Machine learning (ML) is a branch of artificial intelligence that enables computers to learn from data without being explicitly programmed for every possible situation. Instead of hand‑crafting rules, we feed examples and let an algorithm adjust its internal parameters so that it can make predictions or decisions on new, unseen data. ML is typically divided into three main paradigms:
Supervised learning – the algorithm learns from labelled examples (e.g., photos tagged “cat” or “dog”). Unsupervised learning – the algorithm finds hidden patterns or groupings in unlabelled data (e.g., clustering customers by purchasing behaviour). Reinforcement learning – an agent learns by interacting with an environment, receiving rewards or penalties (e.g., a robot learning to walk).
A Brief History of Machine Learning
The roots of machine learning stretch back long before digital computers. In 1805, Adrien‑Marie Legendre published the method of least squares, and in 1809 Carl Friedrich Gauss used it to predict the orbits of celestial bodies – arguably the first data‑driven model. Thomas Bayes’ theorem (1763) later became a cornerstone of probabilistic learning.
The modern era began in the 1950s: Arthur Samuel coined the term “machine learning” in 1959 while building a checkers‑playing program that improved with experience. Frank Rosenblatt’s perceptron (1958) was an early neural network. After a quiet period (“AI winter”), the 1980s brought back‑propagation, decision trees, and support vector machines. The 21st century’s explosion of data, powerful GPUs, and deep learning architectures (e.g., AlexNet in 2012) sparked the current revolution.
Introduction
There are a lot of problems that you can’t solve them without machine learning techniques (easily). For example, suppose a program for face detection purpose. For a 32x32 pixel with 8 bits of depth (grayscale image), there are 32x32x28 = 262,144 different probabilities. Now, suppose a 2MP color image! You may think that you can find the curvature of a face but you find it is also hard to define it. Here machine learning comes to solve the problem. You should prepare several faces and non-face images and write tens lines of code (of course with help of a library such as OpenCV). I should note that do not imagine machine learning as magic to solve everything. It has its pros and cons also.
Today, thanks to many advances in algorithms, electronics, and a huge amount of data expanded over the Internet, developing AI programs are more available. Machine learning techniques are used in finding networking attacks, face detection, spam detection, medicine discovery, law enforcement, and even finding new types of algorithms. Recently a new type of machine learning algorithm can generate and invent something new. For example, finding a new formula for medicines, generating a new face, generating new models of clothes
Smart Cities
By growing population, customary or old approaches are not good enough to handle traffic jams, car pollutions, long queues of people registering for events, etc. How many police officer does it need to do law enforcement in roads? How many workers do you need to manage public parking? To cope with such problems you can use ML. For example, a license plate detection program can be a solution for the problems. ITS (intelligent transportation system) is a solution for car-based problems such as traffic jams, emissions, transportations cost, etc.
 |
Fig. 1, depicts the architecture of hierarchical and networked vehicle surveillance which can be used in ITS. As it can be seen image acquisition in the first layer is a base for electronic toll collection, security monitoring, etc in the 4th layer. In the second layer, we must detect, recognize, or track cars. One property of each car that you can perform detection, recognition, and tracking is license plate. because of this, license plate detection plays a key role in ITS and is a trade in many ML journals, especially image processing related ones.
Medical
Over the past decade, there has been a remarkable increase in the amount of available compound activity and biomedical data owing to the emergence of new experimental techniques. ML can help to test a simulated drug on a simulated disease. For example, chemistry experts can simulate behaviors of a virus using ML and also test a simulated drug on the virus. This process also helps drug specialization for different patients leads to better effectiveness and lower side effects. Medical experts can even predict side effects by ML [https://healthitanalytics.com/news/machine-learning-method-identifies-adverse-drug-side-effects]
In the field of medical imaging, ML helps doctors to find tumors easily, especially when they are tired. Also in many areas, expert doctors are not available at all times. Using a ML program detecting tumors may help people, at least temporarily. In radiology facilities, ML helps to use less radioactive material using image enhancement techniques or you can use less power of x-rays for imaging.
 |
 |
| A | B |
Fig. 2, shows an example ML for image enhancements. Part A in Fig. 2, depicts an unenhanced x-ray image. It can be realized that is matt while enhanced imaged in part B is more pleasant in our eyes. We can see details in B better than A.
Modeling
Many people think that AI or ML applications are limited to classification or detecting something. AI applications are far away from detecting an object, tracking them, or classification. Many approaches of the ML help decide business action. Recently, a new type of ML tool lets you generate something new. For example, you can feed a lot of videos as training samples by caption and then ask a ML program for a new video by a caption. Another example is modeling that you can generate new models.
 |
| A |
 |
| B |
Part of A of Fig. 3, depicts several input image samples. A ML model (generative adversarial network) can be trained using such images. Part B is a possible output of such networks. As it can be seen there are all new models. So we can see a limited amount of innovation here.
