deep learning, feature learning, deep belief nets, convolutional nets
See all
Members (35,276)
Yann LeCun's profile photo
Derek Farren's profile photo
Cesar Pilemyr's profile photo
Kyungmin Park's profile photo
Irina Sedenko's profile photo
Jelili Adebello's profile photo
Ethan Gao's profile photo
Liang-Chieh Chen's profile photo
Srijit Sanyal's profile photo
马剑竹's profile photo
Kermit Lopez's profile photo
Marco Briscolini's profile photo
宮城伸洋's profile photo
iceflame co's profile photo
Jordan Muraskin's profile photo
Nikolai Varankine's profile photo
pawan singh's profile photo
Muhammad Aqib's profile photo
Sri chakra Kumar's profile photo
Bharath bettagere's profile photo
nyemba godlive's profile photo
Hao Huang's profile photo
Arshad Mayet's profile photo
Mouad Jaudar's profile photo

Stream

Join this community to post or comment
 
Emerging new topic: International Workshop on Deep Learning and Music, Call for papers/extended abstracts (in conjunction with IJCNN), Anchorage, Alaska, May 14-19. 
1
Add a comment...

Avani Gadani

Discussion  - 
 
 
New Salk advance uses #deeplearning to analyze #networks, published today in @mitpress http://bit.ly/2kkJnhb #computerscience
2
Add a comment...

Rishab Goel

Discussion  - 
 
Could someone point to good papers for QA using DL?
1
Claude Coulombe's profile photoRishab Goel's profile photo
4 comments
 
+Claude Coulombe and +Patrick Ehlen thanks for the help! :)

Add a comment...

Amr Ahmed

Discussion  - 
 
Siraj Raval asking Udacity Founder & President Sebastian Thrun 67 questions, many about AI!

https://www.youtube.com/watch?v=jfjwXJ6QnsY
5
3
Add a comment...

Mundher Alshabi

Discussion  - 
 
I am much more worried about what people like Hitler or Mussolini might do if they had armies of intelligent robots
1
1
Sarah Rosen (Silencieux)'s profile photoClaude Coulombe's profile photo
2 comments
 
Greetings to Geoffrey Hinton, a great scholar, researcher and a nice guy with ethical concerns. Unfortunately the little kinglets of this world, so full of themselves, do not recognize any rights, treaties or conventions, and much less the United Nation (UN).
Add a comment...

Pip Curtis

Discussion  - 
 
Ofir Nachum works at Google Brain as a Research Resident. His research focusses on sequence-to-sequence models and reinforcement learning. We interviewed Ofir ahead of his session at the Deep Learning Summit in San Francisco, 26-27 January, to find out what started his work in deep learning, what UREX is, the main challenges being addressed in the deep learning space and his prediction for deep learning in the next 5 years.https://www.re-work.co/blog/under-appreciated-reward-exploration-google-brain

The most widely-used exploration methods in reinforcement learning today (like entropy regularization and epsilon-greedy) have not changed much in the last 20 years. Google Brain argues that these exploration strategies are naive and misguided in large action spaces. Google Brain present UREX, a policy gradient algorithm that explores more in areas of high reward.
9
3
Claude Coulombe's profile photo
 
+Pip Curtis Thank you for sharing! A scientific paper about UREX (Under-Appreciated Reward Exploration).:
openreview.net - goo.gl/ZomnVu
Add a comment...

Claude Coulombe

Discussion  - 
 
Arterys’s medical imaging platform is the first FDA approval for a deep learning application to be used in a clinic. Arterys helps physicians understand how a heart is functioning, by providing accurate measurements of the volume of each ventricle.
The FDA approval of a cloud based machine learning application to be used in a clinical setting to help physicians understand how a heart is functioning signals a major breakthrough. Cutting examination time from up to an hour to just 15 seconds paves the way for more AI algorithms in healthcare.
8
3
John Fisher's profile photo
 
Interesting, but very difficult to get the bandwidth in and out of hospitals. disclosure- work for a company that does medical robotics

Add a comment...

Claude Coulombe

Discussion  - 
 
The startup Deepgram has released Kur, a new open source Python DL tool to design, train, and evaluate models without ever needing to code. And the companion GitHub repo:https://goo.gl/b2TLGZ How does Kur compare to other tools like Keras? Although image recognition examples are presented, how well Kur performs outside audio models? That remains to be seen.
25
9
Yagnesh Revar's profile photo
 
I think it's an alternative to NVIDIA DIGITS like tools. 
Add a comment...

Claude Coulombe

Discussion  - 
 
Yet another DL framework... but this time based on regular CPU's. Intel has launched a new open source deep learning framework called BigDL which runs as a Spark job. Modeled after Torch DL library, BigDL was designed to take advantage of hardware acceleration built into Intel's Xeon CPUs. Furthermore, BigDL can load pre-trained Caffe or Torch models. 
BigDL: Distributed Deep Learning Library for Apache Spark
10
4
Benjamin Ellenberger's profile photoClaude Coulombe's profile photo
4 comments
 
+Benjamin Ellenberger Indeed, you are right. Generally, prediction with a trained model is a lot lighter process.
Add a comment...

Claude Coulombe

Discussion  - 
 
On the DL applications scene, Pinterest announces that they use Deep Learning to generate Related Pins, an item-to-item recommendations system that uses collaborative filtering. They built a Pin2Vec embedding representation analog to Word2Vec words embedding and they look up for the nearest Pins by their distances.
One of the most popular ways people find ideas on Pinterest is through Related Pins, an item-to-item recommendations system that uses collaborative filtering. Previously, candidates were generated using board co-occurrence, signals from all the boards a Pin is saved to.
3
Rishab Goel's profile photo
 
Quite interesting approach! :)
Add a comment...

About this community

News about deep learning, feature learning, deep belief networks, convolutional nets, and related topics.

Claude Coulombe

Discussion  - 
 
A note of caution: Libratus IS NOT a Deep Learning software. Nevertheless Tuomas Sandholm and his PhD student Noam Brown from Carnegie Mellon University built a very impressive demonstration of a clever «heuristic brute force» approach using a supercomputer to win against world’s top poker players. More precisely, Libratus is based on a technique called counterfactual regret minimization (CFR) to cope with the 10^160 possible plays of the game. 
1
1
Claude Coulombe's profile photo
 
An article in IEEE Spectrum https://goo.gl/RQHPHj and at CMU https://goo.gl/HWFl9F
Add a comment...
 
Inhaled corticosteroids (ICS) have been used as first line treatment of asthma for many decades. ICS are a form of exogenous glucocorticosteroids that can suppress the endogenous production of glucocorticosteroids, a condition known as adrenal suppression ...
1
1
Condon Freeman Brown's profile photoClaude Coulombe's profile photo
6 comments
 
+Condon Freeman Brown That's OK. It happened to everyone.
Add a comment...

Terry Chen

Discussion  - 
 
 
A TensorFlow implementation of Facebook's Unsupervised Cross-Domain Image Generation.

https://github.com/yunjey/dtn-tensorflow
dtn-tensorflow - domain transfer network. tensorFlow implementation of unsupervised cross-domain image generation
11
5
Claude Coulombe's profile photo
 
The companion scientific paper: https://goo.gl/sGhH5y
Add a comment...

Saikat Basak

Discussion  - 
 
Small (8 layers) neural net delivering >98% accuracy on handwritten digit recognition from the MNIST dataset. A lot of knobs can be tweaked, suggestions are welcomed.

https://github.com/saikatbsk/ml-playground/tree/master/mnist-deep
4
Hossein Hasanpour's profile photoSaikat Basak's profile photo
3 comments
 
Thanks for the suggestions. I'll definitely try conv nets and more datasets, CIFAR10 and notMNIST.
Add a comment...

Kilian Mie

Discussion  - 
 
How to decide deep net architectures?

For binary classification deep nets (not CNNs), what are best practices to decide
A) how many layers the deep net should have?
B) how many hidden units per layer one should use?
C) if each layer should have the same number of hidden units or not?

Brute force cross-validation various net architectures with dropout kinda works but is costly/time-consuming.

Are there any guidelines or paper references on this topic?
Ideally, A/B/C above as a function of number of samples and number of features?
6
Claude Coulombe's profile photoKilian Mie's profile photo
9 comments
 
+Claude Coulombe, +Dominik Andreas -- thanks again to both of you!
I will study up and experiment with the idea to start really small to optimize things like dropout/activation functions, and similar
Add a comment...

Claude Coulombe

Discussion  - 
 
Funny Friday! Googles's co-founder, Sergey Brin said that when he was in charge of Alphabet’s research group he didn’t pay much attention to a team working on AI. Now the technology touches nearly every piece of Google's business.
Alphabet Inc. President Sergey Brin discusses advancements in artificial intelligence at the World Economic Forum in Davos, Switzerland. (Source: Bloomberg)
14
3
Add a comment...

Claude Coulombe

Discussion  - 
 
Following its acquisition of Montréal-based Maluuba, Micro$oft announces additional investment in the city which is becoming an AI hub.
3
Add a comment...

Matt Kuenzel

Discussion  - 
 
Neural Programmer-Interpreters
Scott Reed, Nando de Freitas

OK, they "trained" this network by showing it the desired step-by-step behavior. Why is it impressive that it learns to repeat those steps?
Abstract: We propose the neural programmer-interpreter (NPI): a recurrent and compositional neural network that learns to represent and execute programs. NPI has three learnable components: a task-agnostic recurrent core, a persistent key-value program memory, and domain-specific encoders that ...
1
Claude Coulombe's profile photo
 
Particularly they've done that work with very few training observations / data samples.

Add a comment...

Claude Coulombe

Discussion  - 
 
Ten applications of Deep Learning and AI in use today by R.L. Adams in the Forbes magazine: Siri, Alexa, Tesla, Cogito, Boxever, John Paul, Amazon, Netflix, Pandora and Nest. Ok, mixing businesses and products.. ;-(
There are many examples of artificial intelligence being used today to enhance and improve our lives, but these are some of the most potent applications of A.I.
8
8
Add a comment...