It should come as no surprise that academia is a very different world from business. While the former focuses on understanding the world such as it is and aims for accuracy and reliability, the latter is focused on innovation and speed. These goals lead to very different philosophies, methodologies, and outcomes.
Of course, academia and business aren’t completely disconnected, and in fact, some of the most impressive scientific achievements of the modern era can be attributed to companies that focus on research like Bell Labs or IBM. It’s the best of both worlds.
But that kind of balance is difficult to achieve, and the differences can have a very real impact on our projects, especially when many academics are opting for data science as a career path.
| Aspect | Academic Machine Learning | Production-Oriented Machine Learning |
|---|---|---|
| Focus | Theoretical research | Practical application |
| Goal | Advancing knowledge | Solving real-world problems |
| Data | Benchmark datasets | Diverse and messy data |
| Performance Metrics | Evaluation measures | Business-specific metrics |
| Model Complexity | Emphasis on complexity | Balanced complexity and efficiency |
| Scalability | Not a primary concern | Scalable to handle large datasets |
| Implementation | Prototypes and experiments | Robust and well-tested code |
| Deployment | Limited or ad-hoc | Deployable to production systems |
| Speed | Focus on accuracy | Efficient and fast inference |
| Interpretability | Interpretability studies | Trade-off between accuracy and explainability |
| Maintenance | Focus on research papers | Ongoing maintenance and updates |
| Collaboration | Research community | Cross-functional teams |
| Error Tolerance | Tolerant of minor errors | Rigorous error handling |
| Integration with tools | Varied tools and libraries | Seamless integration with workflows |
Let’s get one thing straight: a person who has dedicated their lives to research is an extremely valuable asset for any project, but it’s very important to understand the baggage they bring with them so we can manage our expectations.
Academic and Business Approaches
Search for any academic paper on the internet, and you will quickly notice some key differences between how academia presents information and how data scientists present information.
Academic papers have a focus on methodology; it’s not what we find but how we find it. The value of an academic paper lies in the possibility of peers understanding the process and being able to reproduce it.
On the other hand, data scientists presenting their work to product owners and decision-makers tend to have a results-first approach. An executive summary translates the results to an audience that may not understand the intricacies of the methodology.
We trust that the data scientist knows what they’re doing, just like a developer knows how to code and a manager knows how to organize a team. Instead what we want to know is what they found and what we can do with that information.
If I had to pick a word to describe the business world it would be pragmatic. If information cannot help us make better decisions then why did we spend time and resources researching it in the first place?
Case in point, it’s perfectly feasible for an academic to spend time and resources designing new algorithms even if the results are similar to other established methodologies. This kind of research opens the field to new possibilities and might push innovation in the future.
The same research in a business environment is a harder sell. Unless we can show the decision-makers that a new algorithm is going to significantly change our cost or outcomes, odds are it’s going to be a steep climb for the data scientist.
Time is another relevant factor. By business standards, academia moves at a glacial pace. Publishing an article, no matter how innovative, can take months of peer reviews, edits, and rewrites.
Compare that to a world of fast deliveries, innovation, digital acceleration, and early adoption. Markets are full of competition, and that competition drives companies to adapt and change with the times.
The World of Machine Learning
In a way, we could say that the business world tends to be risk-acceptant, being a market leader means that we can accept certain levels of uncertainty and hope for the best. Academia on the other hand is risk-averse, methodology isn’t widely adopted until it’s been proven with a higher degree of certainty.
These approaches are reflected in how each area adopts machine learning and will help us understand how almost identical tools can lead to widely different results.
Academia values confidence and accuracy, the former means that we want a measuring tool that delivers constant and predictable results. The latter means that scientists want their results to be as accurate as possible.
If you’re thinking, “yeah, but’s that’s true for the business world as well” you’re absolutely right. The difference is that in academic fields, accuracy and confidence take precedence over time and costs. As long as it’s within their budget, scientists will push for perfection.
Of course, we want to make the most accurate algorithm possible, but we also have to take into account that in production machine learning is part of a pipeline, and as we all know, a process is only as fast as the slowest step.
Accuracy and confidence often come at the cost of speed. The faster we create results, the less reliable the models become, especially if corners are cut. The question we have to ask ourselves is, how much risk are we willing to accept?
Imagine that you want to launch a revamped product, but you are not sure how your audience is going to react, so you ask two teams to make projections. You are 95% certain that the results provided by team A will be extremely accurate, but they are going to deliver their results a day before launch.
As for team B, you’re about 80% certain that you can trust their results, and they can deliver weeks before launch, giving your team time to make changes. Who do you pick?
Academic models aren’t built with scalability and long-term goals in mind. It doesn’t matter if a program is memory inefficient as long you can use it long enough to finish your research. The same can’t be said for programs in a pipeline.
Resource-intensive tasks are costly, and that might be fine for small bursts, but for long-term projects and automation, we need efficient solutions. In other words, we have to take into account how much it costs to keep upgrading or scaling a model and to keep it running.
Should We Bring In the Academics?
So far, I’ve made it sound like we should torch every curriculum with an academic background. Not so fast. There are plenty of good reasons to invite former academics to your project.
While the initial culture shock might be surprising, once the new data scientist understands the process and protocols, they adapt very quickly. Good management and coaching will definitely ease the transition.
Academics have to face some very hard and abstract problems, which often require creative solutions. A biologist might not seem like the best person to analyze market trends. Until you realize that they have used graph theory to model behaviors in very innovative ways that can be transplanted to your project.
And to be quite honest, nothing says tolerance to frustration as being an expert who teaches undergrads. All jokes aside, teaching is related to communication skills, leadership, and empathy, all good skills to have for teamwork.
