The Landscape of Modern AI: A Map for Students

We end the Foundations track by stepping back and looking at the whole map. When students first enter AI, they often feel lost among acronyms — CNN, RL, LLM, RAG, MLOps. This lecture gives you a mental atlas. Return to it whenever you feel lost in later tracks.

Three paradigms, one goal#

Every AI system you meet uses some mixture of three paradigms:

Search and reasoning — explore possibilities explicitly (A*, alpha–beta, MCTS, planners, SAT solvers).
Probabilistic modelling — represent and update uncertainty (Bayesian networks, HMMs, Gaussian processes).
Learning from data — fit functions from examples (regression, trees, neural networks).

Modern breakthroughs often combine them: AlphaZero combines search and learning; a self-driving stack combines learned perception, probabilistic tracking and planning; an LLM agent combines a learned language model with tool calls and search.

The machine learning taxonomy#

Paradigm	Data	Goal	Example
Supervised learning	Inputs with labels	Predict labels	Spam detection
Unsupervised learning	Inputs only	Discover structure	Customer segmentation
Self-supervised learning	Inputs; labels derived from data	Learn representations	Predicting masked words
Reinforcement learning	Interaction and rewards	Learn a policy	Game playing, robotics
Semi-supervised / weak supervision	Few labels, many unlabelled	Leverage both	Medical imaging

The deep learning stack#

Deep learning dominates perception and language. Its components:

Architectures — MLPs, CNNs (images), RNNs (sequences), Transformers (almost everything now), GNNs (graphs), diffusion models (generation).
Training — backpropagation, stochastic gradient descent and Adam, normalisation, regularisation.
Scale — GPUs/TPUs, distributed training, mixed precision.
Paradigms — pretraining on huge unlabelled data, then fine-tuning or prompting for tasks.

Application domains#

Computer vision — classification, detection, segmentation, medical imaging, remote sensing.
Natural language processing — translation, question answering, summarisation, chat assistants.
Speech — recognition, synthesis, speaker identification.
Science — protein structure, weather forecasting, materials discovery.
Recommendation and search — the most economically significant ML systems in the world.
Robotics and control — manipulation, navigation, autonomous vehicles.
Humanitarian and development work — mapping settlements from satellite imagery, forecasting displacement, multilingual information services for people in need.

The engineering reality#

A research notebook is not a product. Deployed ML requires data pipelines, evaluation, monitoring, versioning and governance. Studies of production systems have repeatedly found that the model code is a small fraction of the whole system. That is why this course includes an MLOps track.

Research frontiers (as of today)#

Reasoning and planning in learned models.
Efficiency — smaller, faster, cheaper models; on-device AI.
Multimodality — joint understanding of text, images, audio, video and action.
Agents — models that use tools, browse, write code and act over long horizons.
Alignment, interpretability and safety.
Data — curation, synthetic data, and the limits of available human-generated text.
AI for science and AI for social good.

How to use this course#

The tracks form a deliberate sequence:

text

AI Foundations ──► Mathematics for ML ──► Machine Learning ──► Deep Learning
                                                                   │
            ┌───────────────┬──────────────────┬───────────────────┤
            ▼               ▼                  ▼                   ▼
     Computer Vision   NLP & Transformers   Generative AI   Reinforcement Learning
            └───────────────┴────────┬─────────┴───────────────────┘
                                     ▼
                     MLOps & Engineering ──► Ethics, Society & Careers

Tools you should install#

Python 3 with numpy, pandas, matplotlib, scikit-learn.
PyTorch (or TensorFlow/Keras) for deep learning.
Jupyter or VS Code notebooks for experiments.
Git for version control — from your very first project.
A free cloud GPU notebook environment for heavier experiments.

bash

python -m venv ai-course
source ai-course/bin/activate        # on Windows: ai-course\Scripts\activate
pip install numpy pandas matplotlib scikit-learn torch jupyter

#overview#roadmap#careers#research areas

Written by

Janin A Apurba

B.Sc. in CSE, AUST · Advanced ICT Officer, CNRS-UNHCR. Teaching AI, ML and Deep Learning to the next generation of engineers and researchers.