Discussion: Cours online & gratuit : Introduction to Artificial Intelligence - Fall 2011

1. Envoyé par Franck Dernoncourt
Ca ne rigole pas par contre :
Ca ne rigole pas aussi du coté d'Udacity. J'ai complêtement foiré le premier homework. J'ai par contre réussi à faire marché le code 10 minutes après la date limite.

Mais bon, c'est mon premier code python
Code python :
```123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101
colors = [['red', 'green', 'green', 'red' , 'red'],
['red', 'red', 'green', 'red', 'red'],
['red', 'red', 'green', 'green', 'red'],
['red', 'red', 'red', 'red', 'red']]

measurements = ['green', 'green', 'green' ,'green', 'green']
motions = [[0,0],[0,1],[1,0],[1,0],[0,1]]
sensor_right = 0.7
p_move = 0.8

def show(p):
for i in range(len(p)):
print p[i]

#DO NOT USE IMPORT
#ENTER CODE BELOW HERE
#ANY CODE ABOVE WILL CAUSE
#INCORRECT

equi_prob = 1./ (len(colors) * len(colors[0]))
mat_x = len(colors)
mat_y = len(colors[0])

p = [[equi_prob for x in range(mat_y)] for y in range(mat_x)]

def sense(p, Z):
q = p
s = 0
for i in range(mat_x):
for j in range(mat_y):
hit = (Z == colors[i][j])
q[i][j] = p[i][j] * (hit * sensor_right + (1-hit) * (1-sensor_right))
s += q[i][j]

for i in range(mat_x):
for j in range(mat_y):
q[i][j] = q[i][j] / s
return q

def moveRight(p):
q = [[equi_prob for x in range(mat_y)] for y in range(mat_x)]
for i in range(mat_x):
for j in range(mat_y):
k = i+1
if k >= mat_x:
k=0
q[k][j] = p[i][j]*(p_move) + p[k][j]*(1-p_move)
return q

def moveLeft(p):
q = [[equi_prob for x in range(mat_y)] for y in range(mat_x)]
for i in range(mat_x):
for j in range(mat_y):
k = i-1
if k < 0:
k=mat_x-1
q[k][j] = p[i][j]*(p_move) + p[k][j]*(1-p_move)
return q

def moveUp(p):
q = [[equi_prob for x in range(mat_y)] for y in range(mat_x)]
for i in range(mat_x):
for j in range(mat_y):
k = j-1
if k < 0:
k=mat_y-1
q[i][k] = p[i][j]*(p_move) + p[i][k]*(1-p_move)
return q

def moveDown(p):
q = [[equi_prob for x in range(mat_y)] for y in range(mat_x)]
for i in range(mat_x):
for j in range(mat_y):
k = j+1
if k >= mat_y:
k=0
q[i][k] = p[i][j]*(p_move) + p[i][k]*(1-p_move)
return q

def process(p):
for i in range(len(measurements)):
if motions[i][0] == 1:
p = moveRight(p)
if motions[i][0] == -1:
p = moveLeft(p)
if motions[i][1] == -1:
p = moveUp(p)
if motions[i][1] == 1:
p = moveDown(p)
p = sense(p, measurements[i])

return p

p = process(p)

#Your probability array must be printed
#with the following code.
#
show(p)```

2. Ah sur Udacity j'ai préféré ne suivre que CS 101 ce semestre
Par contre je suis un assez agacé par le peu de temps qui nous est laissé entre la publication des homework et la deadline pour les soumissions. Là par exemple je viens de voir que le homework #2 a été publié aujourd'hui et la deadline est mardi prochain... cela paraît con mais il suffit d'avoir un weekend prolongé de prévu, un petit coup de bourre au boulot, un concours de prévu qui s'étale sur plusieurs jours, etc., et on rate la deadline. Je préfèrerai que les cours (vidéos et homework) soient moins en flux tendus. Même problème du côté de Coursera. A cause de cela je suivrai moins de cours que ce que je pourrai. Au moins ils pourraient essayer de donner 1 semaine entière si vraiment ils ont peur que le cours n+1 spoile le homework n (et accessoirement d'envoyer un e-mail lorsqu'un cours ou homework est mis en ligne - car actuellement je ne reçois rien, peut-être une option que je n'ai pas vue ?)

