Tid

Lektionen forventes at tage ca. 20 min. at gennemføre. Dertil kommer refleksionsøvelserne, som du med stor fordel kan lave sammen med dine medstuderende.

Om lektionen 'Hvad er kunstig intelligens (og machine learning)?'

Lektionen om kunstig intelligens (AI) og machine learning giver dig en grundlæggende forståelse af AI som et generelt koncept og machine learning som en konkret teknologi, der driver AI. Gennem modulet vil du lære om, hvad vi kan bruge AI til – både som almindelige forbrugere og i videnskabelig sammenhæng. Du vil også blive introduceret til forskellige tilgange til at designe og udvikle AI, herunder skellet mellem generative modeller, der producerer ny data, og diskriminative modeller, der træffer beslutninger.

Læringsmål

Når du har gennemført lektionen, forventer vi, at du:

kan nævne eksempler på, hvad kunstig intelligens (AI) kan bruges til.
kan forklare forskellen på AI og machine learning (ML).
kan forklare forskellen på generativ og diskriminerende AI.
kan identificere de væsentligste kilder til bias i ML-baserede modeller.

Kilde

Lektionen er produceret af Københavns Universitet som en del af Københavns Universitets læringsressourcer til digital dannelse, 2022.

Fagansvarlig:

Silas Nyboe Ørting, postdoc, Datalogisk Institut

____________________________________________________________

Hvad er kunstig intelligens?

Når vi prøver at definere kunstig intelligens (artificial intelligence, AI), kommer det nemt til at handle om at definere 'intelligens' i det hele taget. Hvis vi f.eks. slår op på Wikipedia, får vi at vide, at: "Artificial intelligence (AI) is intelligence demonstrated by machines [...] as opposed to intelligence of humans and other animals", (Wikipedia/Artificial_Intelligence, 15. maj 2023).

Spørger vi i stedet AI-systemet ChatGPT, svarer den eksempelvis: "Artificial intelligence (AI) refers to the simulation of human intelligence in machines that are programmed to think and act like humans", (ChatGPT Mar 14 Version, 2023).

Kunstig intelligens er altså selve idéen om at programmere en computer til at 'demonstrere' eller 'simulere' menneskelig intelligens.

Da vi mennesker kan udvise intelligens på mange forskellige måder og med mange forskellige formål, må vi bryde fænomenet ned og se nærmere på nogle af de grundlæggende metoder, som forskere og udviklere arbejder med for at programmere computere til at løse konkrete opgaver 'intelligent'. Lad os starte med at se på, hvilke typer af opgaver kunstig intelligens kan hjælpe med at løse.

____________________________________________________________

Hvor støder vi på kunstig intelligens i vores hverdag?

Kunstig intelligens (AI) er integreret i mange af de tekniske løsninger, som vi bruger til hverdag, f.eks. når Netflix eller Spotify anbefaler indhold til os, når sociale medier sammensætter vores news feed, når vi bruger Google Translate, og når mobilkameraet optimerer de billeder, vi tager. I alle de nævnte situationer er vi i en vis forstand passive og ubevidste forbrugere af kunstig intelligens.

Derfor skabte det stor opmærksomhed, da den amerikanske virksomhed OpenAI lancerede den avancerede chatbot ChatGPT til offentligheden i slutningen af 2022. Pludselig kunne alle og enhver selv lege med en avanceret 'kunstig intelligens' og bruge den til alt fra programmering til universitetsopgaver.

Ny chatrobot kan klare din eksamen, rette din software og skrive dit essay (DR, 5. december 2022)

ChatGPT er en 'generativ AI' – altså en kunstig intelligens, der kan skabe indhold på brugerens kommando. Mange af de øvrige AI-løsninger, vi møder i hverdagen, bygger på 'diskriminativ AI' – altså kunstig intelligens, der kan skelne mellem forskellige karakteristika i store mængder data og træffe beslutninger på den baggrund.

