From Lille, France: the International Conference on Machine Learning. Plus a short note on the new IBM “superchip”

Home / Uncategorized / From Lille, France: the International Conference on Machine Learning. Plus a short note on the new IBM “superchip”

ICML 2015

11 July 2015 – One of the benefits of living and commuting between Brussels and Paris is that you are a stone’s throw (well, in close proximity) to all of the major, “funky” technology events such as IBM Research/Analytics (Paris FR and Zurich CH), IBM Cyber Security (Ghent BE), Cannes Lions (Cannes FR), and the Federal Institute of Technology (ETHZ) Artificial Intelligence events (Zurich CH).

This past week I was able to enjoy a few days at the International Conference on Machine Learning in Lille, France. It is the premier annual Machine Learning (ML) event supported by the International Machine Learning Society which is a leader in ML research.

NOTE: for an excellent “gentle” introduction to ML via Youtube:

A Gentle Introduction of Machine Learning

 

Google (no surprise) had a platoon of its centurions at the Lille event and there were many opportunities to speak with Googlers. Google displayed a range of ML expertise. Some of it is a typical data crunching exercise like reducing internal covariate shift to research that provides a glimpse of what Google will do with its knowledge. And it has really got into the “game thing” (umm, that’s “game revenue”, folks).

Google hosted scores of workshops. But Adobe, Disney Research, Faceboook, IBM, Microsoft Research, and many more ML’s major players were all there.

And the Lille event was typical for all of these tech events: free (very good) Wifi access, and fabulous lunch bars on site to keep you near all the action. Plus scores of lunch/dinner options within a 5-10 minute walk.

So it was time to grab/download the scores of white papers before you end up having to buy them when the research surfaces in assorted journals.
The Lille event plus the IBM events and the ETHZ events all provide fascinating insights into what is happening with ML and data analytics, especially with some pointed notes abut legal text analytics.

A few random points:

  1. As computers better understand natural language, new frontiers have opened up everywhere: improved user-interfaces for applications, better search engines, better personal assistants like Cortana and Siri, plus tools for figuring out what a given document is really about. For instance, I saw how a news website can enable a more engaging experience for its users, how individuals mentioned in articles can be algorithmically linked to Wikipedia and other databases, etc.
  1. One Google and one ETHZ researcher noted that applications of ML in the law (a spirited discussion of TAR included) have come a long way and yes, legal practice requires advanced cognitive abilities. But higher-order cognition does not remain outside the capability of current ML technology and ML technology is being advanced (employing heuristics, pattern recognition, etc.) to find hidden relationships in legal documents and data, and improve the automated organization of documents. There was also much discussion about developments in predicting the outcomes of legal cases.
  1. Real and virtual are crashing together. On one side is hardware that acts like software: IP-addressable, programmable with high-level procedural languages and APIs, able to be stitched into loosely coupled systems. These are “the mashups of a new era” said one presenter. On the other is software that’s newly capable of dealing with the complex subtleties of the physical world — ingesting huge amounts of data, learning from it, and making decisions in real time. The result is an entirely new medium that’s just beginning to emerge.
  1. The new medium is something broader than terms like “Internet of Things,” “Industrial Internet,” or “connected devices” suggest. It’s an entirely new discipline that’s being built by software developers, roboticists, manufacturers, hardware engineers, artists, and designers. Said one presenter: “Ten years ago, building something as simple as a networked thermometer required some understanding of electrical engineering. Now, it’s a Saturday afternoon project for a beginner”.
  2. For years, the cloud was expected to become the dominant model for applications, overtaking traditional data centers, but … well, this just isn’t so. Instead, the cloud has become just one part of a larger system, which comprises private data centers as well as public and private clouds. A lot of this evolution has to do with the development progress: public clouds are just easier to scale, build, and tear down. But private data centers are cheaper to run. So, we’re seeing this convergence of applications that are built to live across these spaces: “It’s a continuum and people are in different places on that continuum even with different applications in the same portfolio of applications in the same enterprise IT shop.”
  3. And some interesting notes from the IBM Analytics conference and the ETHZ conference, and Cannes Lions:
  • Much has been written about how the New York Genome Center and IBM have partnered in a first-of-a-kind program to accelerate the race to personalized, life-saving treatment for cancer patients based on genome analysis. But I learned the computing resources needed to handle genome data will soon exceed those of Twitter and YouTube, that the discipline is not geared up to cope with the coming genomics flood. Nature had a recent article highlighting a study on the subject. By 2025, between 100 million and 2 billion human genomes could have been sequenced according to the article and the data-storage demands for this alone could run to as much as 2-40 exabytes (1 exabyte is 1018 bytes), because the number of data that must be stored for a single genome are 30 times larger than the size of the genome itself, to make up for errors incurred during sequencing and preliminary analysis.
  • This outstrips YouTube’s projected annual storage needs of 1-2 exabytes of video by 2025 and Twitter’s projected 1-17 petabytes per year (1 petabyte is 1015 bytes). It even exceeds the 1 exabyte per year projected for what will be the world’s largest astronomy project, the Square Kilometre Array, to be sited in South Africa and Australia. But storage is only a small part of the problem. The study argues that computing requirements for acquiring, distributing and analysing genomics data may be even more demanding.
  • At Cannes Lions I learned that some of the analysis is a bit off because it does not address how data is handled. For instance, the study underestimates the processing and analysis aspects of the video and text data collected and distributed by Twitter and YouTube, such as advertisement targeting and serving videos to diverse formats.
  • But … genomics will have to address the fundamental question of how much data it should generate. The world has a limited capacity for data collection and analysis, and it should be used well. Because of the accessibility of sequencing, the explosive growth of the community has occurred in a largely decentralized fashion, which can’t easily address questions like this. Other resource-intensive disciplines, such as high-energy physics, are more centralized. As I learned at a CERN workshop, there is a coordinated approach amongst astronomers and high-energy physicists. They have a coordination and consensus for instrument design, data collection and sampling strategies. Astronomers and high-energy physicists process much of their raw data soon after collection and then discard them, which simplifies later steps such as distribution and analysis. But genomics does not yet have standards for converting raw sequence data into processed data.
  • Genomics poses some of the same challenges as astronomy, atmospheric science, crop science, particle physics, e-commerce and whatever big-data domain you want to think about.

 

Lastly, from the Machine Learning conference, here is a complete troubleshooting session, in which the “NCM” (a computer generated Neural Conversational Model) and an “Actor” (that’s a human, by the way) interact.  As Spock would have said “Fascinating” :

 

Conversation 1: VPN issues.

Describe your problem: i am having issues accessing vpn

NCM: hi

Actor: hello

NCM: could you please let me know what are the operating systems you are using ?

 

Actor: i am using linux

NCM: ok it was in the ssl vpn on which you stated it was broken

NCM: please check if you are logged in to corp network

 

NCM: what is the error that you are running please

 

Actor: i am seeing an error related to vpn

NCM: what is the error message that you are getting when connecting to vpn using network connect ?

 

Actor: connection refused or something like that

NCM: may i know the version of network connect you connect ?

 

Actor: i am not sure i know that

NCM: is the network connect prompting for any user certificate?

 

NCM: ?

Actor: yes

NCM: are you at home or at work at the moment?

 

NCM: office ?

Actor: i am at home

NCM: try this

NCM: goto <URL> page and enter your name and password

 

Actor: done

 

NCM: try logging in with <NAME> and check

if its working fine now

 

Actor: yes , now it works !

 

NCM: great

 

NCM: anything else that i can help

 

Actor: no , that was very useful

 

NCM: you are other than <NAME> : )

 

NCM: have a nice night
POSTSCRIPT

Much chattering this week about IBM’s announcement about its seven nanometer chips. IBM says it has built a computer chip whose smallest features are about seven nanometers wide, meaning it could provide about four times the capacity of today’s chips. Basically, it packs far more transistors into the same space, and that means data can travel more quickly between those transistors.

The announcement, made on Thursday, was hailed as a breakthrough, evidence that semiconductors will continue to improve in the years to come.

NOTE:  Patrick Moorhead, an analyst who closely follows the chip business and who I avidly read, thinks Intel has done the same thing already … but they’re just not telling anybody.

The company said that it had working samples of chips with seven-nanometer transistors. It made the research advance by using silicon-germanium instead of pure silicon in key regions of the molecular-size switches. The new material makes possible faster transistor switching and lower power requirements. The tiny size of these transistors suggests that further advances will require new materials and new manufacturing techniques. As points of comparison to the size of the seven-nanometer transistors, a strand of DNA is about 2.5 nanometers in diameter and a red blood cell is roughly 7,500 nanometers in diameter. IBM said that would make it possible to build microprocessors with more than 20 billion transistors.
Okay. But …..

 

Does IBM have the capability to manufacture these chips, package them in hardware that savvy information technology professionals will want, and then support the rapidly growing ecosystem?

 

No. IBM Research is much like the old Bell Labs: it can invent or engineer something nifty. But the Bell Labs’ folks were not the leaders in the productization field. Like it Watson PR machine, IBM’s ability to get its tales of technical prowess in front of us may be greater than the company’s ability to generate substantial top line growth and a healthy pile of cash after taxes.

Result? the companies with core manufacturing expertise will be the winner. Once a chip is made smaller, others know it can be done which allows the followers to move forward.

 

Related Posts