Profile

Cover photo
Rupesh Kumar Srivastava
Works at IDSIA
Attends Università della Svizzera italiana
Lives in Switzerland
635 followers|118,357 views
AboutPostsPhotos

Stream

 
 
PostDoc Jobs 2016: Join the Deep Learning team (since 1991) that won more competitions than any other. We are seeking postdocs for the project RNNAIssance based on this tech report on "learning to think:” http://arxiv.org/abs/1511.09249 . The project is about general purpose artificial intelligence for agents living in partially observable environments, controlled by reinforcement learning recurrent neural networks (RNNs), supported by unsupervised predictive RNN world models. Location: The Swiss AI Lab, IDSIA, in Switzerland, the world’s leading science nation, and most competitive country for the 7th year in a row. Competitive Swiss salary. Preferred start: As soon as possible. INTERVIEWS: Berlin (April 29 - May 1), NYC (May 2-7), Beijing (May 12-13), London (May 17-20), or in Switzerland, or by video. More details and instructions can be found here: http://people.idsia.ch/~juergen/rnnai2016.html

#artificialintelligence
#deeplearning
#machinelearning
#computervision
4 comments on original post
9
1
Rupesh Kumar Srivastava's profile photoDeepjoy Das's profile photo
4 comments
 
+Rupesh Kumar Srivastava  Thanks so much for the info
Add a comment...

Rupesh Kumar Srivastava

Shared publicly  - 
 
 
Interesting Application of LSTM

http://arxiv.org/abs/1601.08188
Abstract: Lipreading, i.e. speech recognition from visual-only recordings of a speaker's face, can be achieved with a processing pipeline based solely on neural networks, yielding significantly better accuracy than conventional methods. Feed-forward and recurrent neural network layers (namely ...
1 comment on original post
1
Add a comment...

Rupesh Kumar Srivastava

Shared publicly  - 
 
New work from Edinburgh on speech recognition with highway networks reduces model sizes and explores the effects of the gates: http://arxiv.org/abs/1512.04280
Abstract: For speech recognition, deep neural network (DNN) have significantly improved the recognition accuracy in most of benchmark datasets and application domains. However, compared to the conventional Gaussian mixture models(GMMs), DNN-based acoustic models usually have much larger number ...
1
Add a comment...

Rupesh Kumar Srivastava

Shared publicly  - 
 
Again and again, signs that Nature needs to reevaluate what it is.
 
On Dark Matter And Dinosaurs

Let me begin by saying there is no evidence that dark matter killed the dinosaurs. None whatsoever. Unfortunately the idea was posted on Nature’s blog, and from there it went to Scientific American and elsewhere. The various social media took the story and it has spread like a prairie wildfire. The actual preprint is much less sensational (and doesn’t mention dinosaurs) but it is still very speculative.

The idea comes from the fact that the Sun does not follow a flat orbit around the galaxy. Instead, its motion wobbles above and below the galactic plane, crossing the galactic plane every 35 million years. This isn’t unusual, as lots of stars follow similar paths, but it has led some to speculate that perhaps this periodicity could explain periodic mass extinctions in the geologic record.

The problem is, there isn’t any strong evidence for cyclic mass extinctions. Some analysis of the data has hinted at a pattern, but the correlation isn’t very strong. Of course that hasn’t stopped people from proposing everything from companion stars to Nibiru to explain these periodic extinctions. There been similar proposals that every time the Sun crosses the galactic plane the Oort cloud would be disrupted, causing comets to sweep into the inner solar system and bombard the Earth.

What’s new here is that the authors propose that dark matter within the plane of the galaxy is doing the disrupting. As I wrote about last week, there is a hint of dark matter seen in gamma ray observations of the center of our galaxy. One model that could account for these gamma rays is type of dark matter that would lie within the galactic plane. So if this type of dark matter exists, and if it disrupts the Oort cloud when the Sun crosses the galactic plane, and if that caused comets to fling into the inner solar system and bombard the Earth, and if that bombardment caused periodic mass extinctions, then you should see some evidence in the geologic record.

