NIPS 2014 - highlights

This year, NIPS is in my home town Montréal!
Here is the highlights https://nips.cc/Conferences/2014/Program/

Most interesting invited talks:

John Carlos Baez

University of California, Riverside

Networks in Climate Science

Cynthia Dwork

Microsoft Research

Privacy in the Land of Plenty

Arunava Majumdar

Google Inc. & Stanford University

Role of Coupled Networks in 21st Century Energy Infrastructure

Yurii Nesterov

Catholic University of Louvain (UCL)

Subgradient Methods for Huge-Scale Optimization Problems

Posner Lectures

Michael Kearns

University of Pennsylvania

Games, Networks, and People

John J. Hopfield

Princeton University

Using the Emergent Dynamics of Attractor Networks for Computation

Friday Dec 12 highlights workshops

• Personalization: Methods and Applications
• Deep Learning and Representation Learning
• Modern Machine Learning and Natural Language Processing
• Learning Semantics

Saturday Dec 13 Workshop

1. Human Propelled Machine Learning
2. Advances in Variational Inference
3. 4th Workshop on Automated Knowledge Base Construction (AKBC)

-------------

more details:

Workshop Friday 12 dec : 8:30am – 6:30pm Friday, December 12, 2014

8:30am – 6:30pm	Yoshua Bengio, Adam Coates, Roland Memisevic, Andrew Y Ng Deep Learning and Representation Learning
8:30am – 6:30pm	Reza Bosagh Zadeh, Ion Stoica, Ameet S Talwalkar Distributed Machine Learning and Matrix Computations

8:30am – 6:30pm

Cedric Archambeau, Antoine Bordes, Leon Bottou, Chris J Burges, David Grangier Learning Semantics

8:30am – 6:30pm

Ankur P Parikh, Avneesh Saluja, Chris Dyer, Eric P Xing Modern Machine Learning and Natural Language Processing

8:30am – 6:30pm

Khalid El-Arini, Yisong Yue, Dilan Gorur Personalization: Methods and Applications

 Workshop Saturday 13 dec : 8:30am – 6:30pm Friday, December 12, 2014

8:30am – 6:30pm	Shakir Mohamed, Tamara Broderick, Charles Blundell, Matthew D Hoffman,David Blei, Michael I Jordan Advances in Variational Inference
8:30am – 6:30pm	Shivani Agarwal, Hossein Azari Soufiani, Guy Bresler, Sewoong Oh, David C Parkes, Arun Rajkumar, Devavrat Shah Analysis of Rank Data: Confluence of Social Choice, Operations Research, and Machine Learning
8:30am – 6:30pm	Richard Baraniuk, Michael C Mozer, Divyanshu Vats, Christoph Studer, Andrew E Waters, Andrew Lan Human Propelled Machine Learning
8:30am – 6:30pm	David S Choi, Aaron Clauset, Edo M Airoldi, Leto Peel, Johan Ugander Networks: From Graphs to Rich Data

Conference other interesting talks:

Tuesday Dec 9th

4:15 – 5:40pm	Oral Session 4Session Chair: Ranzato Marc'Aurelio
	·         Y. Tang, N. Srivastava, R. Salakhutdinov Learning Generative Models with Visual Attention ·           o    J. Yosinski, J. Clune, Y. Bengio, H. Lipson How transferable are features in deep neural networks?

Wednesday Dec 10th

2:00 – 3:30pm	Oral Session 7Session Chair: Geoffrey J Gordon
	·         Posner LectureJohn J. Hopfield Using the Emergent Dynamics of Attractor Networks for Computation

Agile tour 2014 highlights/notes

I went to my first "agile tour" conference. Why: I want to hear Jean-Marc De Jonghe (transform management to succeed an agile transformation) and David Hussman (Lessons in gravity: what keeps you down?).

