RobotnixAcademy
← Courses

AI

AI for Beginners

An 18-week introduction to artificial intelligence that moves from symbolic reasoning, neural networks, computer vision, language, and responsible AI into machine learning, evaluation, clustering, production systems, edge deployment, and an evidence-based capstone. Weeks 1-12 preserve Microsoft's AI for Beginners source material and attribution; Weeks 13-18 continue the learner experience with Robotnix-original applied chapters and interactive laboratories.

Length
18 weeks
Audience
Grades 9-12
Entries
114
Interactives
7 labs
Provenance
Sources and credits

Course licensing

Microsoft source and license

Weeks 1–12 preserve the source-faithful Microsoft foundation. Weeks 13–18 are the Robotnix applied extension. The source material is from Microsoft AI for Beginners by Microsoft and the upstream contributors, pinned at commit 33e781bf7bfb9b39fd27c4e4a3e592669b52cb4b, and provided under the MIT License.

View complete sources, credits, modifications, and reuse limits.

Week 01

  1. overviewWeek 1: What Can AI Systems Do?What kinds of jobs can an AI system perform, and what must we inspect before we trust its result?
  2. Lesson 1Introduction to AIArtificial Intelligence is an exciting scientific discipline that studies how we can make computers exhibit intelligent behavior, e.g. do those things that human beings are good at doing.
  3. Lesson 2Knowledge Representation and Expert SystemsThe quest for artificial intelligence is based on a search for knowledge, to make sense of the world similar to how humans do. But how can you go about doing this?
  4. Lesson 3Investigation: Where System Behavior Comes FromTwo systems can accept similar inputs and produce similar-looking outputs while working in completely different ways. In this guided investigation, you will trace three mechanisms before the Lab asks you to compare…
  5. LabLab: Build an AI Field MapRun four supplied systems through the same trace. Your job is not to decide which one is “most intelligent.” Your job is to identify the task, mechanism, test, and failure boundary of each one.
  6. reflectionWeek 1 Wrap-Up: What Can AI Systems Do?Use your AI Field Map Evidence Record. Respond in 75–125 words or a one-minute recording. Use at least two exact labels from the artifact.

Week 02

  1. overviewWeek 2: Perceptrons and Layered NetworksA neural network is not a digital brain or a box that “thinks.” It is a parameterized calculation. This week begins with Microsoft's pinned perceptron and multilayer-perceptron chapters, then slows the mechanism down…
  2. Lesson 1Introduction to Neural Networks: PerceptronOne of the first attempts to implement something similar to a modern neural network was done by Frank Rosenblatt from Cornell Aeronautical Laboratory in 1957. It was a hardware implementation called "Mark-1", designed…
  3. Lesson 2Introduction to Neural Networks. Multi-Layered PerceptronIn the previous section, you learned about the simplest neural network model - one-layered perceptron, a linear two-class classification model.
  4. Lesson 3Investigation: Find the BoundaryA perceptron combines represented inputs into one score and compares that score with a threshold. The arithmetic is simple enough to inspect:
  5. LabLab: Build a Two-Layer XOR NetworkBuild a hand-checkable layered calculation that matches all four XOR cases. You will not hide a failed row or claim that four binary cases prove anything about a deployed neural network.
  6. reflectionWeek 2 Wrap-Up: Perceptrons and Layered NetworksUse your week evidence record. Respond in 75-125 words or a one-minute recording. Use at least two exact labels from the artifact.

Week 03

  1. overviewWeek 3: Frameworks and the Vision PipelineNeural-network frameworks make tensor operations, gradient calculation, and hardware acceleration manageable. Computer vision begins by turning captured light into numeric arrays. This week connects those ideas without…
  2. Lesson 1Neural Network FrameworksAs we have learned already, to be able to train neural networks efficiently we need to do two things:
  3. Lesson 2Introduction to Computer VisionOne of the simplest tasks of computer vision is image classification.
  4. Lesson 3Investigation: Transformations Change the EvidenceA computer-vision pipeline often transforms an image before producing a result. A transformation can make a feature easier to measure, but it can also remove information or expose a fragile shortcut.
  5. LabLab: Inspect the Pixel PipelineRun the local Pixel Inspector Lab. Trace one supplied scene from RGB values to grayscale and edge evidence, compare controlled conditions, and export a Pixel Evidence Record.
  6. reflectionWeek 3 Wrap-Up: Frameworks and the Vision PipelineUse your week evidence record. Respond in 75-125 words or a one-minute recording. Use at least two exact labels from the artifact.

