Downstream Task is the target task that a pre-trained model is applied to after self-supervised or supervised pre-training — used to evaluate the quality of learned representations and measure how well the pre-trained features transfer to practical applications.

What Is a Downstream Task?

• Examples: Image classification (ImageNet), object detection (COCO), semantic segmentation (ADE20K), action recognition, medical imaging.
• Evaluation Protocol: Freeze pre-trained backbone -> train a task-specific head (linear probe or fine-tuning).
• Metric: Performance on the downstream task benchmarks the representation quality.

Why It Matters

• Representation Benchmark: Downstream task performance is the ultimate test of self-supervised learning methods.
• Transfer Learning: Good representations transfer to many downstream tasks, even with limited labeled data.
• Practical Value: The pre-trained model's usefulness is entirely determined by how well it performs on real downstream tasks.

Downstream Task is the final exam for pre-trained models — the real-world challenge that determines whether the learned representations are actually useful.