Dataset bias in vision in Computer Vision - Model Pipeline Trace

Model Pipeline - Dataset bias in vision

This pipeline shows how dataset bias in vision can affect model training and predictions. It demonstrates how biased data leads to skewed learning and poor generalization.

Data Flow - 5 Stages

1Raw Image Dataset

1000 images x 64x64 pixels x 3 channels→Collect images mostly of cats in indoor settings→1000 images x 64x64 pixels x 3 channels

Image of a cat sitting on a couch indoors

↓

2Data Preprocessing

1000 images x 64x64 pixels x 3 channels→Resize images to 32x32 pixels and normalize pixel values→1000 images x 32x32 pixels x 3 channels

Resized and normalized cat image

↓

3Train/Test Split

1000 images x 32x32 pixels x 3 channels→Split dataset into 800 training and 200 testing images→800 training images x 32x32 pixels x 3 channels, 200 testing images x 32x32 pixels x 3 channels

Training set mostly indoor cat images, test set includes some outdoor cat images

↓

4Model Training

800 training images x 32x32 pixels x 3 channels→Train CNN to classify cat images→Trained CNN model

Model learns mostly indoor cat features

↓

5Model Evaluation

200 testing images x 32x32 pixels x 3 channels→Evaluate model accuracy on test images→Accuracy score (percentage)

High accuracy on indoor cats, low accuracy on outdoor cats

Training Trace - Epoch by Epoch

Loss
1.2 |****
0.9 |***
0.7 |**
0.6 |*
0.55|*
    +------------
     Epochs 1-5

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.45	Model starts learning basic features
2	0.9	0.60	Model improves recognizing indoor cats
3	0.7	0.72	Model fits well to biased training data
4	0.6	0.78	Loss decreases steadily, accuracy increases
5	0.55	0.82	Model converges on training data with bias

Prediction Trace - 5 Layers

Layer 1: Input Image

Layer 2: Convolutional Layer

Layer 3: Pooling Layer

Layer 4: Fully Connected Layer

Layer 5: Output Layer with Softmax

Model Quiz - 3 Questions

What is a main cause of dataset bias in this vision model?

A Most training images show cats indoors

B Images are resized to 32x32 pixels

C The model uses ReLU activation

D Test images are fewer than training images

Key Insight

Dataset bias in vision models causes them to learn features mostly from the dominant data type, reducing accuracy on underrepresented cases. Careful dataset design and diverse data collection are essential to build fair and robust vision models.