A Typology of AI

A Typology of AI

Al is a suite of different technologies - much more than just the Large Language Models (LLMs) that have captured the lion's share of attention in the past year. Different kinds of Al can support specific tasks for individuals and organizations to complete at scale.

The following is a taxonomy of Al, broken down according to the type of task completed. These technologies are available today and are used by organizations in the private and public sectors at different scales. In assessing the potential impact of Al on the operating model of government, we focused on how these capabilities match the typical functions of a government department and the improvements they can bring.

NARROW Al

Narrow AI refers to AI systems that are designed to operate within a predefined and constrained domain, performing specific tasks with a high degree of expertise. While generative AI creates new content or data that did not previously exist, narrow AI is focused on achieving particular goals and solving specific problems using a fixed set of guidelines.

STATISTICAL Al

• Classification: assigning categories to data points.

• Clustering: grouping similar data points together, for example through photo tagging.

• Anomaly detection: identifying unusual data points.

NATURAL LANGUAGE PROCESSING (NLP)

• Sentiment analysis: determining the sentiment expressed in a piece of text.

• Machine translation: translating text from one language to another.

• Part-of-speech tagging: identifying words as nouns, verbs, adjectives and so on.

• Text summarization: generating a concise and coherent summary of a larger text.

• Language modelling: predicting the next word in a sentence.

COMPUTER VISION

• Image classification: identifying the main subject of an image.

• Object detection: locating objects within an image and identifying them.

• Face recognition: identifying or verifying a person's identity using their face.

• Image segmentation: dividing an image into parts to be analyzed separately.

• Motion detection: identifying movements within a video sequence.

RECOMMENDATION SYSTEMS

• Content-based filtering: recommending items similar to those a user has liked before, based on item features.

• Collaborative filtering: making recommendations based on the preferences of similar users.

• Personalized recommendations: tailoring suggestions to individual user profiles.

SPEECH RECOGNITION

• Automatic speech recognition (ASR): converting spoken language into text.

• Voice-command recognition: understanding and executing spoken commands.

• Speaker identification: determining who is speaking.

• Speech-to-text transcription: transcribing audio content into written text.

• Voice-activity detection: detecting when someone is speaking in audio data.

TIME-SERIES SYSTEMS

• Identifying trends: detecting long-term increases or decreases in data, such as rising sales trends or declining product demand.

• Understanding abnormal fluctuations: spotting cycles not tied to a fixed calendar schedule, such as economic expansions and recessions.

• Spotting outliers: identifying unusual data points that deviate significantly from the norm, which could indicate errors, extraordinary events or opportunities for further investigation.

• Predicting future values: estimating future data points, like stock prices or weather conditions, based on the identified patterns and relationships in the time-series data.

• Simulating new scenarios: predicting the likely changes over time of key variables, based on various parameters.

GENERATIVE Al

Generative AI refers to AI applications that can generate new content, data or information (text, images, audio, video, music, 3d objects, architectural and engineering drawings/blueprints, computer programs etc.) that is similar to human-created content. Unlike discriminative models that classify input data into categories, generative models can produce data that is not present in the original training set. They are general purpose in nature, and outputs from the same inputs are not always exactly the same.

TEXT-TO-IMAGE/VIDEO

• Image synthesis: generating images that visually represent the content described in text inputs. This includes creating artworks, product designs and realistic scenes based on descriptive text.

• Video synthesis: generating short videos or animations based on a text description. This involves combining elements of text-to-image generation with motion and transition models to create dynamic scenes.

TEXT-TO-AUDIO

• Speech synthesis (text-to-speech): converting written text into spoken words. This is widely used for creating voiceovers, reading text aloud in accessibility tools and virtual assistants.

• Speech recognition: transcribing spoken language into text. While the input is audio rather than text, this technology is crucial for enabling further generative tasks, such as translating spoken words or generating text-based responses.

• Music generation: composing music based on textual descriptions of mood, genre or specific musical elements.

TEXT-TO-TEXT

• Content creation: writing articles, stories or poetry based on prompts or outlines provided in text form.

• Translation: translating text from one language to another while maintaining the original meaning and context.

• Paraphrasing: rewriting text to alter its form without changing its original meaning.

TEXT-TO-3D MODELS

3D-model generation: creating 3D models from textual descriptions. This can be used in game development, architecture and product design to visualize objects and environments described in text.

TEXT-TO-CODE

Code generation: producing executable code from natural language descriptions. This aids in software development by allowing developers to describe functionalities in plain English and automatically generate code snippets.

SYNTHETIC DATA

Synthetic-data production: creating artificial data that can be used in place of real data for various purposes. Examples include creating new health data to protect privacy and simulating real-world environments for driverless cars.

Al SAFETY

Each of these models is designed to address specific ethical and practical challenges in Al, such as understanding Al decisions, ensuring consistent performance, protecting user privacy and treating all users fairly.

EXPLAINABILITY

Interpreting Al decisions: using dynamic sampling to provide insights into why an AI model made a certain prediction, helping users understand the decision-making process.

ROBUSTNESS

• Detecting unusual patterns: employing anomaly-score technology to identify out-of-the-ordinary data points or shifts in data patterns, useful for spotting potential fraud or errors in data.

• Assessing prediction trustworthiness: implementing credibility scores to evaluate how much confidence users should have in the predictions made by an Al model, based on its past performance with similar data.

PRIVACY

• Generating safe-to-use data: creating synthetic data that mimics real-world data while preserving individual privacy, ensuring that sensitive information remains confidential when the data is used for testing or development purposes.

• Sharing data privately: developing algorithms that allow for sharing data in a manner that upholds individual privacy, enabling collaborative use of data without compromising personal information.

FAIRNESS

• Correcting biases in Al: applying fairness-correction algorithms to adjust Al models and mitigate biases, ensuring that decisions are equitable without the need for extensive retraining.

• Ensuring equitable Al: using model-agnostic methods to analyses and improve fairness throughout the life cycle of Al models, aiming for unbiased and fair outcomes in Al predictions and decisions.