DEEP LEARNING IS THE FUTURE: Q&A WITH NAVEEN RAO OF NERVANA SYSTEMS

CME Group was one of several companies taking part in a $20.5 million funding round for the San Diego startup, Nervana Systems. The company specializes in a biologically inspired form of artificial intelligence known as deep learning. Based on neuroscience, the technology uses algorithmic models to learn representation from large datasets. It is presently used in voice command and face recognition capabilities, but Nervana Systems CEO Naveen Rao anticipates its application across a wide range of industries including financial services, agriculture, and medical diagnostics.

Rao began his career as a computer engineer, received a PhD in computational neuroscience, and worked in a neuromorphic design research group at QUALCOMM before founding Nervana Systems in 2014. According to Venturebeat, the company has accumulated a total of $24.4 million to date.  Nervana Systems also recently released its Neon deep learning software under an open source license. To learn more about Nervana Systems and the place of deep learning in the global economy, we spoke with Rao. This is an edited version of our conversation.

Most people are now familiar with the idea that data affects everything. What is deep learning, and how is it different?

Deep learning is really the latest iteration of neural network approaches to machine learning problems. Basically we took some very high level abstraction of how neurons work and how neurons purportedly process information and we try to build mathematical models out of it. This is not a new field — people have been doing this since the fifties when people discovered the basis for how information is transmitted in neurons.

Deep learning is the end of the evolutionary process of learning how neural networks learn to represent information and actually find structure within data.

What does Nervana do to make deep learning more accessible?

One thing about deep learning, there are a lot of aspects to it that are difficult to get it to work on real data problems. We’re trying to make that very easy by wrapping software layers around it so that people can safely board its problem spaces.

The other piece – deep learning fundamentally requires a lot of compute resources, and new compute resources that people don’t really have. Right now, it’s in the realm of the Googles, Facebooks, Microsofts of the world, and we want to bring those capabilities to everyone else.

Naveen Rao co-founded Nervana Systems in 2014.

In which industries will deep learning have the most impact in the near future?

In the next 5 to 10 years the world is going to look very different due to these machine learning techniques. But one early area I see that is going to be really important is finance.  Regulatory enforcement is something that financial services companies are very interested in and that’s something we’re seeing from multiple places.

Other verticals are agriculture—agriculture is kind of on this tipping point where all of a sudden we’re applying technology to these very old problems. We need to feed the world and we’re trying to figure out new ways to do that using technology.

Medical diagnostic space is something that I personally would love to see – my whole family are MD’s.  I would really love to see some of these problems that can become standardized instead of having a doctor somewhere who is the best at reading an X-ray, and if you don’t get in front of his eyes you don’t get the best. We should be able to standardize that and build it into an algorithm. You’re going to have criteria for when a cell looks cancerous or not; we can standardize that and code it to an algorithm so it’s definitive, and not an opinion.

Where are we already seeing these kind of algorithms at work?

Every smartphone out there —  Windows phones, Android phones, or iPhones — the voice personal assistants use deep learning to decode your voice, and the reason that these devices now become useful is because of deep learning.

Voice recognition has been around for a long time – even five years ago it was kind of useless.  It made so many errors it wasn’t above a line that made it useful yet. But now, I can talk to my phone, I can ask it a question and it gets it right most of the time.

That also frightens some people, how smart machines are becoming. Is that a misconception with artificial intelligence, that it’s a threat?

We’re pretty far away from that. I’ll never say never but at this point in time, there’s not a good path from here to there. The things we’re building now are better tools, they extend our capabilities just like I can’t push a ton of dirt but I can build a bulldozer that will push a ton of dirt.

In the same way, we can make an algorithm sift through 10 million pictures in a few seconds and actually figure out what’s in them, whereas I can’t do that as an individual. It’s more of a tool in my mind; it’s really something that I think will actually make lives better in multiple ways.

Where do you see it having the biggest economic impact?

It’s kind of like what we had at the industrial revolution.  We had the mechanization of farming and other industries that were disrupted by tractors and other mechanical implements. We’re going to see labor markets shift away from humans doing what we would now consider mundane tasks but there’s no other way to do them.

An example would be people who have millions of loan contracts sitting on a desk and they literally need to just score them for risk. Sounds like a simple problem but when you start scaling it up to a million different contracts you need a lot of humans to do that and it takes a long time. Something like that we can simply build an algorithm. It’s standardized, it works flawlessly and we can do that in minutes instead of a year.

How did you decide to start Nervana?

In 2007, I quit my job as a computer engineer and went back to school to get a PhD in computational neuroscience. The reason was to understand what computation really means in a biological context.

Looking at visual data and integrating that with sound and smell and touch and all these things; putting it all together into one cohesive version of the world – that’s what our brains do.  That’s exactly what we want to do with large datasets.

I was actually working in a neuromorphic design research group in QUALCOMM. Neuromorphic design is basically taking biological inspiration at the lower level, on the circuit level, and trying to build a synthetic machine out of that. That was more of a research project, but I started talking to potential clients who might use our stuff and there’s a lot of growing demand for deep learning.

Now we have the data, we have the computational techniques, which is deep learning, and we have a market need. We have tons of data and we need to make sense of it. Companies need to find insights in the data. With those things coming together was when we launched Nervana.

What are the next steps for Nervana?

With the closing of this round, it gives us the fuel to continue the development process. We’re hiring quite a lot. We’re going to be launching our cloud service to make it accessible for people. The idea is really to bring this state of the art performance and capability to anyone who wants to pay for it on a small scale.

We’re very excited about it because there are certain kinds of techniques or models that people haven’t even tried because it’s just not possible on today’s architectures. For instance, it can take a year for us to train certain algorithms on some large datasets. If it’s going to take a year, no one’s going to do it. But we can literally take those things that would take a year and run them in a day or even hours. We see that opening up a whole new set of possibilities. The next year and a half is going to be very exciting in this field for us.

 

DEEP LEARNING IS THE FUTURE: Q&A WITH NAVEEN RAO OF NERVANA SYSTEMS的更多相关文章

  1. 博弈论揭示了深度学习的未来(译自:Game Theory Reveals the Future of Deep Learning)

    Game Theory Reveals the Future of Deep Learning Carlos E. Perez Deep Learning Patterns, Methodology ...

  2. (转) Deep Learning Research Review Week 2: Reinforcement Learning

      Deep Learning Research Review Week 2: Reinforcement Learning 转载自: https://adeshpande3.github.io/ad ...

  3. (转) Deep Learning in a Nutshell: Reinforcement Learning

    Deep Learning in a Nutshell: Reinforcement Learning   Share: Posted on September 8, 2016by Tim Dettm ...

  4. A Statistical View of Deep Learning (III): Memory and Kernels

    A Statistical View of Deep Learning (III): Memory and Kernels Memory, the ways in which we remember ...

  5. 课程一(Neural Networks and Deep Learning),第一周(Introduction to Deep Learning)—— 1、经常提及的问题

    Frequently Asked Questions Congratulations to be part of the first class of the Deep Learning Specia ...

  6. 【深度学习Deep Learning】资料大全

    最近在学深度学习相关的东西,在网上搜集到了一些不错的资料,现在汇总一下: Free Online Books  by Yoshua Bengio, Ian Goodfellow and Aaron C ...

  7. Deep Learning 19_深度学习UFLDL教程:Convolutional Neural Network_Exercise(斯坦福大学深度学习教程)

    理论知识:Optimization: Stochastic Gradient Descent和Convolutional Neural Network CNN卷积神经网络推导和实现.Deep lear ...

  8. (转) Awesome - Most Cited Deep Learning Papers

    转自:https://github.com/terryum/awesome-deep-learning-papers Awesome - Most Cited Deep Learning Papers ...

  9. (转) The major advancements in Deep Learning in 2016

    The major advancements in Deep Learning in 2016 Pablo Tue, Dec 6, 2016 in MACHINE LEARNING DEEP LEAR ...

随机推荐

  1. SqlBulkCopy高效能批量插入SQL SERVER

    what SqlBulkCopy是.NET提供的用来批量插入数据的一个类,特别是将内存中的数据一次性插入到数据库,目前只能插入到SQL SERVER数据库,数据源可以是DataTable.IDataR ...

  2. C++面试题算法

    #include <iostream> #include <string> using namespace std ; /* 题目:给一个字符串.例如 "ababc& ...

  3. 在文件地理数据库中使用 SQL 进行报告和分析 (转)

    ================以下摘自ArcGIS10.1帮助=================== 文件地理数据库允许在 QueryDef 中通过 SubFields(字段列表)方法使用表达式和别 ...

  4. javascript进击(八)JSON

    JSON 是存储和交换文本信息的语法.类似 XML. JSON 比 XML 更小.更快,更易解析. 什么是 JSON ? JSON 指的是 JavaScript 对象表示法(JavaScript Ob ...

  5. openwrt chinadns

    opkg update opkg install ip ipset resolveip iptables-mod-tproxy resolveip libopenssl luci-i18n-base- ...

  6. Linux基础命令(三)

    一.常用命令—文件目录类命令 1.ls 列出指定或默认目录的文件信息 使用形式: ls [选项] [目录名] 实例: $ls $ls –als $ls /home/sq/Desktop $ls ./D ...

  7. Bootstrap的竞争对手Zurb Foundation

    Bootstrap并不是唯一的前端开发框架,比如还有JQuery UI.HTML5 Boilerplate等等.但对于Bootstrap来说,真正的竞争对手是Zurb Foundation.Boots ...

  8. C#扫盲之:前台线程后台线程

    1.线程分类 线程由程序员创建,可是创建的方式不同,总体来说有两种,一种是个人构造,也就是使用thread类new线程对象创建,这一类线程是大部分程序员知道的,也叫专用线程;还有一种是由CLR创建,这 ...

  9. java Spring配置数据单元

    基本原理 - 容器和bean 在Spring中,那些组成你应用程序的主体(backbone)及由Spring IoC容器所管理的对象,被称之为bean. 简单地讲,bean就是由Spring容器初始化 ...

  10. Sun开发的JINI技术在网络中的应用

    一.概述 Sun公司开发的Jini系统是一种创新技术,听说它的人很多但较少被人理解.Jini系统可利用Java扩展网络计算的范围,并有可能成为最佳解决方案. Jini是Sun公司的研究与开发项目,它能 ...