What to Bring on Your ML Journey
Many CTOs feel they should be doing something with machine learning. Lately customers and partners are asking me how to apply machine learning more than anything else.
Where most struggle is in problem selection. CTOs worry if they pick the wrong project and it doesn’t work, that’s it—their one shot is done. Often the goal has to do with increasing efficiency or revenue, which sounds pretty straightforward. But if you don’t get it right, you can spend time and money on what might feel like a vast abyss of complexity, with little to show for your voyage.
I’m not sure the perfect problem, the one with the high-assurance outcome, exists—at least not yet. That’s why we see so many early applications of ML to such a wide variety of problems, but far fewer at production scale. But there’s a lot of good that can come from an ML initiative. ML is a new frontier, and it requires a different mindset. With ML, the process and the outcomes can be incredibly challenging and rewarding.
The Road to Kilimanjaro
About ten years ago, I decided to climb Mt. Kilimanjaro. At first, I wasn’t sure why I wanted to go. I appreciated my life and family, and there were no business or financial outcomes tied to the trek. But sometimes the intangible moves us, and I found myself researching and eventually booking the trip.
My climb has several parallels to embarking on an ML initiative. Success isn’t assured. The published “summit rate” for Kilimanjaro is roughly 65%. To land on the right side of that probability, it took a lot of preparation, learning, clarifying my destination and re-committing to the goal with the help of fellow trekkers. If you’ll allow me to be your guide here, you’ll see how you can apply good mountaineering to an ML initiative and tilt the probability of success in your favor.
What you’ll need:
1. A problem definition. We touched on this already, but you need to have some idea of your destination—the problem you’re looking to solve with ML and the goal you're trying to reach. On my first mountaineering trek, I weighed several options. Being a first-time climber, I realized I didn't need to tackle Mt. Everest, with a very real 1-2% chance of dying. The same goes for your ML project. Identify a project that’s significant and relevant to your business, but don’t pursue the highest-stakes problem first. If you need a framework for project selection, Andrew Ng has you covered with this practical approach.
2. Good data. Think of your data and the process of data transformation as the thousands of individual steps you take up the mountain. At Google, it was only because we had accomplished the foundational work—curating, extracting, transforming, loading, sharing and experimenting with data—that we were able to test the frontier of machine learning in our organization. Maybe one of the most significant opportunities for organizations today will be how they leverage data and make it easy to share with groups that have high sophistication levels, like data scientists, analysts and business intelligence professionals, who are trying to answer a difficult question quickly for the head of a line of business. Unless you have that level of data sophistication, machine learning will probably be out of reach for the foreseeable future. Here’s some quick advice from our Google Developers community for thinking through the data section of your journey.
3. A cross-functional team. When you’re hiking Mt. Kilimanjaro, you need people on your team who can tie knots and know what food to bring, which boots to buy and how long to break them in. There were points along the hike at higher altitudes where I experienced severe sleep apnea, meaning my breathing basically stopped. But thanks to team members who were experienced mountaineers and medical professionals, I could manage this dangerous condition. You need a cross-functional team for ML too. You may need people from the Tensorflow community, the microcontroller community, the Python community. A diverse group gives you a wealth of knowledge, skills and perspective. If you’re looking for ideas about who to bring along on your ML journey, the suggestions here are a great start. You’ll be grateful when you’re halfway up the mountain.
4. An experimentation environment. You need to practice before the big day. Before my trip, I mapped out distances, worked out on the treadmill with 50- to 70-pound backpacks and followed a training schedule. I learned the language of mountaineering. I tied knots at night and on weekends. I bought my boots and ensured all the equipment worked. You need an excellent experimentation environment, one where you can figure out which development tools, frameworks and technologies work best for you, how to get comfortable transforming data, and where you can see how others have solved similar problems. You need to practice. Fortunately, there’s a wealth of free resources available online. Just pick a topic that’s close to the business problem you’re trying to solve.
5. Off-ramps. Kilimanjaro has two peaks: Gilman’s Point and Uhuru Peak. Many people who take the trek make it to Gilman’s Point. The last climb is only about 150 meters, but the air is so thin, many experience severe issues finishing the climb. Sometimes all the preparation can’t overcome unexpected and uncontrollable complications, such as underlying medical conditions. Your ML initiative can have two peaks, where, even if you don’t reach the ultimate goal, you’ll have learned enough that you can coach to the peak you’ve reached and be better prepared to reach the second in future endeavors. Either way, you’ve accomplished something memorable and valuable.