Stock Trading
Many have seen time-machine-related movies. Imagine you can travel to a future time and see what will be happed to stock prices! ML is a tool for time travel! To be more precise you extract a model behavior of stocks. Such problems are known as time-series problems, simply because the price varies according to time. you can suppose it as a regression problem.
 |
Fig.4, depicts a simple prediction using an AI model which is downloaded from https://www.blockchain.com/charts/market-price. As you can see it is not accurate in terms of predicting details. If you need to predict more details you need complex models and more data. for example, you can predict better with more years or knowledge about fundamental events in bitcoin usage.
Steps of a Machine Learning Pipeline
Building a reliable ML system follows a series of well‑defined stages. Although the exact order may vary, a typical pipeline includes:
1. Data collection & gathering – acquiring raw data from sensors, databases, web scraping, or public repositories.
2. Data cleaning – handling missing values, correcting errors, removing duplicates, and dealing with outliers.
3. Data labelling – assigning ground‑truth tags (e.g., “spam” or “not spam”). This can be done manually, semi‑automatically, or via weak supervision.
4. Feature extraction & engineering – transforming raw data into meaningful numerical representations (e.g., edges in an image, word frequencies in text, or embeddings from a neural network).
5. Model selection & training – choosing an algorithm (decision tree, neural network, etc.) and learning the parameters from the training data.
6. Decision / inference – using the trained model to make predictions on new data.
7. Evaluation – measuring performance with metrics such as accuracy, precision, recall, F1‑score, or mean squared error, typically on a held‑out test set.
Challenges and How Scientists Mitigate Them
Data gathering & labelling bottlenecks: Obtaining large, high‑quality labelled datasets is expensive and time‑consuming. Researchers tackle this with data augmentation (creating synthetic variations), transfer learning (reusing pre‑trained models), and weak supervision (noisy, programmatic labels). Active learning helps by intelligently selecting which samples a human should label next.
Privacy issues: Medical, financial, or personal data cannot be freely shared. Solutions include federated learning (training models across decentralized devices without centralizing data), differential privacy (adding calibrated noise to obscure individual records), and homomorphic encryption (computing on encrypted data).
Security breaches & adversarial attacks: ML models are vulnerable to adversarial examples – carefully crafted inputs that fool a classifier (e.g., a sticker that makes a stop sign invisible to a self‑driving car). Model extraction and membership inference can leak sensitive training data. Recently, even large language models like Claude have faced sophisticated jailbreaking and prompt‑injection attacks that circumvent safety filters, exposing harmful content. Mitigations include adversarial training, input sanitization, red‑teaming, and continual safety fine‑tuning with reinforcement learning from human feedback (RLHF).
Un‑interpretable features: Deep neural networks often behave as “black boxes,” learning features that are incomprehensible to humans. This lack of transparency is unacceptable in critical domains like medicine or law. Scientists counter this with explainable AI (XAI) methods: LIME and SHAP highlight which input features most influenced a prediction; saliency maps and attention visualizations show where a model focused; and inherently interpretable models (e.g., small decision trees, sparse linear models) can be used when possible.
Other Notable Applications
Natural Language Processing (NLP): Chatbots, translation, sentiment analysis, and text summarisation all rely on ML. Large language models (GPT, Claude) can draft articles, answer questions, and even write code.
Recommendation systems: Streaming platforms, online shops, and social media use ML to suggest movies, products, or connections, dramatically shaping user experience.
Autonomous vehicles: Self‑driving cars fuse data from cameras, LiDAR, and radar, using deep learning to detect objects, plan paths, and make real‑time decisions.
Robotics & manufacturing: Industrial robots learn to grasp objects, sort items, and detect defects with reinforcement learning and computer vision.
Climate science & agriculture: ML improves weather forecasting, models crop yields, monitors deforestation via satellite imagery, and optimises irrigation.
Cybersecurity: Anomaly detection models flag network intrusions, malware, and phishing attempts far faster than manual rules.
Resources
Books:
• Pattern Recognition and Machine Learning – Christopher M. Bishop
• Deep Learning – Ian Goodfellow, Yoshua Bengio, Aaron Courville
• Hands‑On Machine Learning with Scikit‑Learn, Keras, and TensorFlow – Aurélien Géron
Online courses:
• Machine Learning by Andrew Ng (Coursera)
• Practical Deep Learning for Coders (fast.ai)
• Deep Learning Specialization (deeplearning.ai)
Libraries & frameworks:
• Scikit‑learn – general‑purpose ML
• TensorFlow & PyTorch – deep learning
• OpenCV – computer vision
• Hugging Face Transformers – NLP
Datasets:
• ImageNet – image classification
• UCI Machine Learning Repository – diverse datasets
• Kaggle Datasets – community‑driven collection