Week 04

  1. overviewWeek 4: Convolution and Transfer LearningConvolutional networks preserve spatial structure by applying small filters across local neighborhoods. Transfer learning reuses representations learned on one dataset, then adapts the model for another task. Both…
  2. Lesson 1Convolutional Neural NetworksWe have seen before that neural networks are quite good at dealing with images, and even one-layer perceptron is able to recognize handwritten digits from MNIST dataset with reasonable accuracy. However, the MNIST…
  3. Lesson 2Pre-trained Networks and Transfer LearningTraining CNNs can take a lot of time, and a lot of data is required for that task. However, much of the time is spent learning the best low-level filters that a network can use to extract patterns from images. A…
  4. Lesson 3Investigation: From Local Responses to Reused FeaturesA convolution filter examines one local neighborhood at a time. At each position, matching image and kernel values are multiplied and added. Sliding the same kernel across the image produces a feature map.
  5. LabLab: Slide the Convolution WindowRun the local Convolution Window Lab. Move one 3 × 3 kernel across a fixed 6 × 6 binary image, preserve one complete calculation, compare two feature maps, and export a Convolution Evidence Record.
  6. reflectionWeek 4 Wrap-Up: Convolution and Transfer LearningUse your week evidence record. Respond in 75-125 words or a one-minute recording. Use at least two exact labels from the artifact.

Week 05

  1. overviewWeek 5: Compression and Generative ModelsAutoencoders learn an encoder, bottleneck representation, and decoder by minimizing reconstruction error. GANs train a generator and discriminator in an adversarial loop. Both can produce outputs; neither makes those…
  2. Lesson 1AutoencodersWhen training CNNs, one of the problems is that we need a lot of labeled data. In the case of image classification, we need to separate images into different classes, which is a manual effort.
  3. Lesson 2Generative Adversarial NetworksIn the previous section, we learned about generative models: models that can generate new images similar to the ones in the training dataset. VAE was a good example of a generative model.
  4. Lesson 3Investigation: A Generated Output Is Not a Proven OutputAn autoencoder maps an input through an encoder into a smaller latent representation, then uses a decoder to reconstruct the input. Training adjusts parameters to reduce reconstruction loss.
  5. LabLab: Measure Latent Reconstruction LossRun the local Latent Reconstruction Lab. Compare a two-bit bottleneck with a one-bit bottleneck, inspect every reconstruction, and export a bounded evidence record.
  6. reflectionWeek 5 Wrap-Up: Compression and Generative ModelsUse your week evidence record. Respond in 75-125 words or a one-minute recording. Use at least two exact labels from the artifact.

Week 06

  1. overviewWeek 6: Detection, Segmentation, and Localization EvidenceClassification names an image-level category. Detection adds object location with boxes. Segmentation assigns classes at pixel level. More detailed output does not automatically mean more accurate or safer output.
  2. Lesson 1Object DetectionThe image classification models we have dealt with so far took an image and produced a categorical result, such as the class 'number' in a MNIST problem. However, in many cases we do not want just to know that a…
  3. Lesson 2SegmentationWe have previously learned about Object Detection, which allows us to locate objects in the image by predicting their *bounding boxes*. However, for some tasks we do not only need bounding boxes, but also more precise…
  4. Lesson 3Investigation: Measure Localization Before Judging ItA bounding box and a segmentation mask can both localize something, but their geometry differs. A box covers a rectangle. A mask identifies selected pixels. Both can be compared with a reference annotation using…
  5. LabLab: Compare Localization OverlapRun the local Localization Overlap Lab. Compare tight, shifted or under-segmented, and oversized predictions for both boxes and masks. Export a Localization Evidence Record.
  6. reflectionWeek 6 Wrap-Up: Detection, Segmentation, and Localization EvidenceUse your week evidence record. Respond in 75-125 words or a one-minute recording. Use at least two exact labels from the artifact.