Nedenfor har vi bedt ChatGPT om at give eksempler på, hvordan vi som almindelige forbrugere møder AI i hverdagen.

Klik på spørgsmålet for at læse chatbottens svar.

How do ordinary people encounter AI in their daily life?

ChatGPT - answer icon.png ChatGPT (2023, May 12 Version ) svarede:

Ordinary people encounter AI in their daily lives through various applications and services that have integrated AI technologies. Here are some common ways in which people interact with AI:

Virtual Assistants: Virtual assistants like Siri, Google Assistant, Alexa, and Cortana are AI-powered technologies that reside on smartphones, smart speakers, and other devices. They can perform tasks based on voice commands, answer questions, set reminders, play music, provide weather updates, and control smart home devices.
Smart Home Devices: AI is integrated into smart home devices such as thermostats, security systems, cameras, and lighting systems. These devices can learn from user preferences and adjust settings accordingly, providing convenience and energy efficiency.
Streaming Services: Platforms like Netflix and Spotify use AI algorithms to recommend movies, TV shows, and music based on users' viewing and listening history. They analyze user preferences and behavior to provide personalized content suggestions.
Social Media: AI algorithms are employed in social media platforms to curate users' news feeds and recommend content. They analyze user behavior, interests, and engagement patterns to deliver personalized and relevant content.
Online Shopping: E-commerce platforms use AI to personalize product recommendations based on users' browsing and purchase history. AI algorithms can also be used for price optimization, fraud detection, and chatbots for customer support.
Voice Recognition and Transcription: AI-powered voice recognition technologies are used in applications like voice-to-text transcription, voice search, and voice commands in smartphones, voice-controlled appliances, and navigation systems.
Navigation and Ride-Sharing Apps: Navigation apps like Google Maps and Waze utilize AI to analyze traffic patterns, optimize routes, and provide real-time updates on road conditions. Ride-sharing apps like Uber and Lyft employ AI algorithms to match drivers with passengers and estimate fare prices.
Online Customer Support: Many companies use AI-powered chatbots on their websites to provide automated customer support. These chatbots can answer frequently asked questions, provide product recommendations, and assist with basic troubleshooting.
Email Spam Filters: Email providers utilize AI algorithms to filter spam and detect phishing attempts. These algorithms analyze email content, sender reputation, and user feedback to identify and block unwanted messages.
Smartphone Cameras: AI is often used in smartphone cameras to enhance photo quality, apply filters and effects, and enable features like facial recognition, object detection, and augmented reality.

These are just a few examples of how ordinary people encounter AI in their daily lives. As AI continues to advance, its integration into various technologies and services will become even more pervasive.

Du kan læse mere om, hvordan ChatGPT virker på KUnet > Teknologien bag ChatGPT

____________________________________________________________

Hvad kan vi bruge AI til i videnskabelig sammenhæng?

En af styrkerne ved AI er teknologiens evne til at analysere store mængder data og finde mønstre, korrelationer og anomalier. Derfor kan AI bruges i mange forskellige videnskabelige sammenhænge. ChatGPT giver nedenstående bud på, hvordan AI bruges til videnskabelige formål.

How is AI used for scientific purposes?

ChatGPT - answer icon.png ChatGPT (2023, May 12 Version ) svarede:

AI is increasingly used in scientific research and exploration to accelerate discoveries, analyze complex data, and simulate intricate systems. Here are several ways AI is utilized for scientific purposes:

