What are the challenges and shortcomings of machine learning today?
Here are 3 ML challenges and how to tackle them:
1. Building a good enough model:
From our experience working with companies like Uber, Etsy, Zappos Family of Companies, Ancestry, and many more, typically, the biggest challenge in ML is building a model that’s good enough to provide business value.
We often hear that 80% of ML models never make it to production.
But I’m yet to meet a team with a good enough model but couldn’t figure out deployment.
While deployment isn’t trivial, it’s not as different from deploying an application.
2. Identifying the business use case that is feasible from an ML perspective and can provide value:
There needs to be an intersection between the business owners and data scientists within the organization.
It typically means bringing both of these personas into the same room and having a conversation.
That starts by correctly identifying the right business use case and making sure machine learning and data science are part of that process.
3. Lack of predictability:
With machine learning and data science, we often don’t know if we’ll succeed beforehand. So instead of identifying one problem, we take a portfolio approach where ML teams explore multiple projects at once and only double down to extract signal from the data.