Understanding Artificial Intelligence (AI)

What is Artificial Intelligence (AI)?

• AI includes technologies that enable computers to perform advanced functions.
• These functions encompass seeing, understanding, translating languages, analyzing data, and making recommendations.
• AI is concerned with constructing machines and computers capable of reasoning, learning, and acting like humans.
• AI can handle data analysis at scales beyond human capabilities.

Timeline of AI

• Foundations in neural networks were set in 1943 by Warren McCulloch and Walter Pitts.
• They drew parallels between the brain and computing machines.
• Alan Turing introduced the Turing test in 1950 as a way to test the intelligence of a machine.
• The term "Artificial Intelligence" was coined in 1955 during a conference.
• ELIZA, a natural language program, was created in 1965.
• ELIZA was designed to handle dialogue on any topic, similar to modern chatbots.
• Edward Feigenbaum created expert systems in the 1980s, which emulate decisions of human experts.
• In 1997, Deep Blue, a computer program, defeated world chess champion Garry Kasparov.
• iRobot launched Roomba in 2002, an autonomous vacuum cleaner.
• Google developed the first self-driving car in 2009 for urban environments.
• IBM's Watson won the US game show Jeopardy! in 2011.
• From 2011-2014, personal assistants like Siri, Google Now, and Cortana were developed.
• They used speech recognition to answer questions and perform tasks.
• Ian Goodfellow introduced Generative Adversarial Networks (GAN) in 2014.
• AlphaGo beat professional Go player Lee Sedol 4-1 in 2016.
• Most universities offered courses in Artificial Intelligence by 2018.
• Computer vision began switching to neural nets and VAEs around 2012-2014.
• Transformer architecture was introduced around 2017-2018.
• Advancements like GPT-1, BERT, and Graph Neural Networks occurred around 2018-2019.
• This was followed by GPT-2 with improved Generative Models.
• GPT-3, emphasizing self-supervised learning, came around 2019-2020.
• Projects like AlphaFold 2, DALL-E, and GitHub Copilot were developed around 2021-2022.
• ChatGPT and Stable Diffusion emerged around 2022-2023.

Types of AI

• Narrow AI (Weak AI) involves AI that performs specific tasks within limited constraints.
• General AI (Strong AI) refers to AI that can understand, learn, and apply intelligence like a human.
• Superintelligent AI surpasses human decision-making and has superior intellectual processing abilities.

Narrow AI (Weak AI)

• Designed for particular tasks within a limited scope.
• Lacks consciousness, feelings, or the ability to do tasks not specifically programmed.
• Excels in designed tasks but cannot generalize knowledge or learn new things.
• Image and speech recognition software, self-driving cars, and spam filters are examples.

General AI (Strong AI or AGI)

• Possesses the ability to understand, learn, and apply intelligence to any problem like a human.
• AGI would be flexible, capable of reasoning, solving problems, planning, learning, communicating, and operating independently.
• General AI is a hypothetical achievement with no consensus on its capabilities or how to achieve it.
• Some experts believe AGI is possible, while others see it as impossible or undesirable.

Superintelligent AI

• Exceeds human decision-making capacity.
• Exhibits enhanced capabilities in creativity, social adaptation, and intellectual processing compared to humans.

Another Classification of AI

• Reactive machines are limited AI that respond to stimuli based on preprogrammed rules.
• They lack memory and learning capabilities, illustrated by IBM's Deep Blue beating Garry Kasparov in 1997.
• Limited memory AI uses memory to improve via training on new data through neural networks.
• Deep learning, a machine learning subset, is an example of limited memory artificial intelligence.
• Theory of mind AI, currently theoretical, emulates the human mind with equivalent decision-making abilities.
• This includes recognizing emotions and reacting in social situations.
• Self-aware AI, also theoretical, possesses awareness of its existence and human-like emotional/intellectual capabilities.

Basic Operation of AI

• Five steps included are data, algorithms, models, learning, and implementation.

Data

• AI relies on data, the raw material essential for training AI systems.
• Data can be text, images, audio, video, or sensor information.
• The amount and quality of data are crucial for the success of AI systems.

Algorithms

• Algorithms serve as sets of instructions.
• They process data and extract information.
• Algorithms rely on mathematical and statistical techniques.
• There are varied types for different AI tasks like classification, regression, clustering, and anomaly detection.

Model

• AI models are computer programs trained on data for specific tasks.
• The model's parameters are adjusted during training to achieve the highest accuracy.
• AI models include artificial neural networks, decision trees, and probabilistic models.

Learning

• Supervised learning involves training a model with labeled examples showcasing relationships between inputs and outputs.
• Unsupervised learning involves the model that learns from unlabeled data to find patterns and structures.
• Reinforcement learning involves the model learns through environmental interaction, gaining rewards for goal-oriented actions.

Implementation

• AI models are implemented in software or hardware.
• They are designed for real-world applications.
• Implementation can require specialized expertise in software and hardware engineering.

Fields of Study in Artificial Intelligence

• Includes machine learning (ML), deep learning (DL), natural language processing (NLP), computer vision, and planning and automatic reasoning (PAR).

Machine Learning (ML)

• Machine learning is a branch of AI focused on using data and algorithms to imitate human learning.
• Machine learning is a statistical method for learning patterns based on examples.
• Machine learning algorithms are trained on large labeled datasets.
• After this they can be used to predict new data to identify spam, classify images, or predict customer behavior.

Deep Learning (DL)

• Deep learning is an ML technique using artificial neural networks to learn data patterns.
• It is inspired by the human brain.
• It can learn from large and complex datasets.
• Deep learning algorithms are trained on large labeled datasets.
• This data then is used to make predictions on new data to identify image objects, translate languages, or generate text.

Natural Language Processing (NLP)

• NLP focuses on the interaction between computers and human language.
• Its goal is to enable machines to comprehend, interpret, and respond to human texts and voices naturally.
• NLP applications include machine translation, speech recognition, text summarization, question answering, and sentiment analysis.
• Machine translation automatically translates text from one language to another.
• Speech recognition automatically converts spoken language into text.
• Text summarization automatically generates short summaries of text.
• Question answering automatically answers questions posed in natural language.
• Sentiment analysis automatically determines the sentiment of a text (positive, negative, or neutral).

Computer Vision

• Computer vision enables machines to "see" or identify and understand visual content.
• This involves processing images and videos to identify objects, people, scenes, and activities.
• Computer vision applications include object and face detection, scene understanding, medical imaging, and robotics.
• Object detection: automatically identifying objects in images or videos.
• Face recognition: automatically identifying people in images or videos.
• Scene understanding: automatically understanding the content of a scene.
• Medical imaging: using computer vision to analyze medical images.
• Robotics: using computer vision to enable robots to interact with the world.

Planning and Automatic Reasoning (PAR)

• PAR is a field of AI focused on machines representing knowledge and using it to achieve specific goals.
• Automatic Reasoning (AR) is a PAR area for developing systems that can understand and apply logic.
• These systems solve problems, prove theorems, and make decisions based on available information.
• AR aims for computers to "reason" similarly to humans at scales and speeds surpassing human capabilities.
• Self-driving cars use PAR to navigate and avoid obstacles.
• Recommendation systems like Amazon and Netflix use PAR to suggest products.
• Medical diagnosis systems use PAR to analyze symptoms and medical data.

Vocabulary

• Algorithm, Artificial intelligence, BIAS (in IA), Chatbot, Generative pre-trained transformer (GPT), Machine Learning, Neural Networks (NN)

Sources

• Google: What is AI?, CIRCLS, Gemini (Bard), ChatGPT