Data Analysis and Pattern Recognition: AI algorithms can analyze large and complex datasets in scientific fields like genomics, astronomy, and particle physics. AI helps identify patterns, correlations, and anomalies in the data that might not be immediately apparent to human researchers. This can lead to new scientific insights and discoveries.
Drug Discovery and Development: AI is employed in drug discovery to expedite the process of identifying potential drug candidates. AI algorithms can analyze vast amounts of molecular data, predict drug-target interactions, simulate molecular structures, and optimize drug design. This can help in the development of new treatments and therapies.
Genomics and Bioinformatics: AI plays a vital role in genomics research, where it assists in genome sequencing, variant analysis, gene expression analysis, and predicting protein structures. AI algorithms aid in understanding genetic diseases, identifying biomarkers, and facilitating personalized medicine.
Climate Modeling and Environmental Analysis: AI is used in climate modeling to simulate complex climate systems, predict weather patterns, and assess climate change impacts. AI algorithms analyze environmental data from satellites, weather stations, and sensors to generate accurate predictions and models.
Robotics and Automation: AI-powered robots are used in scientific research and exploration, such as space exploration, deep-sea exploration, and laboratory automation. These robots can perform tasks in challenging and hazardous environments, gather data, and assist researchers in their work.
Scientific Image Analysis: AI algorithms are employed to analyze and interpret scientific images, including microscopy images, satellite images, and astronomical images. AI can aid in identifying objects, patterns, and anomalies in these images, enhancing scientific understanding.
Simulation and Modeling: AI is utilized to simulate complex systems and phenomena that are challenging to study experimentally. AI-driven simulations enable researchers to understand and predict the behavior of complex systems, such as protein folding, galaxy formation, and climate dynamics.
Scientific Literature and Knowledge Extraction: AI algorithms can process and analyze vast amounts of scientific literature, extracting relevant information, summarizing research articles, and identifying connections between different scientific studies. This helps researchers stay updated with the latest advancements and facilitates knowledge discovery.
Robot-Assisted Experiments: AI-powered robots can assist scientists in performing experiments, carrying out repetitive tasks, and collecting data. These robots can optimize experimental parameters, conduct high-throughput screenings, and enable faster and more efficient experimentation.
Data Visualization and Exploration: AI techniques, such as machine learning and data visualization, aid in exploring complex scientific data, identifying trends, and extracting meaningful insights. AI algorithms help researchers visualize and interpret data in a more intuitive and comprehensive manner.

These are just a few examples of how AI is used for scientific purposes. The integration of AI technologies continues to expand, enhancing scientific research and enabling breakthroughs across various disciplines.

Som du kan se af ChatGPTs svar, er der også et stort potentiale for at bruge AI i undervisning. Du skal dog være meget opmærksom på, at AI-sprogmodeller kan give ukorrekte svar. Læs mere i lektionen 'Algoritmerne styrer dit søgeresultat'.

______________________________________________________

Refleksionsøvelse

Hvor og hvordan giver det mening at bruge AI inden for dit fagområde?

____________________________________________________________

Hvilke metoder bruger man til at udvikle AI?

Udviklingen inden for AI går meget hurtigt og følger mange spor. Derfor er der også mange forskellige måder at forklare de grundlæggende tilgange og metoder på – og du kan støde på mange betegnelser, som er relaterede, overlappende samt over- eller underordnede i forhold til hinanden. Samtidig bliver forskere og udviklere bedre og bedre til at kombinere metoder, som tidligere blev udviklet hver for sig, så de opnår programmer med stadig mere avanceret 'intelligens'.

Når vi spørger ChatGPT om de vigtigste metoder inden for AI, får vi endnu en liste med 10 (!) punkter:

What are the most important methods used within AI?

ChatGPT - answer icon.png ChatGPT (2023, May 12 Version ) svarede:

Within AI, several important methods and techniques are commonly used. Here are some of the key methods used within AI:

Machine Learning (ML): Machine learning is a core method in AI that enables systems to learn and improve from data without explicit programming. ML algorithms can be broadly categorized into supervised learning, unsupervised learning, and reinforcement learning. Supervised learning involves training models on labeled data to make predictions or classify new instances. Unsupervised learning focuses on finding patterns and structures in unlabeled data. Reinforcement learning involves learning optimal actions through interaction with an environment and receiving rewards or penalties.
Deep Learning: Deep learning is a subset of machine learning that focuses on artificial neural networks with multiple layers. Deep learning models, known as deep neural networks, can automatically learn hierarchical representations from raw data. They excel at tasks such as image and speech recognition, natural language processing, and complex pattern recognition.
Natural Language Processing (NLP): NLP involves the ability of machines to understand, interpret, and generate human language. Techniques within NLP include text classification, sentiment analysis, named entity recognition, machine translation, question answering, and language generation. NLP often utilizes methods like word embeddings, recurrent neural networks (RNNs), and transformer models.
Computer Vision: Computer vision focuses on teaching machines to understand and interpret visual information. It involves techniques such as image classification, object detection, image segmentation, facial recognition, and image generation. Deep learning approaches, especially convolutional neural networks (CNNs), have revolutionized computer vision tasks.
Reinforcement Learning (RL): Reinforcement learning is a learning paradigm where an agent learns optimal actions through interaction with an environment. The agent receives feedback in the form of rewards or penalties based on its actions. RL algorithms, such as Q-learning and policy gradients, aim to maximize long-term rewards by learning an optimal policy.
Probabilistic Graphical Models: Probabilistic graphical models provide a framework for representing and reasoning about uncertainty in AI. They combine probability theory with graph theory and enable modeling complex relationships and dependencies between variables. Examples include Bayesian networks and Markov random fields.
Evolutionary Algorithms: Evolutionary algorithms are inspired by biological evolution and natural selection. They involve generating a population of solutions and applying operators like mutation, crossover, and selection to evolve and improve the solutions over generations. Evolutionary algorithms are often used in optimization problems and search spaces with a large number of possible solutions.
Knowledge Representation and Reasoning: Knowledge representation and reasoning (KRR) focuses on representing and utilizing knowledge in AI systems. It involves capturing knowledge in a structured form, such as ontologies, semantic networks, or rule-based systems, and using logical or probabilistic reasoning methods to draw inferences and make decisions.
Transfer Learning: Transfer learning is a technique that allows models to leverage knowledge learned from one domain or task to improve performance in a different but related domain or task. Pretrained models, such as those trained on large-scale datasets like ImageNet, are often used as a starting point for transfer learning.
Ensemble Learning: Ensemble learning combines multiple models, typically using different algorithms or trained on different subsets of data, to make predictions. By aggregating the predictions of diverse models, ensemble methods can often improve overall performance and generalization.

These methods represent a subset of the techniques used within AI, and there are numerous other specialized algorithms and approaches depending on the specific problem domain and application.

Machine learning (ML) er en helt central tilgang til AI. Du kan lære mere om ML og de underliggende metoder 'supervised learning', unsupervised learning' og 'reinforcement learning' i lektionen 'Hvordan kan vi bruge Machine Learning?'

____________________________________________________________

Almindelige betegnelser for forskellige typer AI

ChatGPT's liste ovenfor mangler nogle ret almindelige betegnelser for forskellige typer af AI, som du også kan støde på:

Mønsterbaseret vs. regelbaseret AI

Se en kort video, hvor professor Thomas Bolander fra DTU Compute fortæller om de to tilgange:

Narrow/Weak AI vs. General/Strong AI

Artificial Narrow Intelligence (også kaldet Weak AI) er AI-systemer, der er udviklet og trænet til at løse specifikke og afgrænsede opgaver som f.eks. at spille skak, analysere billeder eller generere tekstsvar som ChatGPT.

Artificial General Intelligence (også kaldet Strong AI) er idéen om at udvikle AI-systemer, der kan løse enhver opgave og lære nye kompetencer. Der findes endnu ikke AGI-systemer, men mange forskere og udviklere arbejder hårdt på at lykkes med missionen.

Deterministisk vs. ikke-deterministisk AI

AI kan være deterministisk, dvs. at samme input giver samme output hver gang.

AI kan også være ikke-deterministisk, dvs. at der er tilfældig variation i output for det samme input.

Mennesker har svært ved at være deterministiske. Der er næsten altid variation i vores beslutninger, og den variation, der er, er sjældent uden bias. AI kan i princippet være uden bias, men når vi udvikler systemerne til at efterligne mennesker, kommer de nemt til at reproducere den menneskelige bias.

