Comparisons and
Roadmaps
The big table
The Big Table: I. Core
| Course | Approach | Input | Output | Success measure |
|---|---|---|---|---|
| Artificial Intelligence | search, logic, probability and learning combined | problem definition: states, actions, goal | action sequence, plan or decision | reaching the goal, cost, rationality |
| Machine Learning | statistical learning | labeled or unlabeled examples | a model that makes predictions | error on unseen data |
| Neural Networks | network of simple weighted units | numeric feature vector | class or numeric prediction | loss and accuracy |
| Deep Learning | multilayer representation learning | raw data: images, text, audio | learned representation and prediction | accuracy, loss, scaling |
The big table
The Big Table: II. Branches of AI
| Course | Approach | Input | Output | Success measure |
|---|---|---|---|---|
| NLP | linguistics, statistics and deep learning | text | label, translation, summary, answer | accuracy, BLEU, human evaluation |
| Reinforcement Learning | trial and error with rewards | environment: state, reward | policy: state → action | total expected reward |
| Pattern Recognition | statistical decision theory | measurement or feature vector | class label | classification error |
| Fuzzy Logic | partial membership and linguistic rules | measurements, expert rules | control signal or decision | control quality, expert agreement |
| Evolutionary Optimization | population-based heuristic search | objective, solution encoding | good, not proven optimal, solution | quality, evaluation count |
| Expert Systems | symbolic logic and rules | facts and expert rules | conclusion with justification | expert agreement, explainability |
| Graphical Models | probability and graphs combined | variables, dependencies, data | conditional probability, best explanation | likelihood, inference accuracy |
| Generative Models | learning the data distribution | huge, mostly unlabeled data | new samples: images, text, audio | realism and diversity (FID etc.) |
| Multi-Agent Systems | interacting autonomous agents | agents, goals, protocols | joint behavior, equilibrium | social welfare, stability |
| Explainable AI & Ethics | human-centered evaluation | model, its data and context of use | explanation, audit report, principles | interpretability, fairness, accountability |
The big table
The Big Table: III. Signals, Images and Audio
| Course | Approach | Input | Output | Success measure |
|---|---|---|---|---|
| Signal Processing | mathematical transforms and filters | sampled signal | filtered signal, spectrum | signal-to-noise ratio, frequency response |
| Image Processing | pixel-level transformations | image | processed image or regions | visual quality, PSNR |
| Computer Vision | extracting meaning from images | image or video | label, box, depth, 3D structure | accuracy, mAP, IoU |
| Speech and Audio | acoustic signal and language model | audio waveform | text, speaker, sound event | word error rate (WER) |
| Medical Imaging | image processing plus clinical knowledge | CT, MRI, X-ray, ultrasound | segmentation, measurement, diagnostic suggestion | sensitivity, specificity, Dice |
| Remote Sensing | physical sensing and image analysis | multispectral satellite imagery | land cover map, change | map accuracy, kappa |
| Computer Graphics | geometry and light simulation | scene: models, lights, camera | image or animation | realism, frame rate |
| Robotics | sense-plan-act loop | sensor measurements | motor commands, path | task success, safety |
| Biometrics | matching and verification | face, fingerprint, iris, voice | identity decision | FAR, FRR, EER |
The big table
The Big Table: IV. Data
| Course | Approach | Input | Output | Success measure |
|---|---|---|---|---|
| Data Mining | discovery-driven analysis | large tables or transaction data | rules, clusters, outliers | interestingness, support, confidence |
| Data Science | question, data, analysis, story | raw, messy data | report, visualization, model | impact on decisions, accuracy |
| Big Data | distributed storage and computation | huge, fast-flowing data | processed data, query results | scalability, latency, cost |
| IR & Recommenders | matching and ranking | query, documents, user history | ranked result list | precision, recall, NDCG |
| Time Series | time-dependent statistical models | timestamped measurements | forecast, trend, seasonality | forecast error (MAE, RMSE) |
| Social Network Analysis | graph theory and statistics | nodes and links | centrality, communities, diffusion | modularity, interpretability |
The big table
The Big Table: V. Mathematical Foundations
| Course | Approach | Input | Output | Success measure |
|---|---|---|---|---|
| Linear Algebra | vector spaces and transformations | vectors, matrices | solution, decomposition, transformation | accuracy, numerical stability |
