Building a machine learning solution is not the same as building a general software solution. And overseeing a machine learning project involves some unique aspects and potential pitfalls.
In this guide, we’ll go over all the insights you need to deliver a successful machine learning solution.
For a higher-level overview, also take a look at What Is Artificial Intelligence? Everything a CEO Needs to Know and 4 Critical Success Factors in Artificial Intelligence.
Who is this article for?
This article is for people who have been tasked with building a machine learning solution for a medium or large business. If you want to learn machine learning for fun or for general knowledge, building a proof of concept on your own or with an inexperienced team is completely fine. But we’re going to assume you are tasked with building a large-scale machine learning project that takes on real risk (if it fails) and generates real value (when it succeeds).
What is machine learning and why is it important?
While machine learning can add huge value in a large variety of different areas, it’s also often overhyped. We’ve all seen science fiction movies, and it can be tempting to think of machine learning as something that gives machines human-level intelligence. In reality, while machine learning can help you in many ways, it’s better thought of as a specialized tool to analyze data than as a silver bullet to solve any problem.
What is machine learning?
Machine learning is ultimately about finding patterns in structured data and making predictions. These can be (and often are) predictions about what will happen in the future. But this is not the only way you’ll find the term “predictions” used in machine learning solutions. Often it also means predicting answers to questions like: “What kind of dog is in this image?” The latter kind of prediction isn’t a time-based prediction (looking into the future), but rather a prediction in terms of: “What answer would an all-knowing oracle give if asked this specific question?”
How does machine learning differ from traditional software?
Machine learning can also be thought of as “learning from data”. Traditional software solutions are built around deduction (a smart person identifies a set of rules and codes them up as a series of if statements, and these rules can then be applied to data), while machine learning solutions are built around induction (a machine learning algorithm automatically discovers rules by looking at a large number of examples, and these rules can then be applied to still more data).
How does machine learning differ from AI?
As an aside, we’ll use the term “machine learning” throughout this article, but in common usage the terms “artificial intelligence” or “AI” are often used instead. For a more detailed discussion, see Machine Learning vs. Artificial Intelligence. They are not the same.
Why is machine learning important?
Machine learning, especially with recent advances, can almost certainly bring new opportunities to your business, no matter which field you’re in. One of the reasons machine learning has received so much attention is because it’s being used to power breakthroughs in several seemingly unrelated fields, including:
- Image processing (for example, facial recognition);
- Sound processing (for example, auto captioning videos with subtitles);
- Text processing (for example, translating between different natural languages);
- Time Series processing (for example, predicting future energy use);
- Numerical modeling (for example, estimating a fair price for a house, or the probability that a particular customer will buy a certain product).
Each of these fields expands into yet more subfields (for example, the same algorithms we use for facial recognition can also detect cancer in X-Rays), and more importantly, very similar algorithms can be used across all of these fields, meaning that an advance in one algorithm can lead to advances in many fields. This applies to nearly every field, including medicine, marketing, finance, and trading.
Machine learning in each of these fields is enabled by huge (and growing) datasets. Improvements in machine learning unlock more of the value in this data, so data itself becomes more valuable. And as more attention is given to data, machine learning models improve still further, leading to a virtuous cycle.
Is machine learning a good fit for your company and project?
Machine learning has so much potential that nearly every company wants to use it. But that doesn’t mean that it’s always a good idea.
As a business manager, you run two risks:
- Forcing machine learning into a solution where it can’t add value in an effort to stay modern and relevant;
- Failing to take advantage of machine learning in a solution where it would add value.
While every situation is unique, there are some good rules of thumb regarding whether or not machine learning is good for your company or project.
“Pretty much anything that a normal person can do in <1 sec, we can now automate with AI.” Andrew Ng
Common signs that a machine learning solution is likely to be a good fit.
- You already have a lot of clean, well-structured data, and you already rely on it heavily to make decisions. Reliance on data is a good sign that you would derive value from a machine learning solution. For example, you manually inspect hundreds of columns in your KPIs to try to forecast sales for the next quarter.
- You already use (or can easily imagine using) hand-crafted rules or heuristics. If you can already partially solve the problem using heuristics, then machine learning will help you automatically find more complex and accurate rules using existing examples. If you are a travel agency and you want to recommend a specific hotel package to clients based on their browsing history, you might use hard-coded price ranges and location matching. A machine learning solution can greatly improve your existing hard-coded rules by learning them at scale.
Common signs that a machine learning solution is likely to be a bad fit.
- You rely on human labor to carry out what seems like a repetitive task. Often these “repetitive” tasks are not fully defined and are not as repetitive as they seem. For example, if you employ people to write sales emails or handle support cases, a machine learning solution might be able to help them be more efficient, but it wouldn’t be able to replace them, because many of their tasks rely on empathy and creativity.
