24-679-distilbert-lighting

24-679: Light Type Classifier

This model classifies texts as describing Natural Lighting (0) or Artificial Lighting (1). It fine-tunes a pretrained DistilBERT with AutoGluon MultiModal. This is a homework assignment.

Inputs and outputs

Texts that are labeled as indicated below:

Data and training

Source: Lighting Dataset.

Evaluation

Confusion matrix: rows are actual labels; columns are predicted labels.

Each changed prediction moves accuracy by about 6.67 percentage points.

Brief Error Analysis

There were 6 misclassified text cases:

1. Flicking the blinker on my car turns on the high beam, which will illuminate with a radius many times the width and length of the car, lighting down the road. Turning it off leaves the immediate surroundings dark in the night.

• True: Artificial Light
• Predicted: Natural Light

1. The room is dark, but excessive hallway light creeps into the kitchen. The kitchen's extremities are dark, illuminated only by the reflection at the doorstep.

• True: Artificial Light
• Predicted: Natural Light

1. Clicking the button on my car keys makes the light source illuminate orange briefly. At night, you can see the flash radiate a few meters away from the car, but it is instantaneous, returning the immediate surroundings to darkness once it leaves.

• True: Artificial Light
• Predicted: Natural Light

1. The tiny dots scatter across the night sky. They are faint and they don't illuminate the Earth's surface. You can see them from a distance, whether they are a star or a planet.

• True: Natural Light
• Predicted: Artificial Light

1. There is a visible glare as you step outside after being inside for an extended period of time. You walk out and you can't help but squint at this light source, especially.

• True: Natural Light
• Predicted: Artificial Light

1. The light source is so strong it peers through the thin leaves to make them translucent; the leaves themselves look illuminated in the afternoon.

• True: Natural Light
• Predicted: Artificial Light

Some of these are somewhat understandable errors as the:

• Falsely predicted Natural Lighting errors all contain something about the night (moon) or illumination at doorstep (outdoors)
• Falsely predicted Artificial Lighting don't seem to have a trend

Licensing and provenance

Maintainer: OmarK211. The repository had no model license specified when this card was added; this card does not assign one. Refer to the dataset card for image collection, consent, and reuse limitations. Removing camera metadata does not remove identifying details visible in an image.