Generativ vs. diskriminativ AI

Generative modeller kan bruges til at generere nye data, f.eks. når vi beder ChatGPT om at svare på spørgsmål om AI, eller når vi bruger Midjourney til at kreere billeder ud fra vores tekstinput.

Billede skabt af Midjourney version 5.1, 22. maj 2023. Tekstinput: Photo of students in university auditorium with trees growing, scifi

Generative modeller er ofte ikke-deterministiske, dvs. at der er tilfældig variation i det generede output, selvom input er det samme. Variation er med til at gøre generative modeller interessante. ChatGPT havde næppe skabt samme hype, hvis den gav præcist samme svar til det samme spørgsmål hver gang.

Diskriminative modeller kan bruges til at skelne mellem data eller træffe beslutninger, f.eks. diagnosticere lungebetændelse i røntgenbilleder. Diskriminative modeller er ofte deterministiske, dvs. at det samme input giver det samme output hver gang. Deterministiske modeller kan være nemmere at stole på, fordi der ikke er variation i output. Det er dog ikke en garanti for, at output er korrekt.

Den samme model kan være generativ eller diskriminativ alt efter, hvordan den bliver brugt, og nogle modeller er både generative og diskriminative. Men konceptuelt kan det hjælpe at tænke på det som to forskellige klasser af modeller.

____________________________________________________________

Hvordan opstår bias i AI-systemer?

AI-systemer bliver ofte kritiseret for, at de reproducerer bias i de data, de er trænet på, eller hos de udviklere, der har designet algoritmerne. Samtidig er AI-systemerne ofte så store, at det er vanskeligt at gennemskue, hvor og hvordan bias opstår. ChatGPT giver os denne forklaring:

Why are some AI systems biased?

ChatGPT - answer icon.png ChatGPT (2023, May 12 Version ) svarede:
AI solutions can be biased due to various reasons, including:

Biased training data: AI systems learn from data, and if the training data itself contains biases or reflects societal prejudices, the AI system can unintentionally learn and perpetuate those biases. For example, if historical data used to train a hiring AI system has biased outcomes based on gender or ethnicity, the system may inadvertently learn and replicate those biases in its decision-making.
Algorithmic design: Biases can also be introduced through the design and implementation of algorithms. The algorithms may have inherent biases or unintended consequences due to the choices made by developers, such as the selection of certain features or the weighting of different factors.
Lack of diversity in development teams: Bias can arise from the perspectives, experiences, and backgrounds of the development teams. If the teams lack diversity, they may not be fully aware of or sensitive to potential biases in the AI system, leading to biased outcomes.
Inadequate evaluation and testing: Insufficient testing and evaluation of AI systems can fail to uncover biases. If biases are not properly identified and addressed during the development and testing stages, they can persist in the deployed AI system.

Addressing bias in AI solutions requires a combination of careful data selection and preprocessing, algorithmic fairness considerations, diversity in development teams, rigorous testing and evaluation, and ongoing monitoring and feedback loops. It is crucial to have robust ethical frameworks and guidelines in place to mitigate biases and promote fairness and accountability in AI development and deployment.

____________________________________________________________

Refleksionsøvelse om bias

Tag udgangspunkt i ChatGPTs eksempler på kilder til bias. Overvej for hver kilde, om det er muligt at eliminere den kilde til bias fuldstændigt.
Når vi taler om bias. er det generelt med udgangspunkt i kategorier, der dækker store grupper af mennesker, f.eks køn, alder, nationalitet. En af visionerne for AI er personaliseret medicin, hvor behandling bliver målrettet hvert enkelt individ. Overvej, hvad bias betyder i denne sammenhæng.

____________________________________________________________

Hvis du vil vide mere

Her finder du en række ressourcer, hvor du på egen hånd kan lære mere om AI og machine learning:

Data literacy – What is it and why does it matter?
Elements of AI, der introducerer dig til AI

____________________________________________________________