Pt.3 Machine Learning Challenges

Main Challenges of Machine Learning

The insufficient quantity of training data

For very simple problems in machine learning, a simple model, or a linear regression model, would need a lot of data to perform to the best of its capabilities. For typical problems, the model might need thousands of examples, and for complex problems that detect items in images or speeches take millions!

Some machine learning researchers had figured out that data and the way you manipulate that data is more important than the neural network or learning algorithm. Small and medium-sized datasets are still very common and can achieve the same effect.

Non-representative training data

In order to generalize well, we need to make sure that our data and model (learning algorithm) are good enough to make predictions, especially for poor or rich countries.

By using a better representative training sample, the one here, we make sure to take care of the other countries and be less incorrect.

Poor quality data

If some instances are clearly outliers in the data and make no sense, or there is any noise (which means randomness), then the best thing is to clean up the training data. The truth is, data scientists and Machine Learning Engineers spend a lot of their time fixing their data.

underfitting the training data

Underfitting is the exact opposite of overfitting, it occurs when you can make a model to simple to learn the true structure within data. For example, a linear model is extremely bad as it will always produce an error due to it not being able to prioritize clusters! Sometimes we will have to choose a more powerful model.

Overfitting the training data

Overgeneralizing is something we do too often, and an AI model in most cases does the same. Complex models can detect subtle patterns in the data, but if the training data has a lot of features, it might generalize some random detail. It can think that Zimbabwe and Norway have the same life satisfaction index because of the 'w'.

irrelevant features

A very popular phrase in machine learning is the phrase garbage in, garbage out. You have to make sure that when you are looking at the features you need, you only pick the most useful ones, and you can combine features into one (dimensionality reduction algorithms).

Multiple Choice

What is the purpose of a training dataset?

• A training dataset is used to evaluate the performance of a machine learning model.

• A training dataset helps to identify where the model goes wrong.

• A training dataset is used to train a machine learning model by testing it by giving mangos.

• A training dataset provides the input features and corresponding output labels necessary for a machine learning algorithm to learn from and make predictions.

Multiple Choice

What is the purpose of a testing dataset?

• To train the machine learning model.

• To evaluate the performance and generalization of the model.

• To preprocess and clean the data.

• To visualize the data.

Multiple Choice

What is the purpose of a validation dataset?

• To assess the performance of a model on unseen data. To estimate the model's accuracy and compare it with the training results.

• To fine-tune hyperparameters and optimize model performance.

• To test the model.

• To hire a kangaroo.