The Machine learning used in academics/research is quite different from the ML used in Production applications for real usage by end users. Here is a description on what are the differences.
- In academics, the need is to build the next SOTA (State Of The Art) model.
- A 0.1% gain above exiting SOTA is considered exceptional.
- In Production, there is no fixed requirement across all the stakeholders, the Sales team, Product team, Engineering manager, etc. have different requirements.
ML Lifecycle priority
- In academics, GPU/TPU machines with high throughput which can train faster are required.
- In production low latency fast Inference/prediction is required. The users need to be shown the recommendations, the ads fast. A slight delay can reduce the clickthrough rate and thereby revenue drastically.
- In academics, mostly there is a benchmark static dataset on top of which models are built.
- In production, data is constantly getting generated by the users and may have bias.
- Working with shifting datasets make it a challenge.
Bias and Fairness
- In academics, in front of achieving the SOTA model goal, fairness takes a low priority.
- In production, the fairness of the ML model can not be ignored.
- In academics why the model predicts the result is often not a priority.
- In production, explainability is of greater priority on why the model makes this decision and the model should be more than a black box.
We discussed how ML in research is different from ML in production across the following categories:
- Lifecycle priority
- Bias and Fairness
For more such insights follow @soumnedrak_