3. Je suis les deux cours actuellement proposés. J'avoue que peu de temps nous ai laissé entre les cours et les homework. J'ai pu finir le homework pour le self drive car in extremis.

Sinon, les homework pour les units II sont en lignes depuis Mardi. Je crois qu'il faut y aller assez souvent pour voir les mise à jour.

4. De mémoire et d'après ce que je peux voir sur les annoucements, HW2 a été publié mercredi matin. Mais bon, je ne vérifie pas tous les jours et je confonds peut-être avec d'autres deadlines.

http://www.udacity.com/announcements...rseRev/feb2012 :

5. Intéressant : http://tech.mit.edu/V132/N12/agarwal.html

Agarwal is leaving CSAIL to direct MITx
_____________
Steps down as CSAIL director in order to lead MITx team full-time
By Ethan A. Solomon executive editor
March 16, 2012

Anant Agarwal, director of the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), announced yesterday that he would step down from his role as director to fully devote himself to the Open Learning Enterprise (working title), which runs theMITx initiative. Since its announcement last December, MITx has caught the eye of MIT faculty, the world of higher education, and 120,000 people who signed up for the pilot course, 6.002x.

“I am writing to let you know about an extraordinarily difficult decision that I have made to step down as the director of CSAIL in order to serve as full-time director of the new MITx open learning enterprise,” Agarwal wrote in an email to CSAIL yesterday evening. “As some of you may know, I have been trying to lead both CSAIL and MITx these past several months, but I have come to realize that this situation cannot continue indefinitely.” Agarwal was named CSAIL head last July.

Though Agarwal has been spearheading the development of MITx and its prototype course since last year, he now takes the reins of the project full-time as it begins to get more attention from students and faculty, and as it prepares to expand its course offering for the fall term. Faculty devoted their January/February newsletter to discussion on MITx, and the Undergraduate Association has started to consider the project’s impact on students. Chancellor Eric Grimson PhD ’80 and Provost L. Rafael Reif will hold “overview of MITx” sessions with students next week and in early April.

6.002x, an online prototype version of MIT’s introductory circuits and electronics class, went live to the world on March 5, and Agarwal, who also co-teaches the course, says 120,000 people have signed up so far. But less than three months from unveiling to prototype seems fast — why so quick?

“I think MITx is very, very important for MIT,” says Agarwal. “It’s critical that we get out there and do it fast.”

The rapid pace of technical development, Agarwal adds, is thanks to the course development team. In addition to Agarwal, Gerald J. Sussman ’68, Christopher J. Terman PhD ’83, and Piotr Mitros ’04 are working on 6.002x. The course was a “heroic team effort,” according to Agarwal.

Agarwal plans to run MITx like a nonprofit startup, inspired by the small-team energy that has come to typify Internet startups from Silicon Valley to Kendall Square. The MITx team, in fact, is looking for office space in Kendall Square, either in an MIT building or not.

Agarwal says he’s been surprised by what he sees as a community that has developed around 6.002x. 6.002x students — from all over the world — have taken advantage of tools like a discussion forum and wiki to help each other learn the material, much more than expected. The need for 6.002x course staff to handle questions is mitigated by more advanced students helping others.

“I was petrified about what was going to happen when people jumped on the discussion forum,” said Agarwal. “Whether we could scale, or whether we could handle all the questions.”

User response to 6.002x has been positive, said Agarwal. Reviewers on AIQUS.com — originally established to host questions and answers about Stanford’s online Introduction to Artificial Intelligence course — gave MITx glowing reviews.

“I just peeked at [6.002x] and I assure you guys, this course is the best among all in terms of course content and layout,” wrote one reviewer.

“This is miles ahead of both Coursera and Udacity. Nicely done MIT,” wrote another, referencing two other online educational platforms.

Still, users have also asked for improvements. Homework deadlines were changed from Friday to Sunday to better accommodate work schedules, and some users have asked for better navigation of the online textbook, for instance.

And though many faculty have greeted MITx warmly, some have raised concerns over its implementation. Agarwal acknowledges that there is a debate to be had over what form an online educational experience should take.

“Clearly there are concerns, and some very valid concerns,” he said. “We’re breaking new ground and we’re trying to figure out what works best.”

MITx has garnered “a lot of interest” from other universities since December’s announcement, though Agarwal declined to say precisely who was interested in the platform. MIT has billed MITx as an open system that other institutions can leverage to offer their own courses.

Agarwal will continue to serve as CSAIL head until his successor is picked by a search committee — the same committee that picked Agarwal, according to his email to CSAIL yesterday.

http://tech.mit.edu/V132/N12/agarwal.html

6. Impressionnant. C'est le second, après Sebastian Thrun, à quitter son université pour se consacré aux cours en ligne.

7. Enfin, pas vraiment quitter l'université, juste abandonner quelques fonctions qu'ils avaient (Thrun, sa prof. tenure à Stanford, Agarwal, son rôle de directeur du CSAIL)

8. "A real Caltech course on Machine Learning, not a watered-down version": http://work.caltech.edu/telecourse.html

A real Caltech course, not a watered-down version
• Broadcast live from the lecture hall at Caltech
• Introductory Machine Learning course
• Taught by Caltech Professor Yaser Abu-Mostafa [article]
• Prerequisites: Basic probability, matrices, and calculus
• Live Q&A and online homeworks
• Scoreboard ranking of participants

Outline
This is an introductory course on machine learning that covers the basic theory, algorithms, and applications. Machine learning (ML) enables computational systems to adaptively improve their performance with experience accumulated from the observed data. ML techniques are widely applied in engineering, science, finance, and commerce to build systems for which we do not have full mathematical specification (and that covers a lot of systems). The course balances theory and practice, and covers the mathematical as well as the heuristic aspects. Detailed topics are listed below.

Schedule
All lectures are 5:30-6:30 PM GMT, followed by Q&A
• Lecture 1: The Learning Problem (Tuesday, April 3)
• Lecture 2: Is Learning Feasible? (Thursday, April 5)
• Lecture 3: The Linear Model I (Tuesday, April 10)
• Lecture 4: Error and Noise (Thursday, April 12)
• Lecture 5: Training versus Testing (Tuesday, April 17)
• Lecture 6: Theory of Generalization (Thursday, April 19)
• Lecture 7: The VC Dimension (Tuesday, April 24)
• Lecture 8: Bias-Variance Tradeoff (Thursday, April 26)
• Lecture 9: The Linear Model II (Tuesday, May 1)
• Lecture 10: Neural Networks (Thursday, May 3)
• Lecture 11: Overfitting (Tuesday, May 8)
• Lecture 12: Regularization (Thursday, May 10)
• Lecture 13: Validation (Tuesday, May 15)
• Lecture 14: Support Vector Machines (Thursday, May 17)
• Lecture 15: Kernel Methods (Tuesday, May 22)
• Lecture 16: Radial Basis Functions (Thursday, May 24)
• Lecture 17: Three Learning Principles (Tuesday, May 29)
• Lecture 18: Epilogue (Thursday, May 31)

9. Très bon article de Wired : The Stanford Education Experiment Could Change Higher Learning Forever

Extrait :

In June he took the next step: cofounding KnowLabs, which he funded with \$300,000 of his own money. He pulled in David Stavens, one of Stanley’s cocreators, as CEO; he tapped Stanford robotics researcher Mike Sokolsky to be CTO. They converted Thrun’s guesthouse into a temporary office. Thus ensconced on a scenic hillside on Page Mill Road near Stanford’s campus, the team began planning. They had eight weeks before the fall term started—not unreasonable given the modest scope of the project. Stavens thought they’d get 500 students. Sokolsky hoped for 1,000. Norvig figured they might hit 2,000.

In late July, Thrun emailed 1,000 members of the Association for the Advancement of Artificial Intelligence, a group that had weathered the AI winter of the 1980s and ’90s only to see the field later revitalized by the likes of Stanley. By the next morning 5,000 students had signed up. A few days later the class had 10,000. That’s when the Stanford administration called. Thrun had neglected to tell them about his plan—he’d had a hunch it might not go over well. The university’s chief complaint: You cannot issue an official certificate of any kind. Over the next few weeks, 15 meetings were held on the matter. Thrun talked to the dean’s office, the registrar, and the university’s legal department. Meanwhile, enrollment in CS221 was ballooning: 14,000, 18,000, and—just two weeks later—58,000.

In all those meetings, not one person objected to Thrun’s offering his class online for free. They admired his vision. The administration simply wanted Thrun to drop the assignments and certificate. He refused. Those two components, he argued, were responsible for driving the sign-ups. Someone proposed removing Stanford’s name from the course website altogether. Eventually they reached a compromise: (1) Offer a Statement of Accomplishment, not a certificate, and (2) include a disclaimer stating that the class wouldn’t count toward Stanford credit, a grade, or a degree.

Thrun didn’t have time to celebrate. By mid-August, word of his AI class went viral after a write-up in The New York Times. Enrollment skyrocketed past 100,000. KnowLabs’ website had been built to handle 10,000 students. Class was starting in a matter of weeks. “That,” Sokolsky says, “is when I stopped sleeping.”

10. Pour les philosophes en herbe : Five critiques of the Open Educational Resources movement .

Abstract :
This post will review existing literature on Open Educational Resources, introducing five critiques: 1.) An under-theorisation of ‘openness’, in which the concepts of positive and negative liberty will be used to suggest a neglect of coherent theorisation concerning the practice of self-directed learning. 2.) The simultaneous privileging and rejection of institutional authority, where OER literature will be shown to endorse the reputations of established institutions while claiming liberation from them. 3.) The diminishing of the role of pedagogy, in which OER will be aligned with an untheorised learner-centred model of education. 4.) Humanistic assumptions of unproblematic self-direction and autonomy, and 5.) an alignment with the needs of capital, in which a Foucauldian interpretation of subjectivity will offer alternative perspectives on the notions of power and emancipation in OER discourse. It is suggested that these critiques may provide a framework for OER to develop a theoretically rigorous area of scholarship.

11. Attention pour ceux qui suivent Udacity CS101, la deadline du final exam est Sunday, 8 April at 23:59 UTC, c'est-à-dire plus tôt que la deadline habituelle pour les homeworks. (peut-être est-ce également le cas pour Udacity CS373).

EDIT :

---------- Forwarded message ----------
From: Udacity E-mail Robot <hal@knowlabs.com>
Date: Sat, Apr 7, 2012 at 19:42
Subject: Final exam reminder and extension

Dear Franck Dernoncourt,

We hope you enjoyed Unit 7! If you haven't had a chance to watch it, all of the course material will continue to be available after the course finishes.

We apologize for the trouble submitting programs yesterday and have extended the deadlines for both CS101 and CS373 by 24 hours until 23:59 UTC on Monday the 9th of April.

Registration for the April 16th offering of CS101 and CS373 is now available for anyone who is not already taking them. We hope you will encourage your friends who missed the classes this time to try them out this time. If you would like to sign up for the same class again you will be able to do so next week after your class has finished.
[...][/url]

12. Encore un article intéressant : MITx aims to enhance on-campus education.

For MIT’s own students, Reif said, materials developed for MITx classes will add an extra dimension to their on-campus experience. In some cases, he said, MITx might replace existing lecture components with online versions students could watch at their own pace, repeating segments if necessary to make sure they understand. Instead of a classroom experience defined by passive listening, MIT students might actually “experience increased face-to-face interactions” with each other and with instructors, Reif said.

