Introduction

When You Want to Easily Explain

Here are 3 examples of when you would have trouble explaining deep learning:

• When you want to describe the top features of your model — the features become hidden inputs, so you will not know what caused a certain prediction to happen, and if you need to prove to stakeholders or customers why a certain output was achieved, it can be more of a black box
• When you want to tune your hyperparameters like learning rate and batch size
• When you want to explain how the algorithm works itself — for example, if you were to present the algorithm itself to stakeholders, they might get lost, because even a simplified approach is still difficult to understand

Here are 3 examples of how you could explain those same situations from above from non-deep learning algorithms:

• When you want to explain your top features, you can easily access SHAP libraries, say for the algorithm CatBoost, once your model is fitted, you can simply make a summary plot from feat = model.get_feature_importance() and then use the summary_plot() to rank the features by feature name, so that you can present a nice plot to stakeholders (and yourself for that matter)

• As a solution, some other algorithms have made it plenty easy to tune your hyperparameters by randomized grid search or a more structured, set grid search method. There are even some algorithms that tune themselves so you do not have to worry about complicated tuning
• Explaining how other algorithms work can be a lot easier, like decision trees, for example, you can easily show a yes or no, 0/1 chart that shows a simple answer for features that lead to a prediction, like yes it is raining, yes it is winter, would provide for yes it is going to be cold

When You Can Use Other Algorithms

Here are examples of why you would want to use a non-deep learning algorithm, becuase you have so many other, simpler, non-deep learning options:

• They can be easier and faster to set up, for example, deep learning can require you to have your model add sequential, dense layers, and compile it, which can be more complex, and take longer than simply having a regressor or classifier and fitting it with non-deep learning algorithms
• I personally find more errors that can result from this more complex deep learning code and documentation for how to fix it can be confusing or old and not be applicable, using an algorithm like Random Forest instead, can have much more documentation on errors that are easy to understand
• Training on a deep learning algorithm may not be complicated sometimes, but when predicting from an endpoint, it might be confusing on how to feed values to predict on, whereas some models, you can simply have the values in an encoded list of ordered values

When You Have a Small Dataset and Budget

Here is when you should second guess using a deep learning algorthim based on costs and data availability:

• Small data availability is usually the case for a lot of companies (but is not always the case), and deep learning performs better on information with a lot of data
• You might be performing a one-off task, as in the model only predicts one time — and you can run it locally for free (not all models will be running in production frequently), like a simple Decision Tree Classifier. It might not be worth investing time in a deep learning model.
• Your company is interested in data science applications but wants to keep the budget small, rather than perform costly executions from a deep learning model, and rather, use a tree-based model with early-stopping-rounds to prevent overfitting, shorten training time, and ultimately reduce costs

Summary

Here are three times/reasons when you should not use deep learning:

* When You Want to Easily Explain
* When You Can Use Other Algorithms
* When You Have Small Dataset and Budget

References