Author(s): Nehdiii

Originally published on Towards AI.

From recognizing faces in photos to detecting objects in real-time videos, computer vision has revolutionized the way machines “see” the world. Tasks like image classification, object detection, segmentation, and even person re-identification (ReID) have seen massive progress thanks to deep learning.

But these advances came with a cost:

Massive datasets need to be collected and annotated.Training models from scratch takes a lot of time and compute.Task-specific fine-tuning limits generalization.

To overcome these hurdles, researchers introduced a new paradigm: Pre-train a model on large-scale data → Fine-tune it for specific tasks.

While this helped, it still relied heavily on labeled data for each task.

Then came a shift sparked by advances in natural language processing (NLP). What if we could train models on image-text pairs and let them generalize across tasks with no fine-tuning at all?

Different Learning Paradigms

Inspired by language models like BERT and GPT, a new class of models emerged: Vision-Language Models (VLMs).

These models are trained on large datasets of image-text pairs. And instead of task-specific tuning, they aim to understand the alignment between visual and textual modalities.

The result? A single model that can:

Classify images without training on any labels (zero-shot classification),Retrieve images based on text,Understand complex visual semantics.

One VLM made headlines in 2021: CLIP… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI