As in the attached diagram in the image below, it shows the amount of data and performance in using traditional methods or deep learning.
What are the limits to the amount of data with which we can make the decision to use traditional methods or deep learning?
Depends on what problem you are trying to solve. Does it require 100 variables to predict in a way that is expected? The decision of using deep learning or other probabilistic modelling techniques should come from this line of thinking. Example:
You want to develop a prediction algorithm for assessing the wear on car tires. You try to estimate the frequency of "bad tires" or "too much wear" events in normal every day use. You estimate that you want your algorithm to have small false positive rate (80%), because more than often the data that is used for model input is of good tires. If you can get to this performance with one measurement variable, traditional statistical models are a good fit. If you can only get to this performance with 100-300 measured variables, deep learning would be a better fit.
In the following articles, it is illustrated that deep learning methods can be used even if the data set is small:
https://beamandrew.github.io/deeplearning/2017/06/04/deep_learning_works.html
https://towardsdatascience.com/does-deep-learning-really-require-big-data-no-13890b014ded
https://medium.com/@snaily16/is-deep-learning-useless-if-there-is-no-large-dataset-available-no-eac26f898968
Also, you can see,
https://medium.com/merantix/applying-deep-learning-to-real-world-problems-ba2d86ac5837
https://medium.com/@ODSC/how-to-choose-machine-learning-or-deep-learning-for-your-business-bea30f0d518f
https://towardsdatascience.com/how-do-you-know-you-have-enough-training-data-ad9b1fd679ee
- Badges
- Science topic
- Similar topics
- Computer Science
- Images