Teaching "How Computers See" to Architect mode (ages 15–18)

Hook & Warm-Up

This age band's hook assumes real technical vocabulary. Read it as-is — it's pitched correctly for this group — then use it to set expectations for the level of the lesson.

"From convolutional neural networks to vision transformers, computer vision has progressed from hand-crafted feature detectors to end-to-end learned representations. Let's examine the pipeline from raw pixels to semantic understanding."

Add context before diving in: "The lesson itself takes maybe 10 minutes in the app. What I want to add today is the history and the 'why' behind a few of those terms, because 'hand-crafted feature detectors' versus 'end-to-end learned representations' is describing an actual shift in how this field works, not just jargon." If any students are considering CS, data science, or engineering paths, mention that computer vision is one of the more mature, well-funded subfields of applied ML — a genuine career on-ramp, not just a classroom topic.

Briefly unpack the two phrases in the hook before moving on, since they set up the whole lesson: "hand-crafted feature detectors" means a human engineer decided in advance what to look for — for example, writing code that explicitly searches for edges at certain angles, or for a specific pattern of light and dark that tends to indicate an eye. "End-to-end learned representations" means the system is instead given a large set of labeled example images and a general-purpose architecture, and it discovers on its own — through training — which patterns in the pixel data are useful for the task, without an engineer specifying what those patterns should look like. Ask students to predict one advantage and one disadvantage of each approach before you cover the rest of the lesson; most will correctly guess that hand-crafted approaches are more predictable but miss unanticipated cases, while learned approaches generalize better but are harder to fully explain or debug — a real, current tension in the field known broadly as AI interpretability.

Main Activity

The in-app scenes here are the same three as every other age band (Pixels, Face Detection, Object Recognition) — but frame each one as a step toward the concepts in this age band's actual quiz, since that's the level you're building toward.

Scene 1 — Pixels as tensors

State plainly: a digital image is a grid of numbers — each pixel typically three values (red, green, blue intensity), so a color image is really a 3D array of numbers (height × width × 3 channels). In ML terms this array is a tensor, and it's the raw input every vision model starts from. Nothing about "recognizing" anything exists yet at this stage; it's just structured numeric data.

Worth noting for students who will go on to actually build something: before that tensor reaches a model, it's normally preprocessed — resized to a fixed dimension the model expects, and its pixel values rescaled (commonly from the 0–255 range into something like 0–1) because neural networks train more reliably on small, well-scaled numbers than on large or wildly varying ones. This is unglamorous but essential — a huge share of real-world computer vision bugs trace back to mismatched preprocessing between training and deployment, not to the model itself.

Scene 2 — From CNNs to attention

Explain the historical arc, matching the hook's framing: for years, the dominant architecture was the convolutional neural network (CNN) — small filters slide across the image detecting local patterns (edges, then textures, then shapes, then objects), a hierarchy of increasingly complex features, layer by layer. More recently, Vision Transformers (ViT) took a different approach borrowed from language models: split the image into patches (like tiles) and apply self-attention, letting the model learn which patches of the image are relevant to each other regardless of distance, rather than only combining nearby pixels the way convolution does. Both approaches are trained on labeled data and both remain in active use — ViT is not a strict replacement for CNNs, and the choice depends on the task, data size, and compute budget.

Connect to face recognition using the base lesson's language: a system measures geometric relationships between facial landmarks (eye spacing, nose width, jaw contour) to build a numeric biometric template compared at each unlock attempt — a task that predates and coexists with the CNN/ViT story, often using specialized architectures rather than general-purpose classifiers.

It's worth being precise about a distinction that trips up even confident students: classification (what is the single dominant object in this image?), object detection (what objects are present, and where, typically drawn as bounding boxes?), and segmentation (which exact pixels belong to which object, down to the outline?) are three different tasks, not interchangeable terms. A self-driving car needs detection at minimum, and increasingly segmentation, to know precisely where a pedestrian's body actually is rather than just a rough box around them. Face unlock is closer to a specialized form of verification (does this face match one specific stored template?) rather than general classification (what/who is this, out of many possibilities?) — a meaningfully different, usually easier, problem.

