AI and deep learning are continuing to advance and will change, or is already changing, how many industries and businesses function. Enterprise level businesses are no different. In fact, one would argue that these large, Fortune 2000 companies are prime candidates for the intelligent, deep learning provided by this growing sector of programming.
Skymind is one such company providing this intelligent framework. Java might not be the go-to language anymore for Silicon Valley, but the fact of the matter is that most big businesses are still using it for enterprise level IT and that’s where Skymind’s Deeplearning4j comes into play. This framework can be used and adapted to fit many needs of large companies – everything from virtual assistants and photo recognition to churn prevention is possible with AI and AI is made possible thanks to frameworks like Deeplearning4j.
But who better to explain all of this in more detail than Chris Nicholson, co-founder and CEO of Skymind. We had the opportunity to speak with him about his past roles and how he ended up co-founding an open-core AI company. He also breaks down the differences between AI, machine learning, and deep learning, as well as going into detail about who Skymind is intended to be used by.
Check it out below!
Care to introduce yourself and your role at Skymind?
My name is Chris, and I’m the CEO of Skymind. I actually used to be a journalist. I wrote for the New York Times, the International Herald Tribune (RIP) and Bloomberg News, covering finance, tech and M&A. At a certain point, after seeing the trouble in the newspaper industry, I taught myself to code, and I eventually came to Silicon Valley with the notion that I would start fresh as a junior developer. I wanted to get on the right side of the robots, since tech had radically changed the industry I knew.
That’s not what happened though — I ended up interviewing with a lot of YC startups, and all they wanted to know was how they should talk to reporters. They wanted to know how to get press, since the difference between life and death, for a startup, is visibility. So I landed in PR, like a lot of other former reporters. That was at FutureAdvisor, a robo-advisor that was backed by Sequoia and bought by BlackRock about a year ago. I was with them for 20 months, saw them grow 45x in that time, and served as their head of communications and recruiting.
All by itself, that was a wild ride. But at the same time, I was working nights and weekends on Skymind with my co-founder, Adam Gibson. He had been working in machine learning for years and we figured out that he could build the product and I could handle the business side. We were living in a hacker house together for the first year, sleeping in bunkbeds and bootstrapping off our salaries. My role at Skymind changes about every three months. We were in the YC Winter 2016 batch, and at the end of the program, I was a fund-raiser. We raised about $3 million in seed from investors like SV Angel, Tencent, Ray Lane and others. These days I’m still a recruiter. But mostly I’m in sales. We’re working with customers to make sure they get the right support when they build deep learning solutions with our open-source code. So I spend a lot of time listening, making sure every one is on the same page, and that nothing’s blocking the team.
What is Skymind, in just a couple of sentences?
Skymind is Cloudera for deep learning. We’re an open-core AI company. We built Deeplearning4j, the most popular deep learning framework for Java, which is the world’s largest programming language. Java isn’t hot in Silicon Valley, but it is the dominant language in corporate IT. Large organizations have made tremendous investments in Java, and in the big data ecosystem based on the JVM, like Hadoop, Spark, Cassandra, Solr and Kafka. So when AI meets the enterprise, it happens in Java or on the JVM, and it happens with our tools.
Companies are using us for fraud detection, recommender systems, predictive analytics like churn prevention or forecasting, and for image recognition. Deep learning capable of superhuman accuracy on many problems, which is a quantum leap over traditional machine learning algorithms. That’s a huge increase in accuracy, and you can attach a dollar amount to that, so a lot of large organizations have plans under way to implement deep learning. We see a broad shift across industries, notably financial services, telecoms, e-commerce and manufacturing.
What inspired the creation of Skymind? And what’s your founding story?
Skymind got started in early 2014. My co-founder and I thought that enterprise needed an open-source AI layer, just like it had open-source layers for big data storage with Hadoop, or Linux for the OS. It seemed like an AI layer had the potential to create even more value than OS or data storage, because AI is telling you what’s happening with your raw data. It’s creating knowledge from noise, and you can act on that.
So we created Deeplearning4j with some very early contributors, and since then it’s become the biggest deep learning framework for the JVM. We’re following the typical open-core playbook: Skymind does support, training and services for our enterprise distribution, the Skymind Intelligence Layer, or SKIL. An intelligence layer takes in data and puts out decisions about that data. Every open-source business draws a line somewhere, and SKIL bundles a couple closed-source packages as well. We help big companies build deep learning solutions with a distro that deploys easily to the stack they have. And since it’s Java, they can leverage their existing teams to use it.
What customers and businesses is Skymind meant for?
Broadly speaking, Skymind brings deep learning to the Fortune 2000. Enterprise customers. But within that, there are nuances. You can imagine the AI market as a pyramid, with a few very sophisticated companies at the top. There are four strata of companies, moving from narrow to wide.
1) At the top you have large, high-tech companies like Google, Facebook, Amazon, Baidu and Microsoft — they created their own deep learning frameworks, and some of them are using those frameworks to lure users to their public clouds.
2) The second layer down are companies with teams that can build deep learning solutions, but they don’t have the time to maintain their own frameworks. It doesn’t make sense for them to do it — it would be a huge waste of time. Those companies adopt external tools and build their own AI solutions. That includes a lot of Wall Street firms, telecoms, tech companies that have hired quants and data scientists. Many of them have already deployed deep learning, and if they’re on the JVM, they did it with us. We support the solutions they built.