- You expect to gather a lot of data. It’s common for teams to implement machine learning preemptively because they expect their app to deal with big data once it’s successful. In these cases, it’s almost always better to focus on gathering and organizing the data first, gaining insights manually, and to think about machine learning at a later stage.
Keep in mind that these are just guidelines. Nearly every company can find machine learning use cases somewhere. The question is whether you need machine learning now and what the ROI (return on investment) will be. If your company is still small or you only have a little bit of data, you might get a bigger ROI from analyzing your data manually. But if not, machine learning is likely to help you create value.
What does a machine learning solution running in production actually look like?
It might be tempting to spend 2–3 weeks and throw together a PoC (proof of concept): a model that can take in some data your company has and produce some potentially useful predictions. There are millions of tutorials showing you how to achieve this in a few lines of code, using some standard libraries, and it’s tempting to think that productionizing your model will be easy. It won’t. While many people might think that creating the first model takes them 80% of the way to a machine learning solution, people who have built and productionized machine learning solutions before will know that the PoC is closer to 10% of the journey.
All machine learning solutions can be loosely broken down into three concepts:
- Data: used to find existing patterns and build the model;
- Code: used to define and serve the model, and integrate it with other services;
- Model: used to generate predictions.
All three of these look very different when comparing a PoC and a production solution.
A production solution also has many more moving parts. A proof of concept often involves building a simple model and verifying whether it can generate predictions that pass a quick sanity-check. Everything can be done on the same machine. By contrast, this is only the first part of a production workflow. At the production stage, you’ll need a beefy training server and a good process for keeping track of different models. You’ll need a way to test the trained models before integrating them with your existing production services, performing inference at scale, and monitoring everything to make sure it’s all holding up. Finally, you’ll iterate on this process many times, since you can improve the data, code, or model components.
What kind of team do you need to productionize a machine learning solution?
When you’re building a machine learning team, it’s tempting to hire researchers from an academic background. If they’ve invented new machine learning algorithms, they must be the best people to put them to use, right? Not always.
Hiring a researcher to build you a machine learning solution is often like hiring someone who designs KitchenAid appliances to cook you a meal. While researchers come up with novel ways of solving problems and have deep theoretical expertise, they often lack more practical expertise in using existing tools, following good engineering practices, and making difficult time/quality tradeoffs.
More often than not, you want someone who can use existing machine learning algorithms and tools – much like a chef expertly whips together a delicious meal using standard ingredients and kitchen appliances – rather than someone to design new tools from scratch.
Machine learning algorithms are ultimately not a competitive advantage. It’s your data that is. This is why companies such as Google and Facebook are happy to open-source their algorithms, but keep their data secret. You’ll often find researchers (and salespeople) talking excitedly about the latest variation of a Neural Network algorithm, but in fact the algorithm is not that important at all. Usually your success is determined by how well you choose the right problem, which data you include, and how well you prepare your data.
The terminology is still not completely consistent, but you’ll usually be looking for at least one machine learning engineer - someone who specializes in building machine learning solutions for business.
Should you find or build a machine learning team?
The main choice you need to make is between building your own machine learning team or hiring a consultancy. Building your own team can take many years, and it’s probably only the right choice if you don’t need to see results urgently and if you expect machine learning to be the key differentiating factor between you and your competitors.
If you build a team, you might also choose to upskill existing employees instead of hiring new ones. In this case, you might wonder whether to teach engineering skills to your mathematicians or to teach mathematical and statistical skills to your engineers. Either way it’s a difficult leap, but a good engineer can often pick up enough mathematics and statistics (especially if they have some fundamental background already) to become a great machine learning engineer in around 2 years, while it usually takes a statistician longer to pick up all the engineering background and toolsets required (often around 5 years).
A good machine learning team doesn’t only rely on expertise in machine learning engineering. Expertise in software engineering, DevOps, and management are all equally as important, and a common structure might look like this:
At Data Revenue, everyone on our team is a machine learning engineer. This gives us an advantage over companies that see machine learning as a means to an end, rather than an end in itself. But it still took us 2 years of trial and error to build a good team, and 4 years to get to the point where we fully trust our hiring process for machine learning engineers (we now hire 1 out of every 150 qualified applicants).
It makes sense to build a team if you think AI is the main factor that will help you leapfrog your competition. If machine learning is simply part of what’s going to get you to the top, then it might be better to bring in external experts.
But hiring an agency can be risky too, especially if you hire the wrong one. So you’ll need to know how to choose a machine learning consultancy.