Here are my highlights notes (yes its notes):
Jean-Marc De Donghe:

• Wall was coming
• slow decline of subscribers (~45 mins)
• increase web (5-15 mins) -> less revenues
• A good hockey player plays where the puck is. A great hockey player plays where the puck is going to be. - Wayne Gretzky- 
• smarthphone web traffic: 86% apps, 14% browser
• experience TV->laptop->tablette->cell (intimité/privacy)
• Agile vs old school
• context vs control
• confidence vs fear
• context and vision
• recruiting=attitude, aptitude, diversité
• agilité bidirectionnel (good luck)
• build trust = quick win greedy approach
• protect your team
• Speed = disorder (tech debt); 
• Environment who  promote agile
• No split by job
• refactoring = speed 
• Estimate added value to set priorities

David Hussman:

• gravity sucks!
• launch rocket -> Egyptian approach = much easier 
• uncertain/overly certain Product(uncertain) /techno (certain)
• y = product learning, x = structure complexity
• Accidental complexity-> incidental complexity 
• mass has consequence, process mass, techno mass, mass of certainty, meeting mass
• http://devjam.com/ + nonban (least process+most measurable value)
• better conversation not stories
• diff between what they need and what you think they need
• gource mine craft (visualize code)
• point = size of effort (useless) size of value
• book: anti fragile (the black swan) + founder at work
• investment discovery vs delivery
• valuable vs feasible

Others: (great speakers -> Michel Céré, Richard Martin)

• value not effort
• Self manage team = team -> scrum master -> boss
• no hierarchy + scrum master = authority 
• People value more the process then the result
• influence silencieuse de l'autorité (manager keep forgetting)
• scrum master should do their work
• propose: auto organize or hierarchy request (its a choice)
• team should force people to take off is required
• culture  = fondement ordre social
• valeurs vécus+artefacts+presuposé (macro culture, sous culture, micro culture)
• Changement culturel & objectif d'affaire
• Culture des opérations (collaboration/communication), culture conception/ingénierie (idea, système élégeant), culture haute direction (finance, image entreprise et de soi)
• Importance of rituals
• voir openagileadoption
• Dynamisme/cohérence globale -> reponse pertinente
• agile: empowering groups+cadre execution aligné+ cible=client
• WIP/JIT/Velocity (work in progress, just in time...)
• TRG = taux de rendement global: toyota = 86% + bureau = 17%
• Leadership mobiliser
• QIX = Qty I.... Max
• Volonté stratégique
  
  

balance between writting tools and research

Do do research, we need tools but need to find the right balance between writing the tool we need and doing the research. Context switching between the 2 isn't instantaneous and I suspect its because it use very different skill-sets and brain region. 

According to me, several reason lead researcher to reinvent the wheel:

• easier + faster at start + no maintenance required vs reuse = harder at first but more productive in the long run (trade-off fast/reusable/maintenance)
• Not invented here culture
• Skill sets distance (distance between coding (programmer) and research is less important than leveraging code (engineering) and research)

In order to get a better balance between writing tools and doing research, I would like to better leverage existing tools to not reinvent the wheel like:

• algebird -> twitter/algebird · GitHub (Abstract Algebra for Scala running on hadoop)
• Vertica (HP) -> Analytics platform (fast and easy data views/pivots)
• Apache Drill(Self-service data exploration on hadoop)
• http://pandas.pydata.org/ high-performance, easy-to-use data structures and data analysis tools for the Python programming language.

1) Algebird is an abstract algebra library for Scala developed at Twitter and released under the ASL 2.0 license and run on hadoop. It has support for algebraic structures such as semigroups, monoids, groups, rings and fields as well as the standard functional things like monads. More interestingly though are the probabilistic data structures and the accompanying monoids that come out of the box. (Big Data )

2) HP Vertica Analytics platform

Big Data Analytics—No Limits, No Compromises

• Live Aggregate projections
• Open Architecture
• Blazing – Fast Analytics
• Massive Scalability

3) Apache Drill:  Apache Drill is an open source, low latency SQL query engine for Hadoop and NoSQL.

Apache Drill - Self-Service Data Exploration

-------------------------------

MAKE YOUR APPS AND YOUR BUSINESS SMARTER

 Learn how to use Prediction APIs and make Machine Learning work for you — without hiring an expert.

http://www.louisdorard.com/machine-learning-book/