“We believe this is the future,” Reif said of interactive online education as a supplement to campus learning. The software platform will be offered to other institutions for use in their classes — providing an alternative to a trend toward online educational materials developed by for-profit companies, he added.

13. Probabilistic graphical models

Salut à tous,

Je suis le cours Probabilistic Graphical Model sur coursera.org, et j'éprouve de très grandes difficultés dans le programming assignement 4 ( j'y ai déjà consacré 8h, je n'ai réussi à implémenter qu'une seule fonction, et je ne comprends pas ce qu'il faut faire dans les suivantes!!!!! ).
Cette fois le forum de ce cours ne m'est d'aucun secours! ( Je dois être trop nul! ).
J'aimerai trouver un bon cours ou tuto sur les "clique tree" ( éventuellement en français, mais si c'est en anglais je prends aussi ), avec des exemples si possible ( si quelqu'un peut me donner une traduction de "clique tree" en français ça m'aiderai aussi )

Merci

Edit : Bon ben j'ai réussi à progresser, finalement, mais en ce qui me concerne je trouves ce cour passionnant certes, mais très difficile ( Bien plus difficile que le cours de Andrew Ng sur le Machine learning de l'automne dernier ( pour donner une idée ) )! De plus, il me demandes de 8 à 12h par semaines... Mais je m'accroche .

14. Désolé, je ne peux pas t'aider, je ne suis pas ce cours notamment car j'avais vu quelque part qu'ils estimaient la durée de travail à 10h par semaine

15. Encore un article intéressant de ComputerWorld résumant bien les différentes initiatives : MIT and others launch a tech education revolution

Il est intéressant de remarquer qu'ils commencent à se demander publiquement comment rentrer dans leurs frais à terme. MITx et Udacity feront peut-être payer l'obtention de certificats.

16. A propos de certificats, voici en PJ ce à quoi ressemblent les certificats de Udacity.

17. Envoyé par Franck Dernoncourt
A propos de certificats, voici en PJ ce à quoi ressemblent les certificats de Udacity.
Oui, j' ai eu le même un moi aussi pour le CS101. Par contre pour le CS373, je crois que j'ai encore des progrès à faire.

18. Envoyé par darrylsite
Oui, j' ai eu le même un moi aussi pour le CS101. Par contre pour le CS373, je crois que j'ai encore des progrès à faire.
Idem, je voulais faire CS373 mais j'ai l'impression qu'il y avait beaucoup de code à écrire... Pas le temps

19. Envoyé par Franck Dernoncourt
Idem, je voulais faire CS373 mais j'ai l'impression qu'il y avait beaucoup de code à écrire... Pas le temps
Oui, il y avait du code à écrire et comprendre les méthodes utilisées n'étaient pas toujours évident. Pour l'examen final, je n'est pas eu le temps de faire les question sur la programmation vu que le temps estimé pour résoudre certains exos dépassait une heure.

Sinon, actuellement je suis le cours sur la cryptographie et la construction d'un langage de programmation. Le prof à l'air très compétent.

20. Je sais pas encore ce que je vais suivre, peut-être Udacity cs253: Web Application Engineering ou Coursera: introduction à la logique. Beaucoup de cours m'intéressent mais à vrai dire je ne suis pas un grand fan de suivre un cours à 100% car il y a en général trop d'overhead (pour moi en tout cas).

Sinon malins les profs du MIT : MIT launches student-produced educational video initiative. Ils délèguent la production de certaines vidéos aux étudiants

“We wanted to help inspire young people to change the world through engineering and science, and realized that the 10,000 superstar students we have at MIT are uniquely positioned to do that,” says Ian A. Waitz, dean of the School of Engineering and the Jerome C. Hunsaker Professor of Aeronautics and Astronautics. “Our students have responded with all the energy and enthusiasm we knew they would. We worked with them to design the program, and the results are fantastic.”