You now know what deep learning is (lesson 01-01) and how it got here (lesson 01-02). Before diving into the internal mechanics of neural networks, it helps to have a map of the territory: what can be done with deep learning today, and in which sectors is it genuinely working? This lesson tours the major application domains and, for each one, answers two practical questions: what would TecnoMarket need from that domain? And in which module of the course will you learn the corresponding technique? By the end, you will see that this course is not a collection of loose topics, but the complete toolbox for the TecnoMarket project.

1. How to read this overview
2. Computer vision
3. Natural language processing (NLP)
4. Audio and speech
5. Recommender systems
6. Healthcare and medicine
7. Autonomous vehicles
8. Fraud and anomaly detection
9. Map: domains, TecnoMarket and course modules

For each domain we will look at three things: which tasks deep learning solves, real-world examples, and its connection to TecnoMarket and the course. Important: we do not explain how each architecture works internally here (CNNs, RNNs, transformers, GANs...); that belongs to modules 3, 4 and 5. This lesson is the "showroom"; the later modules are the "workshop".

One pattern you will see repeated: almost every domain boils down to a few families of tasks:

• Classify: assign a label to an input (is this photo a coffee maker? Is this review positive?).
• Detect/locate: find where something is (where is the defect in this part?).
• Predict values: estimate a number (how many units will we sell tomorrow?).
• Generate: create new content (a promotional image, a chatbot's response).
• Transform sequences: convert one sequence into another (audio → text, Spanish → English).

This is the domain where deep learning struck its great blow (AlexNet, 2012) and probably the most mature one.

Typical tasks:

• Image classification: saying what a photo contains.
• Object detection: locating and labeling several objects in an image (with their bounding boxes).
• Segmentation: delineating each object pixel by pixel.
• Facial recognition and identity verification.
• OCR: reading text in images (receipts, invoices, license plates).
• Industrial visual inspection: detecting manufacturing defects.

Real-world examples: face unlock on your phone, social media filters, satellite image analysis, quality control in factories.

At TecnoMarket: this is the most urgent need. Sellers upload thousands of product photos a day, and today a human team reviews and categorizes them by hand. With computer vision, TecnoMarket could automatically classify each photo into its category (computing, small appliances, home...), detect low-quality photos or ones that break the rules (watermarks, inappropriate backgrounds) and automatically read the text on product boxes.

In the course: all of module 3 (CNNs), with the guided project in lesson 07-01. We will prototype with MNIST and CIFAR-10 before jumping to TecnoMarket's fictional catalog.

NLP applies deep learning to text: written human language.

Typical tasks:

• Sentiment analysis: is this opinion positive, negative or neutral?
• Text classification: assigning topics or categories (is this inquiry about shipping or returns?).
• Machine translation between languages.
• Automatic summarization of long documents.
• Chatbots and conversational assistants.
• Information extraction: pulling structured data out of free text (names, dates, amounts).

Real-world examples: Google Translate, keyboard spell-checkers and autocomplete, assistants like ChatGPT or Claude, spam filtering.

At TecnoMarket: the store receives hundreds of reviews a day in several languages. With NLP it could automatically measure satisfaction per product (sentiment analysis), detect reviews that mention recurring defects ("the battery doesn't last") to alert purchasing, classify support tickets and route them to the right department, and eventually offer a customer service chatbot.

In the course: module 4 (RNNs and NLP applications, with the IMDB movie review dataset as the test bench), lesson 05-05 (attention and transformers, the foundation of today's language models) and the text generation project in 07-02.

Audio is another unstructured signal where deep learning dominates.

Typical tasks:

• Speech recognition (speech-to-text): transcribing audio to text.
• Speech synthesis (text-to-speech): generating natural-sounding speech from text.
• Speaker identification: who is talking?
• Sound classification: detecting alarms, breaking glass, music.

Real-world examples: Alexa, Siri and Google Assistant; YouTube's automatic captions; meeting transcriptions.

At TecnoMarket: the call center records (with consent) customer service calls. Transcribing them automatically would allow later analysis with NLP: the most frequent reasons for calling, dissatisfied customers, script compliance. It is a good example of model chaining: speech → text (audio) and then text → conclusions (NLP).