Week 07

  1. overviewWeek 7: Text Representations and EmbeddingsText must become numeric representations before a neural network can process it. Tokenization chooses units, vocabulary assigns identifiers, and vector representations determine which relationships a model can measure.
  2. Lesson 1Representing Text as TensorsThroughout the first part of this section, we will focus on text classification task. We will use the AG News Dataset, which contains news articles like the following:
  3. Lesson 2EmbeddingsWhen training classifiers based on BoW or TF/IDF, we operated on high-dimensional bag-of-words vectors with length vocab_size, and we were explicitly converting from low-dimensional positional representation vectors…
  4. Lesson 3Investigation: Similarity Depends on RepresentationA token is a selected text unit. A vocabulary maps each known token to an ID. An ID is an address, not meaning. A one-hot vector preserves identity but makes every different token orthogonal. A learned embedding maps…
  5. LabLab: Compare Token RepresentationsRun the local Embedding Similarity Lab. Compare the same three tokens under one-hot and supplied dense vectors, reproduce cosine calculations, and export an Embedding Similarity Record.
  6. reflectionWeek 7 Wrap-Up: Text Representations and EmbeddingsUse your week evidence record. Respond in 75-125 words or a one-minute recording. Use at least two exact labels from the artifact.

Week 08

  1. overviewWeek 8: Language Modeling and Recurrent Neural NetworksLanguage models learn patterns in token sequences. Recurrent neural networks process those sequences in order, carrying a state forward so later steps can depend on earlier input.
  2. Lesson 1Language ModelingSemantic embeddings, such as Word2Vec and GloVe, are in fact a first step towards language modeling - creating models that somehow *understand* (or *represent*) the nature of the language.
  3. Lesson 2Recurrent Neural NetworksIn previous sections, we have been using rich semantic representations of text and a simple linear classifier on top of the embeddings. What this architecture does is to capture the aggregated meaning of words in a…
  4. Lesson 3Investigation: Order Changes Recurrent StateA language model estimates which tokens fit a context. A recurrent network processes one token at a time and carries a state forward. The same update rule is reused at every step, but the state arriving at each step…
  5. LabLab: Trace Recurrent StateRun the local Recurrent State Trace Lab. Compare sequences that reuse the same update rule, preserve every calculation, and export a Recurrent State Trace Record.
  6. reflectionWeek 8 Wrap-Up: Language Modeling and Recurrent Neural NetworksUse your week evidence record. Respond in 75-125 words or a one-minute recording. Use at least two exact labels from the artifact.

Week 09

  1. overviewWeek 9: Generative Networks, Attention, and TransformersGenerative networks produce sequences one step at a time. Attention changes the mechanism: each query can weight relevant context positions directly, while positional information preserves order for transformer…
  2. Lesson 1Generative networksRecurrent Neural Networks (RNNs) and their gated cell variants such as Long Short Term Memory Cells (LSTMs) and Gated Recurrent Units (GRUs) provided a mechanism for language modeling in that they can learn word…
  3. Lesson 2Attention Mechanisms and TransformersOne of the most important problems in the NLP domain is machine translation, an essential task that underlies tools such as Google Translate. In this section, we will focus on machine translation, or, more generally,…
  4. Lesson 3Investigation: Attention Is a Calculated AllocationA generative model repeatedly scores possible next tokens and selects or samples one. Attention solves a different problem inside the model: it builds a context representation by weighting input positions for a…
  5. LabLab: Compare Attention AllocationsRun the local Attention Allocation Lab. Apply two queries to the same three keys, reproduce softmax normalization, and export an Attention Allocation Record.
  6. reflectionWeek 9 Wrap-Up: Generative Networks, Attention, and TransformersUse your week evidence record. Respond in 75-125 words or a one-minute recording. Use at least two exact labels from the artifact.

Week 10

  1. overviewWeek 10: Named Entities and Pre-Trained Language ModelsNamed entity recognition turns text into auditable token labels and entity spans. Pre-trained language models can perform extraction with instructions or examples, but useful output still requires a declared schema,…
  2. Lesson 1Named Entity RecognitionUp to now, we have mostly been concentrating on one NLP task - classification. However, there are also other NLP tasks that can be accomplished with neural networks. One of those tasks is Named Entity Recognition…
  3. Lesson 2Pre-Trained Large Language ModelsIn all of our previous tasks, we were training a neural network to perform a certain task using labeled dataset. With large transformer models, such as BERT, we use language modelling in self-supervised fashion to…
  4. Lesson 3Investigation: Extraction Needs a Schema and an AuditNamed entity recognition is token classification with structure. BIO labels mark the beginning and inside of an entity span; O marks tokens outside the declared entity types.
  5. LabLab: Audit Entity SpansRun the local Entity Span Audit Lab. Compare three supplied prediction sets against one gold BIO sequence, reproduce exact-span metrics, and export an Entity Span Audit Record.
  6. reflectionWeek 10 Wrap-Up: Named Entities and Pre-Trained Language ModelsUse your week evidence record. Respond in 75-125 words or a one-minute recording. Use at least two exact labels from the artifact.

