A4 Machine Learning

Supervised learning

• Trained on labelled data: every example has a known correct output • The model adjusts until its predictions match the correct outputs reliably • Needs labelled data and a defined target output • Use for: predicting a known label (fraud vs legitimate, spam, diagnosis, positive/negative reviews) • Example: medical images labelled fracture / no fracture • Only as good as its training data: biased or incomplete data gives biased predictions; too little data means poor generalisation to new inputs

Unsupervised learning

• Learns from unlabelled data with no predefined outputs • Finds structure, groupings or patterns by itself • Techniques: clustering and association • Use when labels are unavailable or you want to discover groups nobody defined in advance • Examples: customer segments, hidden population groups, patient symptom groups

Clustering

• Unsupervised technique • Groups similar data points together based on shared characteristics • Examples: customers with similar buying behaviour, patients with similar symptoms, unknown DNA subgroups

Classification

• Supervised task • Assigns each input to a category label • Examples: fraudulent / legitimate, positive / negative / neutral, healthy / diseased

Association

• Unsupervised technique • Finds relationships between variables • Example: items that are frequently bought together

Reinforcement learning

• Trains an agent to make decisions by placing it in an environment • The environment gives rewards for good actions and penalties for bad ones • The agent learns the sequence of actions that maximises total reward • Use for: sequential decisions with a clear goal and feedback, often practised in simulation • Examples: robot navigation, self-driving in simulation, chess by self-play

Deep learning

• Branch of machine learning using artificial neural networks (ANNs) with multiple layers • Each layer extracts more abstract features from the input • Handles complex data such as images, audio and text • Needs a lot of data and computing power • Used in: medical imaging, object detection, natural language processing, sentiment analysis

Transfer learning

• Takes a model already trained on one task and adapts it to a different but related task • Needs less new labelled data and shortens training time • Use when labelled data is limited (e.g. a rare disease with few scans) and a related pre-trained model exists • Example: start from a deep model trained on broad medical images, adapt it to one disease

CPU

• Central processing unit: general-purpose processor • Runs the operating system, program logic and input/output • Processes instructions mostly one after another (sequential) • Fast at complex, branching tasks; high flexibility • Weak for ML training: training repeats the same matrix maths millions of times in parallel, so a few sequential cores are far too slow • Enough for small-scale ML or learning projects (a standard laptop) • Primary use: general computing, logic, operating systems

GPU

• Graphics processing unit: thousands of simple cores for parallel processing • Originally built to render graphics (the same calculation on thousands of pixels at once) • Ideal for training deep neural networks (lots of similar matrix calculations at once) • Can cut training time from weeks to hours • Trade-off: high power consumption • Flexibility: medium; primary use: model training, parallel computation

TPU

• Tensor processing unit: an ASIC built to accelerate tensor computation (the maths underneath most ML algorithms) • More energy-efficient than GPUs for these operations • Widely available through cloud platforms for large-scale training and running models • Developed by Google for its deep learning workloads • Why build one: it does only the calculations neural networks need, faster and with much less power than a GPU; at Google's scale small gains per chip save huge amounts of time, electricity and data-centre size and give a competitive advantage • Trade-off: less flexible than a GPU (low flexibility)

ASIC

• Application-specific integrated circuit: a chip designed to do one particular task with maximum efficiency • Cannot be reprogrammed or repurposed • High performance with low power consumption within its task • Example: the neural processing chip in a smartphone (face recognition, image enhancement) without draining the battery; it cannot run a word processor • Flexibility: very low; primary use: specific inference tasks

FPGA

• Field-programmable gate array: a chip whose logic circuits can be reconfigured after manufacture by uploading a new configuration • More flexible than an ASIC • Use when physical access is impossible and the hardware must be updated • Example: a satellite analysing crop imagery; engineers upload an updated configuration from the ground • Trade-off: more complex to program than an ASIC • Flexibility: medium

Cloud-based platform

• Scalable processing power accessed over a network • Rent hundreds of GPUs or TPUs on demand and pay only for what you use • No investment in hardware or data centres, so large-scale training is accessible to small teams • High flexibility • Trade-offs: latency, ongoing cost, requires connectivity; data leaves your premises • Primary use: large-scale training

HPC centre

• High-performance computing centre: a facility with thousands of networked processors • For the most computationally intensive tasks (training large language models, climate models on decades of data) • Maximum throughput • Trade-offs: expensive to access; latency because processing happens remotely

Edge device

• Processes data locally at the point where it is collected, instead of sending it to a remote server • Examples: smartphones, autonomous vehicles, drones, IoT sensors • Benefits: low latency, keeps sensitive data on the device, works without internet • Trade-off: limited processing power and memory compared with cloud or HPC • Best for: running a trained model (inference) locally; flexibility: low