Ml Production Series establish Model Baseline 3
ML Production Series — Establish Model Baseline 3
In this tutorial, we are gonna discuss the best practices for establishing a model baseline. It is important as ML is an iterative process and continuous improvement of the baseline helps to make a high-quality application
Example
Let us understand this concept with a speech recognition system as shown in the pic (fig-1) below, considering we have four scenarios of different types and the respective accuracy of the model in a particular category.
Analysis for — comparison with human-level performance
consider we first think: of establishing a baseline where we want to improve people’s noise and low bandwidth audio accuracy
Fig-1
second thought: it is required to compare our algorithm performance with human-level performance to make a decision for establishing a baseline as In the above picture you can see let’s say you hired humans to do the transcription and you got the above-mentioned accuracy of human-level performance.
Conclusion — we came to make us decide to improve people’s noise and low bandwidth level performance as humans also cannot get much accuracy, instead focus our attention to improve Car Noise level performance
Scenario : Unstructured Data and Structured Data
As described above the establishing basline- there can be a scenario in which data for modeling can be structured or unstructured — as seen in fig-2 below unstructured data include image, audio, text — in which human level performance comparison is best practice to establish a baseline- whereas structured data might include Inverntory system database or other application based databases which also includes excel sheets or spreadsheets in different format contains lot of columns in which humans can have difficulty to make a decisions
fig-2
Now the Question is How to establish baseline for unstructured data?
As fig-3 describes the ways to establish the baseline performance — for unstructured data Literature comparison is one of the best way to establish a basline in which you compare the performance of your model with state of the art.
Another way — is to make a first version of your model and do the continuous improvements in which you will compare the older implementation of the system
fig-3
Benefit of establishing baseline — is to find out the irriducable error.
Real World Problem : Companies and AI team conflict
Andrew Ng said — he have seen some scenarios in which companies with AI teams have a discussion to reach a particular accuracy. for example client ask you to reach accuracy of 80 percent at least. In such cases AI team should ask them to make a baseline before deciding about error tolerance or system performance commitments.