So what evidence is there? None. Well, not quite none. If you assume the model is true, and then look for a periodicity in the cratering record of Earth, you find that the cratering record agrees with the model about three times better that it agrees with random cratering. Scientifically, that isn’t very convincing data. It makes for a mildly interesting paper, but it’s mostly speculation at this point.

But Nature and several other websites have decided to take this speculative idea, add the word dinosaurs to the title, and imply that scientists are proposing dark matter killed the dinosaurs. No one is proposing that. It’s link-bait noise that makes the job of communicating real science all that more difficult. So if you see one of these sensationalized titles, don’t share it on social media. Tell your friends that share the articles that it’s speculative nonsense. Hopefully we can drown this noise and get back to real science.

Because honestly, science is interesting enough without the hype.

Paper: Lisa Randall, Matthew Reece. Dark Matter as a Trigger for Periodic Comet Impacts. arXiv:1403.0576 [astro-ph.GA] (2014).
There is no evidence that dark matter killed the dinosaurs. None whatsoever. It's link-bait noise that makes the job of communicating real science all that more difficult.
21 comments on original post
1
Add a comment...
 
Our new deep learning library is written completely in Python, but without Theano, if that's more your style :)
 
Announcing Brainstorm Open Source Software for Neural Networks

We are open-sourcing a new neural networks library called Brainstorm, developed over the past year at the Swiss AI Lab IDSIA by PhD students Klaus Greff and Rupesh Srivastava: https://github.com/IDSIA/brainstorm

Brainstorm is designed to make neural networks fast, flexible and fun. Lessons learned from earlier open source projects led to new design elements compatible with multiple platforms and computing backends.

Brainstorm already has a robust base feature set, including support for recurrent neural networks (RNNs) such as LSTM, Clockwork, 2D Convolution/Pooling and Highway layers on CPU/GPU. All data is considered sequential, and RNNs are first-class citizens.

We hope the community will help us to further improve Brainstorm.

#machinelearning
#artificialintelligence
#computervision
#deeplearning

http://people.idsia.ch/~juergen/brainstorm.html
4 comments on original post
15
9
Add a comment...

Rupesh Kumar Srivastava

Shared publicly  - 
 
LSTM+CTC will soon be used by hundreds of millions if not billions of devices. Google has made it more efficient and robust to work for voice search too (it already improved offline voice recognition by large margins).

In a 2003 talk on LSTMs, +Juergen Schmidhuber said "Speech (vs HMMs)? One should try it ..." 3 years later, the LSTM+CTC paper (Graves et al.) cited in this post was published at ICML.

The scientific impact of LSTMs is already huge, but I feel that there is much more yet to come.
 
Leaner. Faster. More robust.

Today, we’re happy to announce that we’ve launched improved neural network acoustic models for voice searches and commands in the Google app (on Android and iOS), and for dictation on Android devices. 

Using Connectionist Temporal Classification and sequence discriminative training techniques, these models are a special extension of recurrent neural networks that use much less computational resources, are more accurate, robust to noise, and faster to respond to voice search queries.

Check out the Google Research blog below to learn more. Happy (voice) searching!
16 comments on original post
2
Add a comment...
In his circles
413 people
Have him in circles
635 people
Swiss Antic's profile photo
Anmeda Tony's profile photo
Neeti Birla's profile photo
Abhishek Pathak's profile photo
Ben Bogart's profile photo
sonii minii's profile photo
Ansuman Pradhan's profile photo
Kayla Kate's profile photo
oladipo kayode's profile photo

Rupesh Kumar Srivastava

Shared publicly  - 
 
 
The groups of +Juergen Schmidhuber, Luca Gambardella and myself have joined forces to present the first work using Deep Neural Networks to enable an autonomous vision-control drone to recognize and follow forest trails. The video is narrated, so turn on your loud speakers and enjoy it! More info on the DNN training and testing in the FAQs below. Btw, the paper is currently nominated for the AAAI Best Video Award; the video with the most likes on YouTube wins; so, if you like it, please give us a thumb-up on YouTube!

Paper:
A. Giusti et al., A Machine Learning Approach to Visual Perception of Forest Trails for Mobile Robots, IEEE Robotics and Automation Letters, 2016.
PDF: http://rpg.ifi.uzh.ch/docs/RAL16_Giusti.pdf
Project webpage and datasets: http://www.leet.it/home/giusti/website/doku.php?id=wiki:forest

FAQs:

What is the paper about?
We present the first work using a Deep Neural Networks (DNNs) image classifier running onboard our vision-controlled drone to recognize and autonomously follow forest trails. Unlike previous works, which relied on image salience or low-level features, our DNN-based image classifier operates directly on pixel-level image intensities and outputs the direction of the trail with respect to the heading direction of the drone. If a trail is visible, the software steers the drone in the corresponding direction.

How did we train the classifier?
In order to gather enough data to train our DNN classifier, we hiked several hours along different trails in the Swiss Alps and took more than 20 thousand images of trails using cameras attached to a helmet (Fig. 4 in the paper). This effort paid off: when tested on a new, previously-unseen trail, the DNN was able to find the correct direction in 85% of cases; in comparison, humans faced with the same task guessed correctly 82% of the time.

Real time and onboard?
Yes. The classifier ran in real time and onboard the smartphone processor (Odroid quadcore computer) on our custom-made vision-controlled quadrotor. Both visual odometry (based on SVO) and control were also running onboard.

Why do we want drones to follow forest trails?
To save lives. Every year hundreds of thousand people get lost in the wild worldwide. In Switzerland alone, around 1000 emergency calls per year come from hikers, most of whom are injured or have lost their way. Drones are an efficient complement to human rescuers and can be deployed in large numbers, are inexpensive and prompt, and thus minimize the response time and the risk of injury for those who are lost and those who work in rescue teams.

Is the training and testing data available for research?
Yes, from the project webpage.

More on Deep Learning: http://www.scholarpedia.org/article/Deep_Learning

#computervision
#deeplearning
#machinelearning
#artificialintelligence
#robotics
#drones

https://youtu.be/umRdt3zGgpU
View original post
1
Add a comment...

Rupesh Kumar Srivastava

Shared publicly  - 
 
"Okay Google" now with LSTMs
 
Take a highly engineered state-of-the-art system, strip it down, replace it with a LSTM, and train it using a loss that matches your problem. Does that recipe sound familiar?
Venue. International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE (2016). Publication Year. 2016. Authors. Georg Heigold, Ignacio Moreno, Samy Bengio, Noam M. Shazeer. BibTeX. @inproceedings{44681, title = {End-to-End Text-Dependent Speaker Verification}, author = {Georg ...
3 comments on original post
1
1
Add a comment...

Rupesh Kumar Srivastava

Shared publicly  - 
 
 
How to Learn an Algorithm (video). I review 3 decades of our research on both gradient-based and more general problem solvers that search the space of algorithms running on general purpose computers with internal memory. Architectures include traditional computers, Turing machines, recurrent neural networks, fast weight networks, stack machines, and others. Some of our algorithm searchers are based on algorithmic information theory and are optimal in asymptotic or other senses. Most can learn to direct internal and external spotlights of attention. Some of them are self-referential and can even learn the learning algorithm itself (recursive self-improvement). Without a teacher, some of them can reinforcement-learn to solve very deep algorithmic problems (involving billions of steps) infeasible for more recent memory-based deep learners. And algorithms learned by our Long Short-Term Memory recurrent networks defined the state-of-the-art in handwriting recognition, speech recognition, natural language processing, machine translation, image caption generation, etc. Google and other companies made them available to over a billion users.

The video was taped on Oct 7 2015 during MICCAI 2015 at the Deep Learning Meetup Munich:  http://www.meetup.com/en/deeplearning/events/225423302/  Link to video: https://www.youtube.com/watch?v=mF5-tr7qAF4

Similar talk at the Deep Learning London Meetup of Nov 4 2015: http://www.meetup.com/Deep-Learning-London/events/225841989/ (video not quite ready yet)

Most of the slides for these talks are here: http://people.idsia.ch/~juergen/deep2015white.pdf

These also includes slides for the AGI keynote in Berlin http://agi-conf.org/2015/keynotes/, the IEEE distinguished lecture in Seattle (Microsoft Research, Amazon), the INNS BigData plenary talk in San Francisco, the keynote for the Swiss eHealth summit, two MICCAI 2015 workshops, and a recent talk for CERN (some of the above were videotaped as well).