Where does a machine learning team fit into your organization?
If you do build a machine learning team, you should build it for the whole company, not just one department. A horizontal AI team.
AI experience is very transferable: To a data scientist, your CRM data looks almost the same as your inventory data. And to an algorithm, they look even more similar. So it makes a lot of sense to build one machine learning task force for the whole firm.
What is the ideal structure for a machine learning project?
Structuring just the technical part of a machine learning solution is challenging on its own. It takes a rare mix of business and technical skills to structure the whole project right, and you’ll want to think about all the different stages of a good workflow before you dive in.
One of the most important things in planning a machine learning project is to plan with production in mind. If you aim for a proof of concept, then that’s what you’ll get. While it’s a good start, you want the people working on the PoC to constantly prioritize the goal of getting to production.
A good intermediate phase between a PoC and production is the pilot stage. This looks almost exactly like a production solution, but it might be limited in scope, in how many users it’s rolled out to, or in how much data it uses.
From the get-go, you should carefully structure and timebox the research and experimentation phase. It’s important that everyone knows the goal is: “What is the best model we can get in x days?” rather than “What is the best model we can get?” Otherwise you’ll get stuck in the “just one more experiment” loop forever.
It’s also vital to figure out which internal and external departments of your company the shiny new machine learning solution will impact. If you’re replacing a non-machine learning solution developed by another team, keep in mind that they might be attached to their approach. "Not developed here" is a hard hurdle to cross, be it between teams within your company, or between an external team and yours.
Before you start building, make sure your model will scale. Often a solution works fine in PoC or even the pilot stage, but a machine learning solution has intensive and unique resource needs, and without proper planning, it can easily overwhelm your infrastructure. Make sure that your solution will be efficient enough to scale to training on your full dataset and to handle the number of requests you expect it to get once it’s live.
At a higher level, make sure you’ve considered everything in our Machine Learning Project Checklist before you kick off.
Why you might still fail: common pitfalls
We’ve mentioned the dangers of getting stuck at the PoC stage, having the wrong team, or using machine learning when it isn’t well suited to a particular problem. But there are more potential pitfalls you should be aware of before you move forward:
- People might not trust your model: Even if your model generates very accurate predictions, it might be hard to win your team’s and your customers’ trust if you can’t explain how you’re getting the answer. You should understand any hard requirements regarding explainability and interpretability before you commit to a machine learning solution. In some areas, such as pharmacology, explainability might be a legal requirement too.
- Your data might not be clean and well organized: In sci-fi movies, we see AI that can read whole libraries, or absorb the whole internet. In reality, AI is very sensitive to the quality and the format of data. Machine learning algorithms can usually only work with data in a format similar to an Excel spreadsheet: well defined, quantifiable data.
- You might build an unmaintainable solution: It’s not enough to have a model that works now. Your solution will need to be maintained, updated, and fixed. You’ll need a good pipeline to get changes to the model, code, and data into production in future, or you’ll be left with a model that works today but not tomorrow, and cannot be fixed when your needs change. Having an experienced team who has already made these mistakes – and learned from them – is a good way to avoid this.
- Your model might be over-complicated: Even if your model produces the correct results, it’s surprisingly common for a much simpler solution to be discovered after all the work that goes into building a machine learning solution is done. For example, your research might find that you can predict a price just by looking at the price on the same date a year ago. Therefore, before you build any model, you should first create the most sensible baseline and use it as a benchmark test during the project.
- Your model might hurt or threaten a specific business unit: While machine learning often augments rather than replaces human talent, in some cases people might perceive your solution as a threat to their jobs. For example, a finance professional’s entire career might revolve around finding patterns in data, and perhaps your model automates every aspect of that. Be upfront with what your model aims to do and whether it will impact existing roles in your company. Talk with your teams about how their roles might change, and how using ML for some of the more tedious tasks could free them up to concentrate on the things an algorithm can't do.
- External factors might change: Machine learning solutions are naturally very sensitive to the data they are trained on. If an external event means that current data is very different from historical data, you might see a dramatic decrease in your model’s reliability. For example, models trained on temporal data from April 2019 might be very inaccurate in April 2020, as the COVID-19 crisis broke many patterns and expectations.
- Your model might display bias: It’s common for human biases to be present in the data machine learning is trained on. For example, the COMPAS system used in the US judicial process to determine which prisoners to release early was found to have significant racial biases. If your machine learning solution suffers from bias, it’s vital to detect this as early as possible rather than face shuttering the entire project after it hits production.
Building a machine learning solution can certainly be a daunting task! We hope you’ll feel more confident about it after reading this article. If you have further questions, reach out to find out how we can help you.