AI0-001 AI Models and Data Engineering Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  K-means clustering

  Why it's wrong here

  K-means is an unsupervised clustering algorithm, not a classification model. It groups data points into clusters based on similarity, but it does not use labels to learn a mapping from features to classes. In this scenario, the task is supervised classification of handwritten digits, so K-means is not applicable. Even if used for feature extraction, it would not directly provide class predictions. Therefore, K-means is not the appropriate model here.

• ✗

  Decision tree

  Why it's wrong here

  Decision trees partition the feature space based on simple rules, but they struggle with high-dimensional data like images. Each pixel is a feature, leading to a very large tree that is prone to overfitting. Decision trees do not capture spatial relationships between pixels effectively. While ensemble methods like random forests can improve performance, they still require significant feature engineering and are generally outperformed by CNNs on image tasks. Thus, a decision tree is not suitable for this scenario.

• ✓

  Convolutional neural network (CNN)

  Why this is correct

  Convolutional neural networks are specifically designed for image data. They use convolutional layers to automatically learn spatial hierarchies of features, such as edges, textures, and shapes, from raw pixel values. This makes them highly effective for image classification tasks like handwritten digit recognition. CNNs also benefit from parameter sharing and local connectivity, reducing the number of parameters compared to fully connected networks. Given the image size and dataset, a CNN can achieve high accuracy with reasonable computational resources. Therefore, a CNN is the most suitable model.

• ✗

  Logistic regression

  Why it's wrong here

  Logistic regression is a linear model that can be used for classification, but it requires manual feature engineering to perform well on image data. Raw pixel values are not linearly separable for complex patterns like handwritten digits. While logistic regression can achieve decent accuracy on MNIST with preprocessing, it cannot automatically extract features and typically underperforms compared to CNNs. In this scenario, where the goal is high accuracy and automatic feature extraction, logistic regression is not the best choice.