Machine learning

1. Meaning and conceptual foundations

Machine learning, or ML, develops computational models whose performance improves through learning from data or experience rather than through manually written rules for every situation. Conventional software typically applies explicit instructions to inputs. An ML system instead estimates patterns from examples and applies those patterns to unfamiliar inputs. Both approaches can coexist: a bank may combine fixed transaction limits with an ML model that detects unusual payment behaviour.

Artificial intelligence is the broader endeavour of building systems capable of tasks associated with intelligence, including perception, reasoning, language processing and planning. Not all AI uses machine learning: rule-based expert systems are an important counterexample. Deep learning is a subset of ML that uses neural networks with multiple processing layers to learn increasingly complex representations of information.

A feature is an input attribute, such as rainfall, transaction value or a pixel intensity. A label is the target associated with a training example, such as whether a transaction was fraudulent. Model parameters are internal quantities adjusted during learning. A loss function measures error relative to a training objective, and an optimisation procedure adjusts parameters to reduce that error. The learned relationship is not necessarily a causal explanation of the underlying phenomenon.

• Algorithm: the computational procedure used for learning or prediction.
• Model: the learned mathematical representation produced through training.
• Dataset: a collection of examples used for training, validation or testing.

2. Main learning approaches

Supervised learning uses examples containing both inputs and desired outputs. Classification predicts a category, such as spam or non-spam email. Regression predicts a numerical value, such as crop yield or electricity demand. Common methods include linear regression, logistic regression, decision trees, random forests and support vector machines. Despite its name, logistic regression is generally used for classification.

Unsupervised learning analyses data without predefined target labels. Clustering groups similar observations, such as consumers with comparable purchasing patterns. Dimensionality reduction represents data using fewer variables while retaining important information. Anomaly detection identifies unusual observations and can use supervised, semi-supervised or unsupervised approaches depending on the availability of labels.

Semi-supervised learning combines a relatively small labelled dataset with a larger unlabelled dataset. Self-supervised learning derives training targets from the data themselves: a language model may learn by predicting missing or subsequent text. It is important in modern foundation models because large quantities of unlabelled text, images and audio are available.

Reinforcement learning involves an agent acting in an environment, receiving feedback through rewards and learning a policy that seeks higher cumulative reward. Applications include game-playing, robotic control and resource allocation. It differs from ordinary supervised learning because a correct answer need not be supplied for each action, and the consequences of an action may emerge much later.

3. Model development and performance

An ML project begins with a clearly defined problem and representative data. Data preparation may involve correcting errors, handling missing values, removing duplicates and creating suitable features. Data are commonly separated into training, validation and test sets. Training fits the model; validation guides model selection and tuning; the test set provides a separate assessment of performance. Repeatedly tuning against test results weakens this assessment.

Overfitting occurs when a model learns peculiarities or noise in training data and performs poorly on new examples. Underfitting occurs when it fails to capture important patterns even in the training data. Regularisation, appropriately chosen model complexity, additional representative data and cross-validation can help improve generalisation. Data leakage occurs when information unavailable at the time of real prediction improperly enters training or evaluation.

Performance must be assessed using measures suited to the task. Accuracy alone can be misleading when classes are imbalanced: if only one per cent of transactions are fraudulent, predicting every transaction as legitimate produces 99 per cent accuracy but detects no fraud. Precision measures how many predicted positives are actually positive; recall measures how many actual positives are detected. Medical screening often places strong emphasis on avoiding missed cases.

After deployment, the trained model performs inference on new inputs. Its performance can deteriorate as populations, behaviours or operating conditions change, requiring monitoring and possible retraining. Training large models can demand substantial computing capacity, electricity and cooling. Inference also consumes resources, while smaller models can operate on mobile phones or sensors through edge AI.

4. Deep learning, generative AI and applications

Deep neural networks are particularly useful for complex, high-dimensional information such as images, speech and text. Convolutional neural networks are strongly associated with image analysis, while transformer architectures underpin many contemporary language and multimodal models. Graphics processing units, or GPUs, accelerate many ML operations because they support extensive parallel computation; they are not mandatory for every ML task.

Generative AI produces content such as text, images, audio or computer code. Predictive and discriminative systems instead commonly estimate outcomes or distinguish categories. A large language model generates text based on patterns learned during training and its current context. Fluent output does not establish factual correctness: models may generate plausible but false statements, often called hallucinations. Foundation models can be adapted to multiple downstream tasks rather than being built exclusively for one narrow application.

Indian applications include satellite-image analysis for crop monitoring, disease screening from medical images, fraud detection in digital payments, electricity-load forecasting and speech recognition for public services. BHASHINI, the National Language Translation Mission launched in 2022, supports language technologies for Indian languages. Such systems can widen access, but performance must be checked across dialects, accents and user groups.

ML differs from several associated technologies. Big data concerns the scale and complexity of datasets and their management; cloud computing supplies shared computing resources; and the Internet of Things connects physical devices that collect or exchange data. These technologies can support ML, but none is synonymous with it. Nor does an ordinary ML system require quantum computing or a blockchain.

5. Risks, safeguards and Indian policy context

Biased or unrepresentative data can produce discriminatory outcomes in lending, recruitment or welfare administration. Historical decisions used as labels may themselves embody exclusion. Explainability concerns whether humans can understand the factors behind an output, while accountability concerns who bears responsibility for its use. Human oversight is particularly important when errors can affect health, liberty or access to essential services.

Privacy risks arise from collecting sensitive data, combining datasets and possible disclosure of training information. Security threats include data poisoning, which manipulates training data, and adversarial inputs designed to mislead a model. Federated learning allows distributed participants to train without centrally pooling all raw data, but does not by itself guarantee privacy. Differential privacy can provide quantified protection under specified assumptions by introducing carefully calibrated randomness.

India's National Strategy for Artificial Intelligence, released by NITI Aayog in 2018, emphasised AI for All and applications in healthcare, agriculture, education, smart cities and mobility. The IndiaAI Mission supports computing access, datasets, indigenous AI capabilities, skills, applications, startup financing and safe and trusted AI. The Digital Personal Data Protection Act, 2023 provides a statutory framework relevant to digital personal-data processing; it is not a comprehensive AI-specific law. Responsible adoption requires purpose-appropriate data, independent evaluation, security safeguards and accessible grievance mechanisms.

Previous year questions

No source-linked PYQ is available here yet. Use the practice questions below.

Practice questions