In the course: there is no dedicated audio module, but the underlying techniques are the same ones we will study: audio is processed as sequences (module 4) or as spectrogram "images" (module 3). In professional practice, moreover, speech is usually solved with pretrained models (transfer learning, lesson 05-03).

Recommenders decide which content or product to show each user.

Typical tasks:

• Product recommendation: "customers who bought this also bought...".
• Personalization of the home page, emails and promotions.
• Ranking: ordering search results by relevance for that user.

Real-world examples: Netflix (which series to suggest), Spotify (personalized playlists), Amazon (related products), social media feeds.

At TecnoMarket: the recommender is money, plain and simple: showing the right product to the right person increases conversion. Deep learning makes it possible to combine heterogeneous signals — purchase history, browsing, review text, even product images — into a single representation of the customer's taste.

In the course: it has no module of its own, but it rests on pieces we will cover: learned representations (embeddings, which will appear in module 4 with text) and dense networks (module 2). It is an excellent application to explore on your own once you finish the course.

One of the domains with the greatest social impact.

Typical tasks:

• Medical imaging diagnosis: detecting tumors in mammograms, diabetic retinopathy in retinal scans, pneumonia in X-rays.
• Drug discovery: predicting molecular properties; the AlphaFold case (protein structure prediction) is a recent scientific milestone.
• Patient monitoring: detecting arrhythmias in ECG signals, predicting deterioration in the ICU.

Real-world examples: retinopathy screening systems approved for clinical use, support for radiologists in cancer detection.

At TecnoMarket: directly, not much: it is not their sector. But we include it for two reasons. First: technically, diagnosing an X-ray and classifying the photo of a coffee maker are the same problem (image classification, module 3); mastering one brings you closer to the other. Second: healthcare illustrates better than any other domain the ethical demands of deep learning (mistakes with human cost, biases, explainability), which we will address in module 8.

In the course: module 3 (the technique) and module 8 (the ethical implications).

Autonomous driving is perhaps the most complex deep learning system in production.

Components where DL is involved:

• Perception: detecting pedestrians, vehicles, signs and lanes from cameras, radar and lidar (real-time computer vision).
• Prediction: anticipating what others will do (is that pedestrian going to cross?).
• Planning: deciding the trajectory (here DL is combined with classical control algorithms).

Real-world examples: the driver-assistance systems (automatic emergency braking, lane keeping) that most new cars already carry, and robotaxis in some cities.

At TecnoMarket: it is not going to build cars, but last-mile logistics resembles this more than it seems: warehouse robots navigating between shelves and, in the future, autonomous delivery. Moreover, a warehouse robot's perception uses the same CNNs from module 3.

In the course: module 3 (visual perception) and module 8 (the ethical and liability dilemmas, where autonomous vehicles are the classic case study).

Spotting the odd one out among millions of normal events.

Typical tasks:

• Payment fraud: transactions with stolen cards, compromised accounts.
• Anomaly detection: behaviors that deviate from normal, without knowing in advance what the attack will look like.
• Cybersecurity: malicious network traffic, suspicious logins.
• Fake reviews and accounts: bots that inflate ratings.

The peculiarity of this domain: fraud is rare (maybe 1 in every 1,000 transactions) and ever-changing (fraudsters adapt). That is why, in addition to classifiers trained on known fraud, techniques are used that learn "what normal looks like" and raise an alert when something falls outside the pattern, without needing fraud examples.

At TecnoMarket: thousands of payments are processed every day. A detection system should flag suspicious transactions in real time (unusual amounts, new addresses, strange purchase patterns) for human review, and also detect fake reviews that manipulate product ratings.

In the course: autoencoders for anomaly detection (lesson 05-02) and the complete project in lesson 07-03, where we will build TecnoMarket's anomalous transaction detector.

This table is the course's "treasure map": keep it as a reference.

Two final observations about the landscape:

• The same pieces are reused: notice how many domains rest on module 3 or module 4. Deep learning is modular: a few fundamental ideas, an enormous number of applications.
• Models are chained together: real systems combine several models (speech → text → sentiment; photo → category → recommendation). Think in terms of pipelines, not isolated models.