6. Re-commitment as you go. Hiking up Mt. Kilimanjaro is a completely optional experience. So is taking on an ML initiative. You choose this path, and you’ll need to keep the band together as you go. At some point, everyone will experience frustration and difficulty. You’ll need to be there for each other. I found strength and inspiration from one guide in particular who shared his experiences and took great interest in mine. We motivated each other, even if it came via laughter at my expense. If you’ve climbed the mountain you’ll know one of the key motivational phrases that is both fun and impossible [for me] to say correctly: poa kichizi kama ndizi, ndani ya friji ;)
7. Knowledge and vocabulary. Buy the guide, and learn the terminology. I was the founder of an analytics and ML-centric startup at the time (yes, even in 2012). One of our board members, who invited me on this trek, was an avid mountaineer. The first thing he suggested: get the essential guide to mountaineering and learn the language. Every activity, be it mountaineering or machine learning, has its precise vocabulary. And it can evolve over time. Even if you are an experienced software engineer, you need to speak the specific language of ML, or you'll be lost. Lucky for us, Chris Albon has you covered (the flashcards are highly recommended).
8. The why. When you get to 18,000 feet, you intermittently stop breathing at night, your legs are crushed and oxygen is scarce. Something has to keep you going. And it better be a pretty good reason. Why are you taking on this ML initiative? I’ve always tried to tell my daughters they are capable of anything they set their minds and effort toward. Honestly, despite being an Army veteran, former US olympic development program athlete and generally in pretty good shape, I was very nervous about this climb. So my why? To show my daughters that fear is the beginning of most things worth doing.
The Final Climb
For the final climb, you begin in darkness. To reach the summit at the optimal time, with the optimal beauty, you must start in the dark. It’s hard to describe what you see when you reach the top of Kilimanjaro. You can see the curvature of the earth, and you can see glaciers—which you might not expect when you’re in Africa. All the work it took you to get there, the planning, the tools, the voyage, finding a frontier feels exciting because it’s so rare. It's temporal, and it doesn't last. But it's memorable. You know I'm talking about both now, Kilimanjaro and ML.
Taking on a machine learning initiative requires a voyager spirit, faith, confidence and trust in the tools and capabilities you've built over time. An interesting byproduct is when you finish, you want to take on the next challenge immediately. When you're on the frontier of a technology, whether it's mobile, ML or quantum computing, it's an exciting place to be, and you want to stay there.
Your voyage will inspire others. A lot of people asked me about my trek after my return. I got a lot of free beers for telling that story. It’s similar to the projects we tackle. The path to ML is not just our path; it's blazing someone else's trail as well.
So go ahead and put something at stake. Maybe you don’t have the perfect problem definition; a good problem definition will do. If you prepare, bring your best voyager mindset and travel with the right team, you may discover that you and your organization are capable of great ascents, as well...and that the value is in the journey itself.
Customer-obsessed technical leader, formerly with Google and Amazon. Expert at scaling customer-facing technical teams, optimizing Go-to-Market, and delivering large-scale technology transformation.
7 个月Well done Will Grannis! A much more inspirational (and accurate) analogy than the one I literally dreamt up this morning (“LLMs are like Sawzalls…tools to build things you never even thought of before, but you need to learn how to use them…because turning them on and walking away can be dangerous”).
Course Director: AI Executive Education @ GenAI
7 个月This is incredibly insightful. Thanks for sharing.
CEO at MetroStar & Chairperson at Zoomph
2 年Thank you for sharing this Will. It’s got the right plays to get folks thinking and plan that journey. Like Will said, I’ve found the planning, focus, capacity, and patience to be crucial along our journey.
Founder and CEO @hopr | Engineer | Systems Thinker | Inventor | Faith-driven | Mission-minded | AMTD evangelist. I help enterprises protect their workloads and data.
3 年Climbing a mountain is a familiar analogy for tech innovators. I like the comparison with your ML key elements. I often think the most important step is one of the earliest in planning: determining which side of the mountain and which path you make the climb. I’m just wired to think that way.
Sr. Director of AI, Google || AI CTO | Entrepreneur | ex-Professor
3 年Nice article, Will! Love the analogy (and the pics!) My favorite bit is about the cross-functional team. Too often leaders think they can do ML by hiring ML engineers. Yes, but one also needs the rest of the cross-functional team - not just for the expertise but also for the diversity of perspectives.