The Four Challenges When Deploying In Production
At first sight, the typical lifecycle of machine learning projects looks simple. You select the data, import it in your data science workbench, build your model, test it, build the data pipelines that will feed your model, and deploy it to production. But in real life, the truth is that it can take longer to deploy ML models in production than to develop them.
This entails costs that often make it difficult to create value. In our experience working at various clients, a simple big data application can represent an implementation cost of around 500K€. Secondly, it is difficult to be compliant with data privacy regulations. However, a Ponemon Institute report shows that non-compliance costs an average of $15 million per year.
In his conference talk (available on our YouTube Channel), our research director Sabri Skhiri talks about unique challenges that make it tricky to deploy ML models:
- Disseminated data that need to be collected throughout the company (06:03)
- Regulatory constraints on data usages (08:30)
- The complexity of the technological stack (11:42)
- Operating machine learning models in production (16:19)
Watch the video: