How to construct a Proof of Concept for an AI project ?
1. Clearly define the purpose of the AI task
It is necessary to clarify the background of the problem with the stakeholders, the operating mechanism of the original service and the goals they want to achieve. And can analyze whether there is AI in the middle that can make up the part, and what benefits and costs can be added if necessary.
- Confirm important window contacts (with relevant knowledge/can trust), and establish an instant communication channel
- Effectively record and synchronize project stakeholders’ information
- Able to draw up phased goals
- Confirm whether the data required for the question is reasonable, what is the method of collecting the data or how to build it
- Preliminarily clarify the evaluation index (metric) as the question to improve the standard
2. Data collection and processing
- Must have a reliable and stable way of receiving data.
- If you need to mark the data, how to integrate relevant domain knowledge, or if there is relevant information for reference.
- Check whether the collected information is normal and complete, and whether it meets the requirements of the target task
Ｏthers:
- Assess whether traditional identification methods can be used, and semi-automatic labeling is done first to reduce the cost of manual labeling. Ｅxample: Use template matching to assist label the object.
3. Looking for methods related to building:
- Evaluate the appropriate method based on the current data size/type.
- Try to find existing methods from google, as a baseline reference.
- From Github to find a package open to use by others, if there is a good method and commercial authorization can be used.
- Read related papers/curriculums of major U.S. universities/public competition materials.
Resources:
- Resource websites with AI papers and programs in various fields:https://paperswithcode.com/
- Kaggle: Data analysis/modeling competition platform, sometimes there will be programs for related topics for reference. https://www.kaggle.com/
To Do:
- How to establish a data/model version control mechanism?
- How to quickly deploy the environment? Reduce the time to rebuild the environment
Course
1. CMU AI software engineering course resources
- Machine learning in Ai production https://ckaestne.github.io/seai/
- Software Engineering for Ai-Enabled Systems https://ckaestne.github.io/seai/S2020/#course-content