Parts of these talks (and some of the slides) are also relevant for upcoming talks in the NYC area (Dec 4-6 and 13-16) and at NIPS workshops in Montreal:

1. Reasoning, Attention, Memory (RAM) Workshop, NIPS 2015 https://research.facebook.com/pages/764602597000662/reasoning-attention-memory-ram-nips-workshop-2015/

2. Deep Reinforcement Learning Workshop, NIPS 2015 http://rll.berkeley.edu/deeprlworkshop/

3. Applying (machine) Learning to Experimental Physics (ALEPH) Workshop, NIPS 2015 http://yandexdataschool.github.io/aleph2015/pages/keynote-speakers.html

More videos: http://people.idsia.ch/~juergen/videos.html

Also available now: Scholarpedia article on Deep Learning: http://www.scholarpedia.org/article/Deep_Learning

Finally, a recent arXiv preprint: On Learning to Think: Algorithmic Information Theory for Novel Combinations of Reinforcement Learning Controllers and Recurrent Neural World Models. http://arxiv.org/abs/1511.09249

#machinelearning
#artificialintelligence
#computervision
#deeplearning

2 comments on original post
1
Add a comment...

Rupesh Kumar Srivastava

Shared publicly  - 
 
Researchers from MIT, JHU and MSR have used highway networks to improve distant speech recognition: http://arxiv.org/abs/1510.08983
More info on highway networks: http://people.idsia.ch/~rupesh/very_deep_learning/
Abstract: In this paper, we extend the deep long short-term memory (DLSTM) recurrent neural networks by introducing gated direct connections between memory cells in adjacent layers. These direct links, called highway connections, enable unimpeded information flow across different layers and thus ...
1
Add a comment...

Rupesh Kumar Srivastava

Shared publicly  - 
 
I found that +Hugo Larochelle wrote up these notes on our paper LSTM: A Search Space Odyssey.

We spent countless hours planning, debating and improving the experimental setup in this paper and writing it, so it's very nice to see it being appreciated :)
We hope that apart from serving as a nice reading about LSTMs, it also helps researchers think about doing their comparative analyses right. This is one area in which recent deep learning papers typically don't do well.

Our paper is here: http://arxiv.org/abs/1503.04069
3
Add a comment...

Rupesh Kumar Srivastava

Shared publicly  - 
 
 
Loophole free Bell experiment
Abstract: For more than 80 years, the counterintuitive predictions of quantum theory have stimulated debate about the nature of reality. In his seminal work, John Bell proved that no theory of nature that obeys locality and realism can reproduce all the predictions of quantum theory.
1 comment on original post
1
Add a comment...
People
In his circles
413 people
Have him in circles
635 people
Swiss Antic's profile photo
Anmeda Tony's profile photo
Neeti Birla's profile photo
Abhishek Pathak's profile photo
Ben Bogart's profile photo
sonii minii's profile photo
Ansuman Pradhan's profile photo
Kayla Kate's profile photo
oladipo kayode's profile photo
Work
Occupation
PhD researcher
Employment
  • IDSIA
    PhD Student, 2011 - present
  • Microsoft Research
    Summer Intern, 2014
  • Waseda University
    Summer Research Intern, 2009
Places
Map of the places this user has livedMap of the places this user has livedMap of the places this user has lived
Currently
Switzerland
Story
Introduction
I grew up in the city of Lucknow in Northern India. I am currently a PhD student in Artificial Intelligence at IDSIA. Before this, I studied Mechanical Engineering at the Indian Institute of Technology, Kanpur. I am interested in machine learning, robotics, philosophy, psychology and music, academically and otherwise.
Education
  • Università della Svizzera italiana
    2012 - present
  • Indian Institute of Technology Kanpur
    Mechanical Engineering, 2006 - 2011
  • Saint John's School, Lucknow
    1990 - 1994
  • Saint Dominic Savio College, Lucknow
    1994 - 2005
Basic Information
Gender
Male