| Probability & Statistics | probability models and inference | sample data | estimate, confidence interval, test result | unbiasedness, power, significance |
| Optimization | mathematical model of objective and constraints | objective function and constraints | optimal or near-optimal solution | objective value, convergence rate |
| Information Theory | measuring information with probability | source and channel model | bounds, codes, compression rate | bits, channel capacity |
The big table
The Big Table: VI. Systems and Applications
| Course | Approach | Input | Output | Success measure |
|---|---|---|---|---|
| MLOps | software engineering meets ML | trained model, data pipeline | a running, monitored service | latency, uptime, data drift |
| Embedded AI | resource-constrained inference | trained model, device constraints | small, fast model | latency, energy, memory |
| GPU & Parallel Computing | data-parallel computation | parallelizable algorithm | accelerated program | speedup, memory bandwidth |
| AI in Cybersecurity | security knowledge plus ML | network traffic, logs, files | alert, class, risk score | detection rate, false alarm rate |
| Bioinformatics | biology, algorithms and statistics | DNA, protein sequences, expression data | alignment, structure, function prediction | biological validity |
The big table
Math and Data Needs (1/3)
| Course | Math load | Data needs |
|---|---|---|
| Artificial Intelligence | medium: logic, probability, graphs | little; often a model and knowledge suffice |
| Machine Learning | high: linear algebra, probability, optimization | moderate to large |
| Neural Networks | medium-high: derivatives, matrices | moderate |
| Deep Learning | high: derivatives, matrices, probability | large; needs GPUs |
| NLP | medium | large text corpora |
| Reinforcement Learning | high: probability, dynamic programming | generated through interaction |
| Pattern Recognition | medium-high: probability, linear algebra | labeled examples |
| Fuzzy Logic | low-medium: set theory | little; expert knowledge |
| Evolutionary Optimization | low-medium | none; evaluations suffice |
| Expert Systems | low: propositional and predicate logic | little; a knowledge base |
| Graphical Models | high: probability | moderate |
| Generative Models | high | very large |
| Multi-Agent Systems | medium: game theory | little; simulation |
The big table
Math and Data Needs (2/3)
| Course | Math load | Data needs |
|---|---|---|
| Explainable AI & Ethics | low-medium | the model's own data |
| Signal Processing | high: complex numbers, Fourier | no learning; the signal suffices |
| Image Processing | medium: convolution, Fourier | little; learning not required |
| Computer Vision | high: geometry, linear algebra | many labeled images |
| Speech and Audio | medium-high | many hours of recordings |
| Medical Imaging | medium-high | little; labels are expensive |
| Remote Sensing | medium | plenty; labels limited |
| Computer Graphics | high: linear algebra, geometry | none |
| Robotics | high: transformations, control | physical experiments and simulation |
| Biometrics | medium | few samples per person |
| Data Mining | medium | large |
| Data Science | medium: statistics | moderate to large |
| Big Data | low | very large |
The big table
Math and Data Needs (3/3)
| Course | Math load | Data needs |
|---|---|---|
| IR & Recommenders | medium: vector spaces | lots of interaction data |
| Time Series | medium-high | long history |
| Social Network Analysis | medium: graphs, matrices | network data |
| Linear Algebra | the theory itself | none |
| Probability & Statistics | the theory itself | a sample |
| Optimization | the theory itself | none |
| Information Theory | high: probability | none |
| MLOps | low | streaming production data |
| Embedded AI | medium | training data already exists |
| GPU & Parallel Computing | low-medium | none |
| AI in Cybersecurity | medium | imbalanced and constantly changing |
| Bioinformatics | medium-high | huge sequence datasets |
Pairs
Often-Confused Pairs (1/2)
| Pair | Difference |
|---|---|
| Image Processing ↔ Computer Vision | In image processing both input and output are images; in computer vision the input is an image and the output is meaning: a label, a box, depth. |
| Artificial Intelligence ↔ Machine Learning | AI can find how to reach a goal by search or inference; ML learns the same behavior by generalizing from data. |
| Machine Learning ↔ Data Mining | They share algorithms; ML aims to predict new examples, data mining to discover patterns in data. |
| Data Science ↔ Data Mining | Data mining focuses on finding patterns in large data; data science covers the whole path from question to decision. |
| Neural Networks ↔ Deep Learning | Deep learning moves on to convolutional, recurrent and transformer architectures; this course builds the basic network and training they rest on. |
| Pattern Recognition ↔ Machine Learning | ML studies learning algorithms in general; this course focuses on the design from measurement to class decision and on the Bayes error. |
| NLP ↔ Speech and Audio | Speech recognition goes from the sound wave to text; NLP analyzes meaning once the text is available. |
Pairs
Often-Confused Pairs (2/2)
| Pair | Difference |
|---|---|
| Generative Models ↔ Deep Learning | Deep learning is the general toolkit and mostly makes predictions; this course uses those tools to generate new samples from a data distribution. |
| Computer Graphics ↔ Computer Vision | Computer vision tries to recover the scene from an image; computer graphics produces the image from the scene. |
| Machine Learning ↔ Probability & Statistics | Statistics emphasizes a model's meaning and uncertainty, ML its predictive performance. |
| IR & Recommenders ↔ Data Mining | Data mining discovers patterns in data; this course turns discovery into a ranked list for one user's current need. |
| Expert Systems ↔ Fuzzy Logic | Fuzzy logic lets rules be partly true; this course teaches representing knowledge and chaining rules in general. |
| Evolutionary Optimization ↔ Optimization | Optimization reaches a provable optimum using derivatives and convexity; this course teaches derivative-free search with no guarantee but wide applicability. |
| MLOps ↔ Big Data | Big data systems store and process data at scale; this course manages the lifecycle of models trained on that data. |
Which course?
Which Course? (1/2)
For each problem, decide which course it belongs to, then open the answer.
- Automatically outlining tumours in a hospital's MRI scans.
Answer
Medical Imaging, Computer Vision, Image Processing. Segmentation comes from image processing and computer vision; clinical data, scarce labels and accountability make it medical image processing. - Building a "customers who bought this also bought" list for an online shop.
Answer
IR & Recommenders, Data Mining. Association rules come from data mining; presenting a ranked list to the user is the job of recommender systems. - Controlling an oven's temperature with rules like "a bit hot" and "very hot".
Answer
Fuzzy Logic. Linguistic rules without sharp boundaries, used for control, are the classic application of fuzzy logic. - A drone learning to stay level in the wind by trying.
Answer
Reinforcement Learning, Robotics. No labels, only a reward: reinforcement learning. The physical body and its sensors are the business of robotics. - Storing and querying per-second readings from a hundred thousand sensors.
Answer
Big Data, Time Series. Streaming data too large for one machine belongs to big data systems; analysing it over time is time series work.
Which course?
Which Course? (2/2)
For each problem, decide which course it belongs to, then open the answer.
- Finding the shortest total route plan for school buses.
Answer
Evolutionary Optimization, Optimization, Artificial Intelligence. A combinatorial optimization problem: exact methods come from optimization, good solutions to large instances from metaheuristics. - Auditing whether a credit model treats some groups of applicants unfairly.
Answer
Explainable AI & Ethics, Data Science. Fairness measures, explanations and accountability are the subject of explainable AI and ethics. - Voice commands that work on a phone without an internet connection.
Answer
Embedded AI, Speech and Audio. The recognition model comes from speech and audio processing; fitting it onto the device is the job of embedded AI. - Finding the genes linked to a disease from expression data of thousands of samples.
Answer
Bioinformatics, Probability & Statistics. Biological data and its interpretation are bioinformatics; significance across many tests needs statistics. - Catching unusual connections in a bank's network traffic.
Answer
AI in Cybersecurity, Data Mining. Outlier detection comes from data mining; an attacker who adapts makes it a cybersecurity problem.
Choosing
From Interest to Course
Pick one or more interests; the courses that serve them are listed, ranked by how many of your interests they touch.
Choosing
Roadmaps (1/2)
Each track runs from prerequisites to specialization; the tracks cross, and following more than one is normal.
Choosing
Roadmaps (2/2)
Each track runs from prerequisites to specialization; the tracks cross, and following more than one is normal.
Choosing
Final Thoughts
Linear algebra, probability and optimization are shared by almost every course; with these in place, moving between courses is easy.
Machine learning sits at the centre of the map: the branches rest on it and the applications use it.
Two courses may use the same algorithm; what separates them is the question they ask and how they measure success.
Separate courses today may merge tomorrow; understanding the kinds of relation lasts longer than memorizing the map.