What is Ambient Intelligence?
The core idea: technology that disappears into the environment. The environment becomes intelligent and adaptive — a network of "hidden interfaces" that understands your behavior and adapts to your needs.
AmI is a multidisciplinary area. Its environment is aware of the current context and ready to act on it.
The 5 components of an AmI system
Sensing
Capture info about the environment & users
Acting
Act on the environment (lights, thermostat, robots)
Reasoning
Model & recognize behavior — connects sensing to acting
HCI
Natural interaction: gestures, speech, emotions
Security & Privacy
Protect sensitive sensed data from attacks
Distributed → Pervasive → Volatile
AmI is built on distributed pervasive systems.
- Distributed system: a collection of connected independent computing nodes that appears to users as a single coherent system.
- Pervasive system = distributed system that (1) includes unconventional nodes (mobile objects with computing + communication — phones, smart appliances, sensors) and (2) is adaptive (adjusts behavior to current context to optimize its goal).
- Volatile: pervasive systems suffer device/comm failures, bandwidth changes, and constant creation/destruction of software associations.
IoT & Domotics
Why are objects "smart"?
- Connected to the Internet → remote access, invocation of services, cooperation
- Run algorithms (locally or remotely) to analyze data and understand context using AI
- Offer personalized context-aware services
The AmI / IoT processing flow
IoT layer (sense/respond) · AI & ML layer (process/interpret/decide) · HCI: speech, visual, movement, gesture
Domotics = IoT + Home
Domotics is the use of IoT devices (sensors & actuators) in the home environment, to transform it into a smart-home. The most common application is automations:
But domotics is more than automations: remote healthcare monitoring, energy management, security, anomaly detection.
Context & Context-Awareness
⭐ Heavily testable. Traditional software forces users to make all parameters explicit. AmI systems acquire context automatically.
👉 Simplified: "all the data useful to adapt a service." Early enumeration-based definitions (context = location + time + people…) were rejected because they were not general enough.
Key terms
- Context-aware system: uses context to provide relevant info/services to the user (relevancy depends on the user's task). AmI systems are context-aware systems.
- Adaptiveness: the property of adapting to a given context for a better service. Crucial because of changes in user situation, environment, network connectivity, and energy/battery.
- Awareness mismatch: the gap between human perception and the system's perceived context. Goal: minimize it.
Taxonomy of context (U=user · S=sensors · D=device · O=other)
| Primary dimension | Examples (secondary) | Sources |
|---|---|---|
| User | Identity, physiological, emotional, interests, preferences, social, activity | U, S, O |
| Environment | Weather, temperature, humidity, light, noise, air pollution, surrounding persons/devices | S, O |
| Location | Physical, symbolic | S, O |
| Time | Absolute, granularity-based | D, S, O |
| Device | Capabilities, status | D |
| Connectivity | — | D, O |
Temporal context & context history
Not only time of day/season — context history matters. Past data lets you derive new context (sequence of positions → trajectory) and predict context (recurrent trajectory at a time → predict destination).
Two levels of context data ⭐
| Level | How it's obtained | Examples |
|---|---|---|
| Low-level | Directly from sensors, or by simple processing / fusion of raw data | Raw temperature; bandwidth average over a window; "Hot&Humid" if avg temp > 30°C AND humidity > 80% |
| High-level | Derived by applying AI models | Activity "cooking a meal" from smart-home sensors; mood from activities + Galvanic Skin Response + face recognition |
Applications of AmI
Smart Homes (Domotics)
Activity analysis → automations, ambient assisted living, energy management
Smart Energy
Analyze appliance consumption → suggest strategies, save energy, optimize for smart cities
Smart Buildings
Security, fire safety, lighting, 24/7 monitoring, HVAC, energy
Pervasive Transport
Road/car sensors + V2V comms → self-driving, mobility, emission & hazard reduction
Pervasive Health
Wearables + env sensors → early detection, chronic monitoring, rehab, independent living
Smart Cities
City-scale integration of all the above
Challenges of AI in AmI
⭐ These six challenges connect directly to later lectures and are prime exam material.
1 · Unbalanced data
Human behavior is rarely balanced (e.g., "walking" ≫ "taking stairs"). Consequence: accuracy is misleading — a model predicting only the majority class scores high but is useless. → Use the F1 score.
2 · Labeled data scarcity
Annotating real human data is costly, time-consuming, and intrusive. Unlike computer vision (ImageNet etc.), there are no huge public datasets of real in-home behavior.
3 · Personalization
Models must be tailored to the specific subject/environment — sensor patterns vary a lot between people/homes (unlike CV where pretrained models often transfer). ⚠️ This creates a circular dependency with challenge #2: personalizing needs per-user labels, but labels are scarce.
4 · Continual learning
Behavior & environments drift over time → a model trained once isn't enough. Must update continuously without retraining from scratch. Problem: neural nets suffer catastrophic forgetting of old information.
5 · Privacy
Sensitive data (activities, habits, energy patterns, health) may leak to untrusted cloud services or attackers. Privacy must be designed in from the start.
6 · Resource-constrained devices
Models sometimes must run on phones/microcontrollers (TinyML) instead of the cloud — to reduce latency, work offline, guarantee privacy, and cut costs. Challenge: reduce model complexity while keeping accuracy.
Final Quiz — Exam Style
Answer all questions. MCQs give instant feedback; open-ended questions reveal model answers. Your MCQ score appears at the bottom.
a) Dey: "Context is any information that can be used to characterize the situation of an entity (person, place, or object) relevant to the interaction between a user and an application." Early definitions just listed attributes (location, time, surrounding people…). They were rejected because they were not general enough — any new application with different needs would require a new list, making them domain-specific and incomplete.
b) Low-level context comes directly from sensors or simple fusion — e.g., averaging temperature readings to get the room's ambient temperature. High-level context is derived by applying AI models — e.g., inferring the activity "cooking a meal" by combining motion, oven power, and smart-meter data through a classifier.
a) AmI analyzes data from real users in real environments. Getting labels means recruiting people, monitoring daily life, and manually annotating events — costly, time-consuming, and intrusive. Unlike computer vision, there are no massive public labeled datasets (like ImageNet) of real in-home human behavior, so each new deployment faces scarce labels.
b) Models must be personalized because sensor patterns vary strongly between subjects/environments (one person's "walking" differs from another's). But personalization needs labeled data from that specific user/environment — which is exactly what's scarce. So you need labels to personalize, yet getting per-user labels is the very problem you can't easily solve. Transfer learning and self-supervised learning aim to break this cycle.