3) The third level is companies that have successfully implemented a big data strategy — they know how to gather data, move it and store it. But they haven’t successfully implemented machine or deep learning yet. They’ve invested a lot in data storage, but they’re not seeing the returns yet, because storage alone is not enough. You need to analyze that data. You need to make predictions based on that data so that you know what to do. That’s the business case for AI. It fulfills the promise of big data. We help those companies build deep learning solutions.
4) The fourth level is companies that are interested in AI, but haven’t implemented a big data strategy yet. We introduce those companies to teams and ideas that can help them, and we say we’ll see them in a year or two. For AI to be useful, you have to know which problems you want to solve, and gather the data that’s relevant to those problems.
For someone who has no idea what deep learning or deep learning libraries entail, how would you explain it to them? Why has it exploded so much recently?
First, it’s good to know the difference between AI, machine learning and deep learning: They’re like a set of Russian dolls going from big to small. Machine learning is just a subset of AI. And deep learning is just a subset of machine learning. But it’s a very advanced subset, in the sense that most of the advances in research are coming from deep learning.
Another phrase for deep learning is deep artificial neural networks, which are a set of algorithms that have been around for many years. People were skeptical about neural nets for a long time, but as hardware got more powerful and as big data became available, deep neural nets started breaking records in accuracy. They needed the powerful chips and the large datasets to produce accurate models of the world.
AI is just math and code. Let’s say you put that math and code in a black box. What goes into the box is data, and what comes out are decisions about the data. The kinds of decisions AI can make are classifying data, clustering data by similarities, and making predictions about data. Classification might be: is this transaction fraudulent or not? Is this email spam or not? Or: Which person’s face is in the photo?
Classification means applying a name or a category to data. Clustering can make decisions about data even if it’s nameless. It can cluster data by similarity. If you upload a photo to Google photo search, it will return similar photos. Making predictions about data usually has to do with time series. A time series is a series of numbers or logs. It could be sensor data from a wearable, or quarterly GDP figures in an economic table, or web activity. A neural net can be trained to predict what happens next in the series.
SKIL just launched – can you tell us about that? How does it differ to your other work?
SKIL is the enterprise distribution of the open-source libraries we built. It includes Deeplearning4j, our neural net configuration layer; ND4J, which is Numpy for the JVM; the C++ engine libnd4j to power that; DataVec for data pre-processing and vectorization; Arbiter for model evaluation; and JavaCPP as a bridge to lower-level code. SKIL is the bundle of versioned libraries that we support. It’s like CDH or HDP for Hadoop. SKIL solves a lot of problems that companies run into when they try to deploy deep learning solutions. It’s Dockerized and runs on DCOS and Mesos for resource management.
That means it’s agnostic to your OS and hardware. It works with Spark, so you can throw a ton of GPUs and CPUs at your neural network to train it quickly. And it plugs into Kafka, so you can process data in real-time. With SKIL, we created an AI component that plugs into the production stack. You can auto-scale SKIL models elastically as you need to process higher volumes of data. That’s something no other deep learning tool can do, and that’s why we say we’re deep learning for enterprise.
Anything excited coming to Skymind that you can tell us about?
We just built a way to import models from other deep learning frameworks. So people who run into roadblocks with Caffe or TensorFlow or Torch will be able to import those models and deploy them with SKIL. That’s exciting. It means that teams working in Python don’t have to lose their work with they interface with the JVM stack for deployment. That keeps the data science team happy as well as DevOps. Nobody has to rewrite anything, you can just port it over.
That’s the real workflow in AI right now. Our Scala API is getting better by the month. We just introduced deep reinforcement learning, which was part of the AlphaGo algorithm that beat the human champion earlier this year. So yea, a lot of exciting things are happening with the product. We’ve got some demos on the way to show how all this translates to use cases in financial services and healthcare and security, so that’s exciting. Pretty soon we’ll be able to show people some fancy dashboards on top of the underlying AI.
Thanks for taking the time to answer our questions, anything you’d like to close with?
Yes! People are making huge promises about AI and chatbots and things like that. We’re in a weird position where on the one hand, deep learning is making real breakthroughs, and on the other, there’s a lot of misleading hype. So the people making a big decision about AI should really dig in and learn as much as they can. Ask companies what algorithms they’re using, get some references and figure out which technology is really the most promising.
When people start wondering how they can use AI to transform their business, they should watch out for magical thinking. Deep learning is doing amazing new things, it’s breaking records, but it has some constraints. You have to start out with a problem whose outcomes you care about: e.g. How can I detect fraud? How can I predict which book my customers want to buy? Or which person my user might want to date? Once you know what you want to solve, then you need data that correlates with those outcomes.
If you don’t have the right data for your algorithms to train on, you won’t get an accurate or powerful model. You need to have the data, or have a plan to get the data. It’s not impossible to get the data, but you need to be aware of it and tackle that before your algorithms can do their work. The only way around that is by using deep learning models that trained on someone else’s data. That’s possible too. But the data has to come from somewhere.
The decisions people make now about AI are going to have a huge impact on their careers and their companies, so it’s important to start thinking about it now, and to get it right.
We’d like to thank Chris for taking the time to answer some of our questions. Find out more about Skymind here.
This post is part of our contributor series. It is written and published independently of TNW.
This post is part of our contributor series. The views expressed are the author's own and not necessarily shared by TNW.