’s only a limited number of ways that characters can be combined into words before we’re back to indistinguishable mishmashes of letters and numbers. But in the meantime, the neural network could help us make our solar neighborhood a more pronounceable place.

[syndicated profile] lewis_and_quark_feed

(Marquee graphic generated with RedKid.net’s sign generator)

Neural networks are a kind of machine learning program modeled very loosely after the human brain. By looking at a dataset and tuning the connections between their own virtual neurons, neural networks can learn to imitate the original dataset. Powerful neural networks can do impressive things, like translate languages, recognize faces, and even describe scenes in words.

The neural networks I use, however, have the approximate processing power of an earthworm.

Granted, it’s an earthworm that has put its entire tiny brain to the sole task of learning the dataset: I’ve taught neural networks to invent cookbook recipes, name cats and guinea pigs, and even write Harry Potter fan fiction, all with some degree of success.

image

What would a neural network make of Broadway musicals? Thanks to lyricist Robby Sandler, we have a chance to find out. Robby pulled a list of over 10,000 Broadway plays and musicals (names, starting dates, closing dates, and total number of performances from the Internet Broadway Database, and I gave them to the open-source char-rnn framework written by Andrej Karpathy.

The first thing I noticed was the time travel.

Mh Radpor
Play, Comedy

Opening: Aep 00, 1969
Closing: Ser 30, 1917
Closing: May 11, 1922
Performance Count: 15

Santen Sos    
Play

Opening: Nov 19, 1906
Closing: Cov 19, 1907
Closing: May 29, 1923
Closing: Fec 24, 1929
Closing: Jat 19, 1927
Closing: May 1935
Performance Count: 14 

Apparently, the plays produced by the neural network in its early stages were so bad that multiple teams of time travelers (or perhaps one very, very determined traveler) had to travel back in time to close them not only decades before they opened, but also retroactively in multiple alternate timelines.

One play called “The Gore” was so powerful that even closing the play 17 times over at least three timelines still resulted in 24 performances.

The Gore    
Play, Comedy

Opening: May 19, 1919
Closing: Jay 19, 1920
Closing: May 19, 1922
Closing: May 19, 1924
Closing: May 19, 1932
Closing: May 19, 1919
Closing: Nov 19, 1922
Closing: May 19, 1919
Closing: Nov 19, 1927
Closing: May 19, 1925
Closing: May 19, 1927
Closing: May 19, 1922
Closing: May 19, 1917
Closing: May 19, 1932
Closing: May 19, 1925
Closing: May 19, 1925
Closing: May 19, 1921
Performance Count: 24

As the neural network progressed in its learning, the situation got less dire, but causality remained highly optional. The plays now only close once, but remain still really weird, and probably quite bad.

The Girls of Hurk    
Play, Comedy
Opening: Mar 17, 1991
Closing: May 05, 1980
Performance Count: 151

Meatlick    
Musical, Comedy
Opening: Mar 28, 1928
Closing: Nov 02, 1929
Performance Count: 133

The Wither Bean    
Musical, Comedy
Opening: Sep 11, 1920
Closing: Jan 1936
Performance Count: 25

Mep and the .    
Musical, Revue
Opening: Jan 04, 1906
Closing: Apr 25, 1901
Performance Count: 58

Worms and Ram    
Play
Opening: Nov 24, 1918
Closing: Feb 1919
Performance Count: 12 

Is a Boot    
Play, Melodrama
Opening: Feb 03, 1900
Closing: Feb 1900
Performance Count: 22

Hot Stans    
Play
Opening: Feb 17, 1943
Closing: Mar 20, 1945
Performance Count: 1 

The Burking Ding of 190 Bour Dadige    
Play, Comedy
Opening: Sep 24, 1929
Closing: Jan 1938
Performance Count: 55

The neural network also seemed, for mysterious reasons, to zero in on butt-related plays.

Butt    
Play, Comedy
Opening: Jun 09, 1899
Closing: Nov 1906
Performance Count: 1

Bum    
Play, Play with music
Opening: Feb 12, 1930
Closing: Mar 1922
Performance Count: 1

Fart    
Play
Opening: Oct 13, 1923
Closing: Apr 03, 1927
Performance Count: 23

Buttosty    
Play, Comedy
Opening: May 14, 1907
Closing: Apr 1908
Performance Count: 12

Bun Life    
Play
Opening: Mar 06, 1995
Closing: Apr 22, 1999
Performance Count: 175

The Old Farting    
Play, Melodrama
Opening: May 24, 1927
Closing: Mar 01, 1926
Performance Count: 48

And, in addition to a continuing disregard for causality, there emerged a last category of plays and musicals that seemed strangely significant. 

The Siri    
Play
Opening: Nov 14, 1932
Closing: Sep 1914
Performance Count: 12

Bot Five    
Play, Comedy
Opening: Oct 25, 1958
Closing: Sep 02, 1959
Performance Count: 256

The Romance of the Bot    
Play
Opening: May 14, 1923
Closing: Jun 1924
Performance Count: 109

Perhaps “bot” is simply a highly-probable English word. Perhaps “butt” is, too. Or perhaps the neural network has examined the total creative output of 100+ years of Broadway, and is letting us know what it likes best.

Taking the chance

11 Jul 2017 03:41 pm
[syndicated profile] neatscience_feed

Posted by thebrightspark

Why did Angelina Jolie have a double mastectomy?

Another question from Jess, via Twitter. I’m sure most people heard about Angelina Jolie’s decision to have a double mastectomy after discovering she carries a faulty copy of a gene called BRCA1, at the time, and her piece about her decision seems to have caused a wave of other women to get the same testing done.

Angelina Jolie in one of her most recognisable film roles.

BRCA1 and 2 genes confer a risk of developing breast cancer, and in fact a number of other cancers too, including ovarian cancer. There are some widely varying estimates on how likely these mutations are to cause significant risk, and often these figures are quoted to encourage people to get testing and to undergo preventative procedures, like Jolie’s double mastectomy. For an individual, it typically increases risk of developing breast cancer less than 50% — in other ways, the chances are equally good that someone carrying a faulty BRCA1 or 2 gene will not actually develop breast cancer. BRCA mutations account for only 5-10% of breast cancers, and 10-15% of ovarian cancers.

So should everybody be tested for the faulty BRCA genes? Probably not. Angelina Jolie was tested for a specific reason: her mother died of breast cancer and died before the age of 60. That kind of family history can suggest a risk of carrying a harmful variant of the gene, and indeed, it proved to be true in Angelina Jolie’s case. She was quoted “an 87 percent risk of breast cancer and a 50 percent risk of ovarian cancer”. Given those statistics, it’s not surprising that she decided to take preventative measures.

It’s still important to remember that a percentage risk is not a promise that something will happen. Even when something has a 99% chance of happening, in one case in a hundred (on average), it will not happen. Any operation carries risks to do with anaesthesia and post-operative infections, as well. It’s worth asking whether that risk is worth it; an individual’s situation is always going to be unique.

What would I do if I found out I carried a faulty copy of BRCA1 or 2? I’d educate myself on symptoms and warning signs, make sure I got checkups whenever I found something concerning, and otherwise try and reduce my risk. I wouldn’t go for a double mastectomy unless the risks were judged to be significant within the next ten years. After all, you can always have the operation done later in life as your risk increases due to age. What you can’t do is take it back.


Profile

firecat: red panda looking happy (Default)
firecat (attention machine in need of calibration)

June 2017

S M T W T F S
    123
45678910
11121314 1516 17
18192021222324
252627282930 

Tags

Style Credit

Expand Cut Tags

No cut tags
Page generated 24 Jul 2017 10:46 am
Powered by Dreamwidth Studios