Week 11

  1. overviewWeek 11: Genetic Algorithms and Reinforcement LearningGenetic algorithms search populations using fitness, selection, crossover, and mutation. Reinforcement learning improves policies through interaction and reward. Both expose the same systems truth: optimization follows…
  2. Lesson 1Genetic AlgorithmsGenetic Algorithms (GA) are based on an evolutionary approach to AI, in which methods of the evolution of a population is used to obtain an optimal solution for a given problem. They were proposed in 1975 by John Henry…
  3. Lesson 2Deep Reinforcement LearningReinforcement learning (RL) is seen as one of the basic machine learning paradigms, next to supervised learning and unsupervised learning. While in supervised learning we rely on the dataset with known outcomes, RL is…
  4. Lesson 3Investigation: An Optimizer Follows the ObjectiveA genetic algorithm uses a fitness function to compare candidate genes. A reinforcement-learning agent uses reward to update a policy. These mechanisms differ, but neither can infer an unstated goal. If safety,…
  5. LabLab: Audit an Objective FunctionRun the local Objective Function Audit Lab. Rank the same supplied behaviors under naive and safety-aware objectives, preserve every operand, and export an Objective Function Audit Record.
  6. reflectionWeek 11 Wrap-Up: Genetic Algorithms and Reinforcement LearningUse your week evidence record. Respond in 75-125 words or a one-minute recording. Use at least two exact labels from the artifact.

Week 12

  1. overviewWeek 12: Multi-Agent Systems and Responsible AIMulti-agent behavior depends on individual capabilities, shared protocols, messages, environmental feedback, and authority boundaries. Responsible AI is therefore a system practice: risks and accountability live across…
  2. Lesson 1Multi-Agent SystemsOne of the possible ways of achieving intelligence is so-called emergent (or synergetic) approach, which is based on the fact that the combined behavior of many relatively simple agents can result in the overall more…
  3. Lesson 2Ethical and Responsible AIYou have almost finished this course, and I hope that by now you clearly see that AI is based on a number of formal mathematical methods that allow us to find relationships in data and train models to replicate some…
  4. Lesson 3Investigation: Coordination Needs Contracts and AuthorityAn agent perceives an environment and takes actions. A multi-agent system adds handoffs: one agent's output becomes another agent's input. Complex behavior can emerge from those interactions, but so can ambiguity,…
  5. LabLab: Audit Multi-Agent HandoffsRun the local Multi-Agent Handoff Audit Lab. Compare verified, schema-mismatch, and unauthorized-action traces, inspect four controls, and export a Multi-Agent Handoff Audit Record.
  6. reflectionWeek 12 Wrap-Up: Multi-Agent Systems and Responsible AIUse your week evidence record. Respond in 75-125 words or a one-minute recording. Use at least two exact labels from the artifact.

Week 13

  1. overviewWeek 13: Learning from ExamplesThe Microsoft sequence closes by asking us to think carefully about what AI can and cannot justify. This week continues from that point with a more mechanical question: when we say a model learns from data, what is…
  2. Lesson 1Features, Labels, and the Question a Model LearnsA row in a dataset is not knowledge. It is one recorded case. Learning begins only after we decide what information the model may use and what answer, if any, it is expected to predict.
  3. Lesson 2Training, Decision Boundaries, and Held-Out DataTraining is the process of choosing model parameters from examples. In a modern neural network there may be millions or billions of parameters. In our tiny experiment there is only one: a distance threshold.
  4. Lesson 3Generalization, Leakage, and the Limits of a Good FitThe purpose of supervised learning is not to memorize a table. It is to learn a relationship that remains useful on new cases from the problem we care about.
  5. LabLab — Train, Break, and Defend a ClassifierThe chapter so far used a threshold because it is simple enough to inspect. Now use the same idea as an experiment rather than a diagram.
  6. ReviewReview & Self Study — What Did the Model Actually Learn?Before leaving supervised learning, separate the mechanics from the claims.
  7. reflectionWeek 13 ReflectionIn a short paragraph, explain the difference between fitting a model and proving a model works on new data.

