The baseline model will only take you less than a tenth of the time, but it can derive more than 90% of the results.
Starting from a baseline model, you can first understand the bottleneck of the problem!
Baseline models put complex models in a dilemma
- Default available performance level
- Human performance index
- Deployable performance level
Reasonable automatic generation performance: the result that can be obtained with a relatively simple model. Through this important benchmark value, we can evaluate whether a complex model performs well, and let us no longer struggle between accuracy and complexity.
Baseline models are easier to deploy
benefit:
- Fast training: can quickly get performance feedback.
- Easier to study: This means that most errors encountered can be easier to locate whether it is a defect in the model or an error in the data.
- Quick inference: The deployment of the baseline model does not require too many architectural modifications and does not cause potential risks.
Once you have created and deployed the baseline model, you are at the best decision point for the next step.
The baseline model will let you quickly get performance benchmark results
The baseline model can help you understand the data
If the selected baseline model performs well, then you have helped yourself to solve most of the difficulties in building a complex model; conversely, finding errors in the process of building a baseline model is very constructive for discovering deviations and specific errors in the data.
Many problems that hinder the process of machine learning are often due to failure to understand and prepare the data, rather than choosing a more complex model.
The baseline model can help you understand the task
The baseline model can also help you understand which part of the content is more difficult and which part is easier. In this way, you can also locate which aspect of the model should be improved to better solve the difficult part.
Most machine learning problems follow the " There is no free lunch in the world " theorem: there is no solution that can solve all problems. The real challenge is to choose from various architectures, determine the appropriate strategy, and choose the model that is most suitable for extracting and utilizing the target data structure.
When not to choose a baseline model
For some tasks, it is indeed difficult to construct an effective baseline model. If you try to separate the speech content of different people in a recording ( cocktail effect ), you may have to start with a complex model to get satisfactory results.
In such cases, instead of simplifying the model, it is more appropriate to simplify the data. In other words, let this complex model overfit a very small part of the existing data set. If the performance of the model is strong enough, then this should be easy; but if the performance is not satisfactory, then maybe you need to try another model.
in conclusion
People have a tendency to ignore simple solutions if they have more powerful solutions. But in most areas of machine learning, starting from the basics is often more valuable .
Although learning to implement complex models is indeed difficult, the biggest challenge for machine learning engineers is actually how to choose a modeling strategy for a specific task .
Using a simple model first can help to a large extent; if the performance is not satisfactory, then when using a more complex model, you can avoid the problems that have been found in the baseline model, so as to achieve better modeling effect.
The baseline model represents the control group, the baseline, which is the model to be compared in the future.