Program Synthesis, Self-Programming, DeepCode - CodeGen - Synthesis of Everything - Казбород

Bulgarian research group in Code Synthesis in Switzerland

Do you know about a trendy research group in Code Synthesis that runs in Zurich, Switzerland? Its leaders are several Bulgarian researchers from ETH with a spin-off start-up called "DeepCode". They use "Big Code", such as GitHub and neural networks, however combined with other methods.

A recent article, an interview witn Martin Vechev, one of the leaders of the DeepCode

https://www.republik.ch/2018/06/27/programmier-dich-doch-selbst

(Mein Deutsch ist nicht so gut, ebenfalls, aber... Google Translate from German to English is good enough. :)  )

Older Notes from 5.2.2018 :

http://www.srl.inf.ethz.ch/riai.php

A recent University lecture course on Reliable and Interpretable Artificial Intelligence... (Fall 2017) ... taught in ETH Zurich by several fellow bulgarian researchers:

Deep Code, Code Synthesis, AI

Thanks to Ivan Dzheferov for the links!

...

Todor and Code Synthesis

For the record, I've been in the Code Synthesis domain, too, but it "doesn't count yet" publicly, because it was not in an academic environment and style, it was part of my general AGI notes and research activities and I haven't published it yet, it's not completed as practical yet.

However it aims at "AGI-Complete" code synthesis which is integrated with the "general AGI", language and vision. So far the ideas were not based on DNN, millions-lines of code style, and were more "conceptual", aiming at hitting a lot of targets with a few bullets. For example - by writing a little code and letting it find and write the rest itself.

However I've switched focus. A friend and a developer, Ivan, notified me about that research group some time ago - because he knew my "radical" claims, maybe since 4-5 years*, that I believed that the software development had no future as a domain and profession.

Software should be developed (generated) automatically and certainly not the way it was done either then or now.

It was quite a claim to express in front of developers, so I avoided to do it in public :D, with a few exceptions, since I expected to be ridiculed by people who had no clue. They believed  that "AI"* was a "science fiction", computers "can't do that". "AI" - because even today the "AGI" is not a popular term.
In an interview with Martin Vechev for a Bulgarian media, he mentions AGI at the end as a distant goal with his Bulgarian translation ("Всеобщ изкуствен интелект"; моят термин беше "Универсален изкуствен разум" и др. варианти, както и само "Изкуствен разум" - за да се разграничи от опорочения "интелект").

I'd better focus back. :))

[Article trimmed ... To be continued...]

Read More

Delayed gratification is an ILL-Defined Concept

https://www.technologynetworks.com/neuroscience/news/the-marshmallow-test-todays-kids-show-more-self-control-305353

The article claims that today's children have shown more self control.

I question the abstract settings of that kind of experiments, the generality and the direction of the conclusions.

Comment: 

The setting is ill-defined, the concept of "delayed gratification" itself, too, especially for little children. Too general conclusions. The experimenters assume that the children believe that two candies later are "a bigger gratification" than one. Why should the child do? Why not a child be satisfied with just ONE now and then go do something else, seek for another, different, more meaningful and interesting gratification - such as play? If it's satisfied anyway, why should her wait?

"A higher salary is better" - is it, at what other cost; are the money (or the number of candies) the only unquestionable "gratification" measure.

What if the children didn't understand the question like the experimenter has defined it? What if they waited more, because they were more suggestive and obedient to do what they were told. (It's true that the "schooling" from an earlier age contributes to developing in that direction.)

...Or because they have other rewards, such as "video drugs" and care less about that candy. Or because they could get a candy anyway afterwards and they are not attracted.

Also are those two candies (eaten at once) a bigger gratification? In practice they would be eaten for about the same time - do the little children count that? (If they saved it for another day, that would be a "delayed gratification").

Do the children understand "more" the same way and also *do they believe, that the waiting costs less*, and when they wait, *do they wait because they like more to please the authority figure who gives them the task*?

One thing that the test measures is some "patience", assuming that it's by-definition suffered for "gratification", and namely for the eating of the candy itself.

The article mentions "not on medication for ADHD" and "attention", but I think that's ill-defined too, because the children who don't want to wait for *a candy* may do wait for something else *about which they do care* more and is of a "value" for them, i.e. I don't think the conclusions are transferable by default, *especially* since the child are young and probably do not always generalise themselves.



"Delaying gratification" for a setting of a subjectively accepted higher reward could be interpreted as more "GREED" and for the case of little a few-years-old children: a higher susceptibility to please the authority figure or to answer what she assumed she was expected to say.

Also I suspect that some of the children do not understand all of the conditions of what they were asked to do.

...

It is true that waiting as Patience and "sustained focus" is correlated to "higher IQ" and other test results in *some tests* - studying and "success" require patience.

It is true also, that pleasing the authority figures in human societies usually leads to "success" in the measures and values, defined by those authority figures.

However patience and "delayed gratification" are correlated also with less ability to contradict the order of the authority figure, which is correlated to less inclination towards critical thinking and creativity under authority pressure.

The rewards for those "delayed gratification" ones is to a bigger extend defined by their authorities and these children may have have accepted and adopted more deeply the values of their "experimenters" and are delaying gratification", they question less the truth of the values they have.

Similarly, some of the children who according to their experimenters "lacked self-control" (based on the experimenters definition), may actually have *REJECTED* or ignored the external control, imposed by the experimenter/teacher and thus they did what they wanted, instead of what they were supposed to do according to the experimenter's values and "gratification criteria" etc.

(I may have encountered similar thoughts in the past).

Read More

Кръг Artificial Mind | Circle Artificial Mind - AGI, машинно обучение, правене на филми и игри, събития, конференции, изследвания

Интервю с мен за новия ми проект:

http://mind.twenkid.com 

Фокус: изкуствен интелект, универсален изкуствен разум (Artificial General Intelligence, AGI), машинно обучение, правене на филми и игри; спорт, купон и др.

Продължение на опита с "мини-конференцията", която организирах през 2012 г. в Пловдив и "Дружество Разум" от юношеските ми години в началото на 2000-те + филмови, обществени и др. неща от "моя дух".

Историята е драматична и включва разпадналото се сдружение-хакерспейс и споделено работно място "Хакафе" (Hackafe Plovdiv). Споменават се някои от другите ми творчески занимания напоследък. - прочетете в интервюто.

Търсят се "екшън герои" за партньори. Като може би през есента, ако намеря достатъчно партньори, ще организираме втора мини-конференция. Засега един от старата компания от тогава одобри идеята.

(...)

Р: (...) Бихте ли разказали повече за "конференцията"?

Т: Разбира се, че думата "конференция" и претенциозното й име са употребени с чувство за хумор. Събирането също мина със смях, както си личи и на снимките и по "хавайската ми риза", но участниците бяха сериозни: Светлин и Даниел сега са докторанти по роботика и изкуствен интелект в Единбург; Орлин стана известен инженер роботист. Имахме моралната подкрепа на Петър Кормушев, тогава ръководител на изследователска група по роботика в Италианския научно-изследователски институт в Геноа, а на следващата година - носител на наградата "Джон Атанасов". Там бяхме още моя милост - авторът на първите два университетски курса по AGI в света, - един от най-добрите ми студенти от втория курс и двама други гости.  (...)

Read More