Week 14

  1. overviewWeek 14: Measuring Whether a Model WorksA model score is not a verdict. It is the result of comparing predictions with known outcomes under a specific evaluation procedure.
  2. Lesson 1Four Outcomes Behind a Model ScoreFor a binary classifier, every evaluated prediction lands in one of four places.
  3. Lesson 2Accuracy, Precision, Recall, and Error CostOnce the confusion counts are visible, common metrics become easier to understand.
  4. Lesson 3Thresholds Turn Scores into DecisionsMany classifiers do not begin by producing a hard yes/no answer. They produce a score or probability-like value. A decision threshold turns that score into a category.
  5. LabLab — Move the ThresholdOpen the Model Evaluation Lab.
  6. ReviewReview & Self Study — Which Error Matters?The confusion matrix is the evidence table. Metrics summarize different parts of that table. Thresholds change which cells receive new cases.
  7. reflectionWeek 14 ReflectionChoose either a false positive or false negative as the more important error in one fictional AI decision.

Week 15

  1. overviewWeek 15: Finding Patterns Without LabelsSupervised learning begins with target answers. Unsupervised learning asks what structure can be found when those labels are absent.
  2. Lesson 1Finding Structure Without Target LabelsA dataset can contain useful structure even when no target label is attached to each row.
  3. Lesson 2K-Means, Distance, and IterationK-means is one of the simplest algorithms for understanding how clustering can emerge from repeated computation.
  4. Lesson 3Representation Changes the ClustersA clustering algorithm sees numbers, not the physical meaning of the objects behind them. That makes representation one of the most important choices in unsupervised learning.
  5. LabLab — Break the ClustersOpen the Clustering Playground.
  6. ReviewReview & Self Study — A Cluster Is Not a TypeClustering can reveal useful structure without target labels, but the analyst still chooses the representation, distance assumptions, and often the number of groups.
  7. reflectionWeek 15 ReflectionDescribe one cluster you observed in the playground without giving it a human category name.

Week 16

  1. overviewWeek 16: From Model to AI SystemA trained model is only one component in an AI system.
  2. Lesson 1The Model Is Only One ComponentIt is tempting to draw an AI system as one box labeled MODEL. That picture hides most of the engineering.
  3. Lesson 2Data Contracts, Serving, and MonitoringA model expects inputs in a particular form. A data contract makes those expectations explicit enough to test.
  4. Lesson 3Fallbacks, Authority, and Failure PathsA system is not fully designed when only the success path is defined.
  5. LabLab — Break the System Around the ModelOpen the AI System Pipeline Simulator.
  6. ReviewReview & Self Study — A Score Is Not a System DecisionThis week moved from model evaluation to architecture.
  7. reflectionWeek 16 ReflectionName one failure that could occur outside the model while the model itself remains unchanged.

Week 17

  1. overviewWeek 17: AI at the EdgeWhere inference runs changes the system boundary.
  2. Lesson 1Where Inference HappensEdge AI usually means inference happens close to where data is produced or where action is required: on the device itself or on nearby local compute.
  3. Lesson 2Latency, Memory, Privacy, and Power BudgetsDeployment decisions become easier to defend when vague preferences are replaced with budgets.
  4. Lesson 3Designing for Disconnection and FailureA deployment architecture should be evaluated in the state where something important is unavailable.
  5. LabLab — Choose an Edge or Cloud DeploymentOpen the Edge vs Cloud Deployment Simulator.
  6. ReviewReview & Self Study — Deployment Is a Constraint DecisionEdge, cloud, and hybrid are placement choices. None is automatically more intelligent, secure, private, cheap, or reliable.
  7. reflectionWeek 17 ReflectionChoose edge, cloud, or hybrid for one fictional AI application.

Week 18

  1. overviewWeek 18: Build and Defend an AI SystemThe capstone is not a model demo. It is a system defense.
  2. Lesson 1Build the System Before You Defend ItBefore presenting results, freeze the system you intend to defend.
  3. Lesson 2Test Success, Failure, and RecoveryA capstone that demonstrates only the happy path leaves the most interesting questions unanswered.
  4. Lesson 3Audit Every Claim Against EvidenceBefore presenting the capstone, read every sentence that makes a technical claim and ask: what evidence supports this?
  5. LabLab — Assemble the AI System DefenseOpen the AI System Defense Lab.
  6. ReviewReview & Self Study — Defend the System Without Overselling ItYour final presentation should make it easy to distinguish what you built, what you tested, what failed, what responded, and what remains unknown.
  7. reflectionWeek 18 ReflectionWhat is the strongest claim your capstone evidence genuinely supports?