Natural language processing and computer vision

Natural language processing and computer vision are the two applied branches where machine learning met the messiest data: sentences whose meaning shifts with context, photographs taken in poor light. For a long time both relied on hand-built features and rules. Deep learning replaced most of that by learning features from examples, and transformer models later gave the two fields a largely shared foundation.

How the blueprints frame it

Microsoft's Azure AI Fundamentals used to give vision and language workloads a domain each. The current exam, AI-901, folds them into one workloads area covering text analysis, speech, vision and information extraction, and asks candidates to build small applications on multimodal models in Foundry; the Azure AI Fundamentals page tracks the change of code. AWS's AI Practitioner outline lists vision, NLP and speech recognition among the real applications a candidate should recognise and names managed services for translation, transcription and text analysis. SFIA's machine learning skill does not split the fields out at all. Read together, the blueprints agree: at entry level these are workloads you pick a service for, not models you train. Specialists still earn their keep in four places.

• Building and labelling datasets for a domain general models have not seen, such as defects on one production line or clauses in one kind of contract.
• Fine-tuning or training smaller models where latency, cost or privacy rule out a hosted general model, for example on a camera or a phone.
• Measuring properly: a held-out set that resembles production, error analysis class by class, and checks for performance gaps between groups of people.
• Maintaining everything around the model, from image preprocessing and tokenisation to the rules applied to its output.

Mistakes, tools and next steps

The beginner's mistake is reporting one accuracy figure on a random split. If the same patient, document or camera appears in both training and test data, the score flatters the model; if one class is rare, a model that never predicts it can look excellent. Leakage of this kind is common enough that careful reviewers ask about the split before they ask about the score. Research and production code in both fields runs mostly on PyTorch, with TensorFlow the other framework in wide use, particularly for deployment on devices. The road in runs through machine learning fundamentals. From here it leads to generative AI and LLMs, where prompting now handles many classic language tasks, to MLOps and AI engineering for serving, and to responsible AI, because face analysis and text classification are where complaints about bias tend to land first.