NLP and Computer Vision: What AI Models Solve

Concept

Natural Language Processing (NLP) and computer vision are fields for extracting useful information from language and images. They are common deep-learning applications because raw text and pixels are high-dimensional, variable, and difficult to summarize with a small set of manual features. They are not synonymous with Generative AI, which is introduced in Module 10.

Structured and Unstructured Inputs

A churn table has named columns such as tenure_months and monthly_charge. A review, image, recording, or video begins in a less directly interpretable form. Text contains varying sequences of words and punctuation; an image is an array of pixel values. Useful systems still require careful problem framing, labels where supervised learning is used, valid splits, and appropriate evaluation.

NLP: Working With Language

NLP applies computational methods to language. Common tasks include sentiment classification, spam detection, document categorization, named entity recognition, translation, search/retrieval, summarization, and question answering.

At a high level, a model needs a numeric representation of text:

text → tokens → numeric representations → model output

Tokens are pieces of text processed by a model. Modern systems often map tokens to learned vectors called embeddings, which can encode useful relationships from data. This does not mean the vectors contain perfect meaning, commonsense, or truth. Their behavior depends on training data, model design, and the task.

For example, a support team may classify incoming messages into billing, technical, or account categories. A model can help route messages, but ambiguous wording, new product terms, and unequal class frequencies still require error analysis. It should be evaluated on representative held-out messages rather than judged by a few impressive examples.

Computer Vision: Working With Images

Computer vision extracts useful information from images or video. Three common tasks have different outputs:

Task	Question	Example
Classification	What is in this image?	Is this product damaged?
Detection	What objects are present and where?	Where are defects on a panel?
Segmentation	Which pixels belong to each region?	Which pixels show a road or a tumor?

Images begin as arrays of pixel values. Neural networks can learn representations useful for visual tasks. A convolutional intuition is that local filters can respond to nearby patterns, then later layers combine those patterns. This is only a high-level picture; architecture internals are outside this introductory module.

Transfer Learning

Training a large model from scratch can demand extensive labeled data and compute. Transfer learning starts from a model previously trained on a large relevant dataset, then adapts its learned representations to a smaller related task. A pretrained vision model might be adapted for manufacturing defect classification; a pretrained language model might support domain-specific document classification.

Transfer learning is practical, not automatic. The new data can differ from pretraining data, labels can be noisy, and performance must still be checked on representative held-out cases. A model that performs well on general images may fail on a factory camera with unusual lighting.

Responsible Evaluation Still Applies

Complex deep models do not remove the principles from Model Evaluation. Keep training and test data separate, choose metrics that reflect decision costs, examine class imbalance, prevent leakage, and assess performance after deployment conditions change. Biased training examples, privacy concerns, domain shift, and overconfident errors can be especially consequential in language and vision applications.

For medical imaging assistance, a high overall score may hide weak performance for one patient group or scanner type. For customer-message routing, an incorrect label might delay urgent support. The model's output should support a well-defined workflow, with appropriate human review where the consequences justify it.

Failure Signals

Common Mistakes

Treating any text system as an LLM or any image task as simple classification.
Assuming embeddings provide human-like understanding.
Ignoring deployment differences such as new camera conditions or new language.
Using a pretrained model without validating it on the actual task.
Believing deep learning removes privacy, fairness, or evaluation responsibilities.

Best Practices

Define the task output precisely, collect representative examples, and start with an evaluation plan. Use transfer learning when it is justified by a related source model and a realistic validation setup. Inspect errors by category, context, and affected group rather than relying only on one aggregate score.

Interview Perspective

Question: What is the difference between detection and segmentation?
Answer: Detection locates objects, while segmentation assigns labels to pixels or regions.
What the interviewer is testing: whether you can connect an output requirement to the appropriate task.
Follow-up: Why might transfer learning help with a small labeled image dataset?

Practice Questions

Classify spam filtering, translation, and product-review routing as NLP tasks.
Is identifying a damaged product classification, detection, or segmentation if the location is not needed?
Why must text be represented numerically before a model can process it?
Give one risk of applying a pretrained vision model to a new camera environment.
Which evaluation principle from Module 04 still applies to a neural image model, and why?

Quick Quiz

What are tokens? Answer: Pieces of text processed as model input.
Does object detection only name an image class? Answer: No; it also identifies locations.
Does transfer learning remove the need for validation? Answer: No.

Key Takeaway

Key Takeaways

NLP works with language and computer vision with images or video. Deep learning can learn useful representations for both, but task definition, representative evaluation, transfer risk, and responsible deployment remain essential.

Next Lesson

Next, make a practical choice between classical ML and deep learning based on data, constraints, and validated value.