Despite the seemingly constant presence of AI in news headlines, the range of visual resources accessible to help designers, commentators, reporters, and researchers communicate about AI remains limited. Practitioners complain about the use of frightening images to represent AI in the media, and visual representations including those in iconography often rely on metaphors like robots and lightning bolts to communicate ideas about AI.
The limited range of ways to convey key concepts in AI and machine learning contributes to the sensationalism of this critical topic and does little to demystify AI, or help audiences understand how it works and its potential impact on our lives.
So, in late-Summer 2019, Noun Project partnered with Essence, a data science and measurement-driven agency, for an Iconathon all about Artificial Intelligence. This Iconathon was the first step in a collaborative design process intended to provide the public with a greater — and more inclusive — range of ways to represent AI and machine learning.
More than 30 designers, engaged citizens and subject matter experts gathered at Essence’s office in San Francisco and received briefings from subject matter experts — including representatives from the Center for Human Compatible AI — on the state of AI, the need to treat AI as an interdisciplinary and inclusive field, and the need for a broader, more inclusive set of visual references. Participants then sketched ideas for new icons to communicate topics as varied and complex as AI training data, Natural Language Processing, AI Ethics, and AI bias.
Participant sketches were then used as inspiration for the final collection of 20 icons, detailed below.
“As the leading resource for visual language, it’s important that we not only provide the visual representations that are currently available, but also lead the way for creating new visual definitions for cutting edge concepts and technologies, like AI, that are not yet clearly defined. Iconathons enable us to engage communities in the process of building this omnipresent language, so that we can use it to cross language and cultural barriers, and simplify communication.” — Sofya Polyakov, Noun Project co-founder and CEO
Each icon in the AI Icon set has been designed to communicate the behavior or concepts associated with each topic clearly, with limited reliance on metaphors.
For example, while an icon depicting a robot reading a book could metaphorically communicate “Machine Learning,” it does little to visually explain what the process of “Machine Learning” actually is. In the AI Icon collection, the Machine Learning icon depicts data being taken in by a computer or machine and then output into new data based on what the machine “learned” from the input. The collection has been designed around a consistent visual system, including the recurring use of a microchip element to represent computers or machines.
These icons have been designed to be more technical visually, ensuring they can be used effectively across a range of materials, from presentations to marketing collateral, to communicate these concepts.
The AI Icon Collection and definitions used in the design process:
1. Artificial Intelligence
Definition: (AI) The science of making things “smart” (act like humans). A non-human program or model that can solve sophisticated tasks, such as a program that identifies diseases from radiologic images.
2. Machine Learning
Definition: The science of getting computers to do something without being programmed with explicit rules; a sub-field of AI. Software that makes useful predictions of never-before-seen data based on what it has “learned” from an existing dataset.
3. Deep Learning
Definition: Branch of Machine Learning utilizing algorithms inspired by the multi-layered structure of neurons in the brain.
Definition: All the data that is used for building or testing the Machine Learning model. Sourced from a public resource or specifically collected.
5. AI Training Data
Definition: A dataset that a Machine Learning model uses to detect patterns and determine which aspects are most important during prediction.
Definition: A physical mechanical device that automatically interacts with its environment by sensing, planning and acting.
7. Natural Language Processing
Definition: A common notion for a variety of Machine Learning methods that make it possible for the computer to understand and perform operations using human (i.e. natural) language as it is spoken or written. Such as how Siri or Alexa understand what you mean.
8. Speech Recognition
Definition: Used for determining the text representation of people speaking. Such as Siri or Alexa knowing which words you said.
9. AI Ethics
Definition: A concern with the moral behavior of humans as they design, construct, use and treat Artificial Intelligence as well as concern for the moral behavior of the AI itself.
10. AI Value Alignment
Definition: Getting an AI system to adopt the goals of human users or stakeholders even if these are hard to express exactly.
11. Implicit Bias
Definition: Automatically making an association or assumption based on existing mental models. It can affect how data is collected and classified, as well as how machine learning systems are designed and developed.
12. Confirmation Bias
Definition: A form of implicit bias. The tendency to favor information in a way that confirms one’s preexisting beliefs. Machine Learning developers may inadvertently collect or label data in ways that influence an outcome supporting their existing beliefs.
13. Data Selection Bias
Definition: Errors in conclusions when they are drawn from data that has been treated as random, but it is not random. IE: Concluding something based on survey results when certain types of users opt-out of surveys.
14. Computer Vision
Definition: (CV) A field of Artificial Intelligence concerned with providing tools for analysis and high-level understanding of image and video data.
15. False Negative
Definition: When a model, while predicting classes, incorrectly predicted the negative class. Such as the model inferred an email was not spam (negative class), but it actually was spam.
16. False Positive
Definition: When a model, while predicting classes, incorrectly predicted the positive class. Such as the model inferred an email was spam (positive class), but it was actually not spam.
17. AI Over-Reliance
Definition: The tendency of humans to put too much trust in an AI and automation systems beyond their actual capabilities.
18. Unintended AI Effects
Definition: Effects caused by an AI system that were not preteen by its developers. IE: Social polarization through “smart” social media feeds.
19. Semi-Supervised AI Learning
Definition: When training AI using a data model where some examples are known with labels, and others are unknown without labels.
20. Unsupervised AI Learning
Definition: When training AI using a data model where all examples are unlabeled and AI needs to find the structure or relationships between the data.
Essence, part of GroupM, is a global data and measurement-driven agency whose mission is to make advertising more valuable to the world. Clients include Google, Flipkart, Nando’s and the Financial Times. The agency is more than 1,800 people strong, manages $4B in annualized media spend and deploys campaigns in 106 markets via offices in Bengaluru, Chicago, Delhi, Düsseldorf, Jakarta, London, Los Angeles, Melbourne, Minneapolis, Mumbai, New York, San Francisco, Seattle, Seoul, Singapore, Sydney, Tokyo and Toronto.
Visit essenceglobal.com for more information and follow us on Twitter at @essenceglobal
Want more content like this? Subscribe to our monthly newsletter, the Noun Gazette. 🎉
Senior Director of Marketing at Noun Project