Aerial Cactus Identification

What I Explored

I explored binary image classification through Kaggle's Aerial Cactus Identification challenge, where the objective is to determine whether an aerial image contains a cactus.

The project helped reinforce the complete supervised image-classification workflow: preparing image data, constructing a CNN-based classifier, training the model, evaluating predictions, and generating inference results.

What Confused Me

A key question was how much information a neural network actually needs to distinguish two visually similar categories.

It also highlighted the difference between memorizing training examples and learning visual features that generalize to unseen images.

What Clicked

I developed a stronger understanding of hierarchical feature learning.

Early convolution layers detect relatively simple patterns such as edges and textures, while deeper layers combine those features into increasingly meaningful visual representations.

My Mental Model Now

For image classification, I think of a CNN as progressively transforming:

Pixels → Edges → Textures → Shapes → Semantic Features → Class Probability

The network does not explicitly receive rules describing what a cactus looks like. Instead, it learns useful visual patterns through optimization.

Implementation / Example

I created a notebook for the Kaggle Aerial Cactus Identification competition and trained and saved a model for cactus-presence classification.

Repository: Aerial Cactus Identification

What I Want to Explore Next

I want to compare custom CNN architectures with transfer learning, experiment with stronger augmentation, and study how pretrained representations affect performance when training data is limited.