Scene 3 — Deployment realities: YOLO, transfer learning, data augmentation

Use the object-recognition examples (self-driving cars, medical imaging, cashier-less retail, plant ID) as the jumping-off point for three deployment concepts this age band's quiz actually tests:

• YOLO ("You Only Look Once"): a family of object-detection models built for real-time speed — it performs detection in a single forward pass through the network, rather than scanning a region, checking it, and repeating for many candidate regions. This single-pass design is why it's viable for something like a self-driving car that must detect pedestrians in milliseconds.
• Transfer learning: rather than training a new model from scratch (which needs enormous labeled datasets and compute), teams reuse the weights of a model already trained on a large general dataset and fine-tune it on a smaller, task-specific dataset. This is how a plant-ID app or a medical-imaging startup can build a working model without Google- or Tesla-scale data.
• Data augmentation: artificially expanding a training set by applying transformations to existing images — rotating, flipping, adjusting brightness or contrast, cropping — so the model generalizes better and doesn't overfit to the exact conditions of its original training photos.

Have students complete the lesson individually, then briefly discuss which of the three concepts above they think would matter most for a hypothetical Philippine agritech startup building a rice-pest identification app on a limited budget (expected answer: transfer learning and data augmentation, since neither requires massive proprietary datasets).

Discussion

• "What's the practical tradeoff between a CNN's local convolution and a Vision Transformer's self-attention across the whole image?"
• "Why does a real-time application like a self-driving car need something architecturally different (like YOLO) from a plant-ID app that can take a couple of seconds to respond?"
• "Transfer learning lets a small team build on a large lab's pretrained model. What are the risks of inheriting biases or blind spots baked into that original model?"
• "Data augmentation can't invent information that was never in the training set — it can only vary what's already there. What kinds of real-world variation could augmentation NOT fix?"
• "If you were advising a hospital about deploying an AI system to flag possible tumors in scans, what would you want to know about its training data before trusting it?"
• "Classification, detection, and segmentation are three different computer vision tasks with different levels of difficulty and precision. If you were scoping a computer vision feature for a product, how would you decide which one you actually need — and why might picking a harder task than necessary be a costly mistake?"

These questions are meant to be argued, not just answered — this age band should leave able to defend a technical opinion with a specific reason, not just recall a vocabulary term. Push for "why" on every answer, and don't let a plausible-sounding buzzword substitute for an actual mechanism.

Quiz Walkthrough

This age band's quiz is entirely its own override, distinct from the base lesson — it assumes the vocabulary covered above (ViT, transfer learning, YOLO, data augmentation) and each question includes at least one deliberately plausible-sounding wrong answer, so walking through the reasoning matters more than just confirming the correct letter.

Wrap-Up & Extension

Close with the age band's own summary: "Modern CV uses ResNet for classification, YOLO for real-time detection, and Vision Transformers for attention-based processing. Transfer learning lets you reuse pre-trained models on smaller datasets. Data augmentation improves generalization." Add: "None of these ideas require a research lab budget to use — transfer learning and augmentation specifically exist so smaller teams, including ones you could be part of, can build real computer vision products."

Extension activity: Have students pick one real, publicly documented computer vision product (a specific self-driving perception stack, a medical-imaging FDA-cleared tool, a well-known app like Google Lens or PlantNet) and research, in a short write-up or presentation, which of today's concepts it plausibly uses — CNN vs. ViT-style architecture, real-time detection needs, transfer learning, or data augmentation — based on publicly available technical blog posts or papers from the company. Push them to cite what they found rather than guessing, and to flag clearly where they're inferring versus where a source states it directly.

For a class with more time or a project-based component, consider a second extension: have students design (on paper — no coding required) a computer vision system for a real problem relevant to their own community, such as identifying counterfeit or damaged produce at a public market, flagging jeepney or tricycle overloading from a fixed camera, or sorting recyclables by material type. For each, have them specify: what task type it is (classification, detection, or segmentation), whether transfer learning from an existing pretrained model would likely be a realistic starting point given a small budget, what data augmentation techniques would make sense given the conditions the camera would actually operate in (variable outdoor lighting, motion blur, dust), and one ethical or accuracy risk worth flagging before deployment. This forces synthesis of every concept in today's lesson against a concrete, locally grounded scenario rather than an abstract one.