Profile

Cover photo
Claude Coulombe
Works at Université de Montréal - MATI
123,630 views
AboutPostsPhotosYouTube

Stream

Claude Coulombe

Discussion  - 
 
An appeal to all... I am looking for a scientific paper which would explain « It's well known or established that a whole class of problems cannot be solved using actual deep learning techniques due to a lack of data. »  Something like « not enough data => no deep learning ».  I thought I read that somewhere but but despite my efforts Googling it again, I can't locate the related paper or blog post.

That said, I know that it depends on the problem... I'm also aware of the rule of thumb « one parameter => one training sample ». But there is much differences between few thousands training data samples and billions. 

In fact, I'm looking a paper which explains limits of actual deep learning approach beyond brute force and huge amount of data. It could also help to fix some limits or rule of thumb to know if deep learning is appropriate or you should use another more classical learning techniques with features engineering. Or maybe, trying « data augmentation » or put more regularization to overcome eventual overfitting.

Sometimes, deep learning approaches work even with small amount of data. I think of some Kaggle competitions won by deep learning techniques. For instance the one on molecules with Merck, which had only several thousand data training samples. But I think that the particular architecture had a reduction effect.

Finally, it seems obvious, but I would need an academic reference that the human brain operates from very few samples. Maybe the brain is sharing data or models with other brain regions, so the same data or models would be used for several different purposes. 
6
Michele Pratusevich's profile photoJohn Taylor's profile photoClaude Coulombe's profile photo
13 comments
 
+John Taylor Right! I've just read something about « transfer learning »... Yes indeed it's a great idea. That could be seen as prior knowledge. Maybe something similar could explain the impression we have that brain learns with only few examples. 

But unless Google or another DL's mammouth is willing to open source their models, you still need a huge amount of data to build an operational deep learning system.  

I agree, there is no rule of thumb to cover all the cases in a rapidly evolving field nor an algorithm better than all others on all datasets (no free lunch ?). But a warning could be valuable in order to avoid that people who have no access to huge datasets taking too much risk. 

Anyway, better to have huge amount of data, read « millions » of training samples to back you, before going on off-road tracks of the Deep Learning. As you said, Deep Learning should be regarded as still being at the research stage. 
Add a comment...
 
Un article de Paul Krill dans Infoworld qui explique l'évolution de GWT 3.0  (Google Web Toolkit) annoncée par Ray Cromwell de Google en marge de la conférence GWT-create. Cette évolution vers le support de ECMAScript 6, de Elemental 2.0 et de Java 8 entraînera des problèmes de non-compatibilité avec les version plus anciennes de GWT. Cela dit une branche GWT 2.0 sera toutefois maintenue. #gwt   #singlepageapps   #java   #gwtcreate  
 ·  Translate
Two new versions of Google Web Toolkit are due, featuring support for Java 8 but sacrificing backward compatibility
1
Add a comment...

Claude Coulombe

Shared publicly  - 
 
Trois excellentes actions suggérées par le blogueur et journaliste Martin Lessard pour augmenter l'attrait de la langue française. Je constate que les CLOM en font partie! #clom   #moocfr   #mooc  
 ·  Translate
 
Mes 3 actions pour augmenter l’attrait de la francophonie : traduire Wikipédia, offrir des MOOCs, libérer les données.

Au boulot!  
 ·  Translate
Lorsqu'on parle d'actions, Wikipédia, les Moocs et les données ouvertes semblent un territoire facile à conquérir.
View original post
2
Add a comment...

Claude Coulombe

Shared publicly  - 
 
Un article intéressant de James G. Mazoue sur les tendances au-delà du modèle actuel des CLOM / MOOC: les CLOM certifiants, les CLOM à base de compétences, la formalisation des contenus et la réforme des institutions. #clom   #moocfr   #mooc  
 ·  Translate
1
Add a comment...

Claude Coulombe

Shared publicly  - 
 
Un billet de David Glance de la University of Western Australia sur la « science de l'abandon » qui tente d'expliquer le décrochage dans les CLOM (MOOC), les gymnases et les diètes. Tous ces abandons présentent d'intéressantes similarités. #clom   #moocfr   #mooc  
 ·  Translate
One of the characteristic features of Massive Open Online Courses is the observation that no matter how many students enroll in a course, only between 5 to 10% of them will ever complete it. Setting aside…
1
Add a comment...

Claude Coulombe

Shared publicly  - 
 
Manifestation pour le climat à Montréal - 21sept2014 - Parc Jeanne-Mance. Nous cherchons les autres manifestants... Ça nous apprendra à prendre des raccourcis. ;(
 ·  Translate
3
André Cotte's profile photo
 
T'as fini par trouver, j'espère! 
 ·  Translate
Add a comment...
 
An article by Paul Krill in Infoworld which explains the future evolution of GWT 3.0 (Google Web Toolkit) announced by Ray Cromwell from Google at the GWT-create conference last week. This trend towards support for ECMAScript 6, Elemental 2.0 and Java 8 will cause problems of backward compatibility with older versions of GWT. That said a GWT 2.0 branch will however be maintained. #gwt   #gwtcreate   #singlepageapps   #java  
Two new versions of Google Web Toolkit are due, featuring support for Java 8 but sacrificing backward compatibility
2
Add a comment...
 
Première image panoramique prise par l'instrument CIVA de l'atterrisseur Philae depuis la surface de la comète Churyumov-Gerasimenko. En bas au tiers de largeur de l'image on remarque un pied de Philae.
Rosetta’s lander Philae is safely on the surface of Comet 67P/Churyumov-Gerasimenko, as these first two CIVA images confirm. One of the lander’s three feet can be seen in the foreground. The image is a two-image mosaic. The full panoramic from CIVA will be delivered in this afternoon’s pr…
1
Add a comment...
 
An interesting article by James G. Mazoue about the trends beyond the current MOOC model: MOOC-based degrees, competency-based teaching, formalization of teaching content and institutional / regulatory reform. #mooc  
1
Add a comment...

Claude Coulombe

Shared publicly  - 
 
A post from David Glance, University of Western Australia, on the  « Science of why we give up » that attempts to explain the drops in MOOCs, gyms and diets. All these dropout's processes have interesting similarities. #mooc  
One of the characteristic features of Massive Open Online Courses is the observation that no matter how many students enroll in a course, only between 5 to 10% of them will ever complete it. Setting aside…
1
Add a comment...

Claude Coulombe

Shared publicly  - 
1
Add a comment...
 
Allons tous marcher pour le climat!  Dimanche, le 21 septembre, 13h00, au Parc Lafontaine à Montréal. Cela promet d'être la plus grande mobilisation de l'histoire contre les changements climatiques.
 ·  Translate
Help us turn more people out for the People's Climate March!
1
Add a comment...
Work
Employment
  • Université de Montréal - MATI
    present
Links
Contributor to
Basic Information
Gender
Male