From Model to Production

1. Cost/Benefit analysis of an AI implementation:
   My first thought after completing the second lesson on fastai was, what is the best way to do a cost/benefit analysis of working on an AI project. Although it’s becoming easier to do such a project, it still takes effort, especially at a corporate level (cost of a data scientist, infrastructure, involvement of business stakeholders, etc.). There are certain cases where the net benefit is clear cut, but in other cases AI implementation needs to be faster and cost effective to be considered by business. This is where AutoML tools come in (for example Datarobot, Dataiku, Alteryx).
2. Human involvement in AI implementation:
   Transfer learning with pre-trained models, existing architectures have really simplified the process of model building. A lot of the heavy work in statistics has been done before. This is good, why try to reinvent best in class architectures (I understand though that transfer learning is not applicable in certain scenarios). Model building also includes experimentation with hyperparameters. I feel selection of hyperparameters can be automated, or at least a visual map of hyperparameters vs model accuracy can be provided automatically. So what’s left to actually do?
   The human or role of an AI practitioner will be reduced to matching the right model to the business, making certain choices along the way (keep/exclude data, choose hyperparameters), understanding impact of bias/feedback loop, how the model will be finally used, etc. Also, not all implementations can use transfer learning. Overall, it’s good that the more technical aspects of AI implementation will be reduced and practitioners can work on what practically matters.
   PS: The future has arrived faster than I thought. Just the next day after publishing an earlier version of this article, I found out about Microsoft Lobe https://lobe.ai/, a software that enables anyone (democratization) to train models (image classification) using an easy tool. Although the process of training models has been simplified, the conceptual knowledge required to train a good model remains.

Training a model to detect when you are drinking water, using Lobe

3. Reflections on the fastai platform:
From my initial experience with fastai, it looks like a set of libraries/functions built on top of other libraries/functions from PyTorch, Python. This “aggregation” makes AI implementation faster with fewer lines of coding required. I liked the part where widgets are combined with Jupyter notebook, which gives an interactive feel to the process:

Widgets used with fastai/Jupyter notebook

Come to think of it : the fastai platform, AutoML tools, Lobe are all some sort of ML “aggregators” meant to make the model training process more easy.

I look forward to my following project.

Sahil