Artificial Intelligence — a top-down view
An introduction to basic concepts in AI
Ever since I got interested in Artificial Intelligence (AI), I have been following the topic with keen interest. But in most cases, I’ve seen a relatively narrow take on AI by the author, probably based their knowledge or interest. Hence, this write-up on AI with a top-down view of basic concepts.
AI is a branch of computer science that deals with methods to make machines do things that are normally expected from human intelligence. Ability to perceive, converse, know, learn, reason, plan and move are some areas. Machines that bring aspects of AI are called intelligent agents (or simply ‘agents’).
The chart below provides the top-down view of AI concepts:
1. Perception
While humans have 5 basic senses of sight, hearing, smell, taste and touch, machines can mostly mimic only 2 or 3 out of these.
Computer vision is a field used to analyse visual inputs with sub-problems such as facial recognition, object recognition, optical character recognition (OCR) and gesture recognition.
Speech recognition is a field used to convert recognise and translate spoken language into text. Such text usually goes NLP to understand the intent.
Touch: though machines cannot sense touch the way humans do, there are many related ones like touchscreen sensor, thermometer, barometer, and humidity.
Other sensors like accelerometer, compass, gyroscope, proximity, GPS, fingerprint, ambient light, are usually available in a smartphone.
2. Natural Language Processing (NLP)
The semantics and lexical affinity of words play a big part in NLP to understand the intent. For example, “Is it cold?” may appear to be a very normal question to us, but a machine would need the context or knowledge base to determine that we’re asking about the temperature in a room.
Applications of NLP may include machine translations, parsing documents to retrieve information, text mining or analytics, and answering questions in natural language such as chatbots to order pizza, book flights and hotels, or get an insurance quote or claim, and virtual assistants such as Alexa and Siri.
3. Knowledge Representation
Knowledge representation and knowledge engineering become extremely important to address solutions in a domain involving cognitive skills. For example, an AI system at a hospital would need to know that pug is a breed of dog in order to associate it with the right documentation about dogs. Knowledge repositories capture domain-specific objects, properties, categories and relations between objects, situations, events, states and time, causes and effects. Such a representation is known as an ontology.
Common knowledge repositories for general knowledge, common sense, lexical database, etc. are available for use in many use cases. AI systems need also the ability to auto-update their ontology model based on experiences.
4. Learning
Machine Learning (ML) is a vast and highly developed area of AI, primarily based on big-data and statistical methods. Classification, prediction, fraud detection, etc. are common ML problems. Popular ML approaches are broadly classified as supervised (requiring human input) and unsupervised. Other approaches include reinforcement learning (based on reward and punishment), feature learning, anomaly detection, and association rules.
Google search autocomplete is learned from past searches. eCommerce sites learn what is frequently bought together to make a recommendation.
Machine learning is usually achieved by creating a model, which is first taught through a training data set and is then validated with a test data set.
Deep Learning is an ML approach, based on artificial neural networks (ANN) that deduce the features and then use them to solve a problem.
Symbolic Learning is based on signs that can be perceived by humans as against data and statistics. Look for my top-down view on Machine Learning.
5. Reasoning, Problem Solving
To automate some type of logic or decision, a reasoning system may use a knowledge-base and logical techniques such as deduction and induction. There are also several problem-solving strategies to achieve a goal based on a given set of rules. Mechanical robots and game bots use such techniques.
Heuristic algorithms can imitate step-by-step reasoning to give an advantage without necessarily analysing all the possibilities. Probabilistic methods can be used to deal with uncertain or incomplete information. However, a machine still can’t make intuitive judgements like humans.
6. Planning
AI Planning or Automated Planning and Scheduling involves the realization of strategies or action sequences by autonomous agents. The intelligent agent has to visualize the future state of the world based on how their actions will change it, and make choices that maximize their utility value.
When there are extraneous factors, the agent should assess the environment to make predictions and also evaluate the predictions and adapt based on their assessment. The solutions are usually complex and must be discovered and optimized in multidimensional space. Multi-agent planning uses the cooperation and competition of many agents to achieve a given goal.
AI Planning has applicability in several areas such as route or schedule optimisation and response prioritising.
7. Motion and Manipulation (Robotics)
A mobile robot is expected to perceive, navigate and move things within a physical environment. Some of the applications include industrial automation and robots in unsafe environments. Combining perception and mobility to achieve physical dexterity a very complex task and machines can seldom match to a human child. Motion planning involves breaking down a movement into individual joint or component movements and feedback.
A software robot simulates a computer user’s input through keyboard and mouse clicks to achieve Robotic Process Automation (RPA). Current RPA technologies go beyond screen scraping to achieve the same effects in a virtual environment (they do not require a physical screen). Look for my top-down view on RPA.
8. Social and General Intelligence
Humans are social beings. Affective computing is a broad field of study to recognize, process or simulate human affects such as feeling or emotion. Being able to predict the actions of others by understanding their motives and emotional states would allow an agent to make better decisions. Textual sentiment analysis and recognising human expressions are examples. Machines are still evolving to understand the intent behind human communication based on tone and accent or appreciate sarcasm.
Artificial General Intelligence (AGI) is the capacity to understand or learn any intellectual task that a human being can. Having commonsense, compassion, a concept of time, dealing with unexpected circumstances while solving any real-world problem.
Use cases in AI will typically have more than one of the above concepts. For example, a virtual assistant like Alexa would need speech recognition (Perception), NLP, and a Knowledge repository. Alexa also uses Learning techniques to respond to follow-up queries or to improve an action over time. The real power of AI comes from the combined application of these concepts.
Key challenges with AI have been to ignore fake data and avoid bias in decisions. To address these, Explainable AI or XAI is gaining prominence.
While areas such as Machine Learning, Perception (visual and speech), NLP and Robotics have documentation aplenty, I hope we see more in other areas as well — especially around Knowledge, Reasoning, and Planning. Please share your thoughts on this, and I shall soon attempt to write in more detail on each of the topics.