what you need to do for a project
- steps of serious ML apps:
- select the problem(supervised) (case voice activation) gets an audio and 0/1 to if it contains the target word
- get data
- design model
- train model
- test model
- QA (continuous)
- deploy
- maintain
- we talk about QA .(continuous)deploy . maintain and select the problem
- we need 2 target words.
- list of 5 bullet points for a success DL
- project
- Data
- Interest
- Domain Knowledge
- Utility
- Feasibility
- Get data (how many days and how you do collect?)
- 1,2 days get data
- it's difficult to know what is hard about the problem
- keep clear notes on experiments you ran. or have spreadsheet that what worked.
- number 6 - deployment
- large neural net
- lots of process on edge devices in contrast to clouds
- Voice activity detection (VAD) . if there is any audio then encode it.
- Non-ml, see if value > epsilon (do this because it's faster)
- train small neural net in human speech
- when ship the product the data changes
- accent for eg.
- different back ground noise
- new mic
- 7 maintain
- the world changes and you need to maintain
- web search changes
- self driving car (light changes)
- inspection
- edge / cloud
- does a cloud or edge deployment make maintenance easier, and why?
- cloud / detect the problem faster
- does a cloud or edge deployment make maintenance easier, and why?