• Mistake: believing that each application needs a completely different technology. As the table shows, a handful of architectures (CNNs, RNNs, transformers, autoencoders, GANs) covers almost every domain. Learning the fundamentals well pays off everywhere.
• Mistake: choosing a project for how spectacular it is rather than for its return. For TecnoMarket, a "boring" photo classifier that saves 200 hours of manual work a month is worth more than a dazzling chatbot nobody asked for. Prioritize by business value.
• Mistake: ignoring the cost of errors. Getting a coffee maker recommendation wrong is not the same as getting fraud detection wrong (or a medical diagnosis). The domain determines how much accuracy you need and how much human oversight you must keep.
• Tip: when you read about a product "with AI", try to break it down into the basic tasks from this lesson (classify, detect, predict, generate, transform sequences). You will see that almost everything fits that scheme, and you will know which course module covers it.
• Tip: revisit this lesson at the start of each course module: it will remind you what the thing you are about to learn is for, out in the real world.

For each need, indicate the application domain and the type of task (classify, detect, predict a value, generate or transform sequences):

1. TecnoMarket wants to estimate how many units of each TV it will sell next week.
2. TecnoMarket wants a system that converts recorded call center calls into text.
3. TecnoMarket wants to create Christmas promotional images from its catalog photos.
4. TecnoMarket wants to know whether a freshly posted review is genuine or generated by a bot.

TecnoMarket's management can only fund two deep learning projects this year: (a) automatic product photo classifier, (b) conversational customer service chatbot, (c) fraudulent transaction detector, (d) promotional image generator. Choose two and justify your choice considering: business value, maturity of the technique, data availability and cost of errors.

Design (in words or with a diagram, no code) the model pipeline TecnoMarket would need to: "automatically detect products whose recent reviews mention defects, and alert the purchasing department". Indicate which domains from this lesson are involved and in what order.

Solution 1:

1. Time series / demand forecasting — a predict a value task (regression on sequences). Covered in module 4.
2. Audio and speech — a transform sequences task (audio → text).
3. Content generation (generative vision) — a generate task. Covered with GANs (05-01, 07-04).
4. NLP + anomaly/fraud detection — a classify task (genuine/fake), possibly supported by anomaly detection if there are few labeled examples of fake reviews.

Solution 2 (suggested answer; what matters is the justification):

A defensible choice is (a) the photo classifier and (c) the fraud detector:

• (a) has immediate value (it saves daily manual review), the technique is very mature (image classification, the best-solved problem in DL), the data already exists (a history of photos already categorized by humans, which serve as labels) and the cost of errors is low (a misclassified photo is easily corrected).
• (c) has a direct monetary return (every fraud prevented is a loss avoided), and although the cost of errors is higher, it can be mitigated by keeping human review of flagged cases.
• (b) the chatbot is attractive but risky: open-ended conversation with customers, highly visible mistakes, and it requires more team maturity. (d) the image generator adds value, but it is less critical to operations and can be covered in the meantime with commercial tools.

Solution 3 (suggested answer):

1. Collection: each day, new reviews for each product are gathered (text).
2. NLP - classification 1: a sentiment analysis model filters out the negative reviews.
3. NLP - classification/extraction 2: on the negative ones, a second model detects whether they mention a product defect (and of what kind: battery, screen, damage in shipping...), distinguishing it from complaints unrelated to the product (a courier delay).
4. Aggregation and anomaly: defect mentions are counted per product and week; if a product deviates from its normal pattern (a spike in defects), an alert is generated.
5. Action: the alert reaches purchasing with the example reviews attached.

Domains involved: NLP (steps 2-3) and anomaly detection (step 4). It is a pipeline: several simple models chained together, not a single giant model.

In this lesson you have seen the full landscape: computer vision, NLP, speech, recommendation, healthcare, autonomous driving and fraud, and you have confirmed two key ideas: almost all applications boil down to a few fundamental tasks, and a few architectures — the ones you will learn in modules 3 through 5 — hold up the entire building. You also now have the map connecting each TecnoMarket need to the course module that solves it.

So far we have talked about deep learning "from the outside". In the next lesson we will open the box for the first time: we will see what an artificial neuron is, how neurons are organized into layers and what "training" a network really means — the essential vocabulary you will use for the rest of the course.