Summary
Highlights
To participate, learners need a device with a webcam (laptop, tablet, or phone) to use NearPocket, an application for designing image classifiers. NearPocket can be accessed online or via Windows and Android apps; iPhone users need to use the website. An optional project worksheet is provided to guide the AI classifier design process, though a plain sheet of paper works too. The course features symbols: a book for teaching, a running person for interactive activities, and a cloud for reflection.
The course, developed by Henry Lee and learnaianywhere.org, aims to make AI literacy accessible globally. It distinguishes AI as a product of human choices, not a human-versus-machine conflict. The primary goal is to equip learners with the knowledge and judgment to be responsible creators and critical users of AI tools. Key learning objectives include understanding AI, machine learning, its mechanisms, and responsible AI principles, all demonstrated by designing an AI classifier with NearPocket.
Lesson 1 defines AI by focusing on two key qualities: autonomy (acting and making decisions independently) and adaptivity (learning from experience and improving). AI aims to imitate these intelligent behaviors in machines. Examples like Google's search engine demonstrate AI's autonomous suggestions and adaptive learning from user data, while a calculator is presented as non-AI due to its lack of independent operation and learning. The lesson emphasizes that AI is a tool, and humans are responsible for its use.
Learners are tasked with designing their own AI classifier on a chosen theme, similar to the Plant Village Nuru tool used by farmers. Examples of student projects include classifying plant leaves, different materials, or even sign language gestures. The project involves completing sections of a worksheet after each lesson, either individually or in groups, culminating in a presentation of the AI classifier and worksheet responses.
Lesson 2 explains the core components of machine learning: model, data, training, and trained model. Human learning, involving collecting notes (data), studying (training), and having a prepared brain (trained model), is analogous to machine learning. A model captures patterns, data is the input for training, training is the process of finding patterns, and a trained model is the result, ready for predictions. NearPocket is introduced as the tool to experience these parts hands-on.
Learners are guided step-by-step through using NearPocket to create an image classifier. The example involves classifying 'scissors' and 'pens' by creating groups, adding photos (data) to each, and then training the model. The exercise demonstrates how the AI classifies new images. It highlights that AI tools can produce unexpected results because they analyze every detail without human-like understanding or intuition, relying solely on analyzed patterns.
AI systems are heavily limited by their training data. An example shows a plant disease classifier failing to identify grass because it was never trained on grass, misidentifying it as a 'healthy peach.' Similarly, language AI only responds in languages it was trained on. This illustrates the principle: 'what comes in, comes out.' Limited data leads to limited and potentially inaccurate predictions. Learners are then prompted to collect data for their own AI classifier project, ensuring they consider various items for classification.
Lesson 3 delves into how machine learning training works, focusing on algorithms. An algorithm is defined as a clear set of step-by-step instructions to complete a task, applicable to everyday activities and machines. The 'Describer-Drawer' game is used to demonstrate how clear verbal instructions (algorithms) are crucial for tasks. This highlights that unlike humans, machines require explicit, detailed instructions because they don't 'understand' intuitively.
The 'Which group does it belong to?' game illustrates human learning as an iterative process of guessing, checking answers, and adjusting focus based on feelings and curiosity. Humans naturally seek meaning and connect ideas. Machines, however, don't have feelings; they make calculated predictions. They convert features into numerical data, multiply them by importance values, and adjust these values based on error measurements to refine their mathematical equations. This repetitive process eventually yields a trained model capable of accurate predictions.
A trained AI model is essentially a complex mathematical equation. Every detail in the input data is converted into numbers for calculation, emphasizing why data quality and variety are crucial. AI's adaptiveness and autonomy stem from human-written learning algorithms and human-provided data and feedback. The environmental cost of AI, running on vast data centers, is also noted. Ultimately, AI is presented as 'math guided by human choices,' underscoring human responsibility in its development and impact.
Learners are instructed to identify and record the key features (e.g., color, shape, size) they will use to distinguish between the groups in their AI classifier, building upon the data collected in Lesson 2. They are encouraged to deepen their thinking about these features and continue to improve their AI classifier by collecting more data if allowed.
Lesson 4 explores the ethical implications of AI, specifically data bias and privacy, and emphasizes human responsibility. The 'Barely a Dog Challenge' with NearPocket demonstrates how human bias in data collection (e.g., only using black bears and white dogs) leads to algorithmic bias and inaccurate predictions for varied inputs. Bias is a natural human trait, but it becomes problematic when imposed on others without considering their experiences, potentially leading to harmful outcomes like inaccurate medical diagnoses, wrongful arrests, or misinterpretations in language.
AI tools are ultimately just tools, devoid of emotions or responsibility; humans are accountable for their actions. To reduce harm, AI creators must aim for diverse, high-quality data, listen to different perspectives, and be transparent about tool limitations. Users should double-check AI results, especially when risks are involved, and provide feedback. The lesson encourages thinking about how to improve the 'bear and dog' classifier by considering diverse features like age, environment, and various breeds to reduce bias.
The concept of data privacy is introduced, distinguishing between public data (accessible with usage rules) and private data (personally identifiable information requiring permission). Examples of private data include faces, contact information, financial details, and medical records. Learners are prompted to consider their feelings if personal photos were used without permission, stressing the right to control one's data. Cases of creators misusing private data and facing consequences are discussed, highlighting the importance of clear consent and data removal for unnecessary information.
To protect privacy, creators must ask for permission, be clear about data usage, and remove unneeded personal information. Users should read data policies, ask questions, and exercise their right to limit data sharing. An example demonstrates how to ensure privacy when adding a photo to an AI classifier by asking permission or blurring/cropping out private details like faces, addresses, or car license plates. The lesson concludes by defining 'responsible AI' as making sure people are responsible creators and critical users, recognizing that everyone plays a role in shaping AI, which is a product of human choices.
For the final lesson, learners are to expand their perspective on features, adding variety, exploring new characteristics, or seeking feedback to reduce bias in their AI classifier. They must explain their classifier's bias and how they plan to reduce it. The importance of asking permission when using private data is reiterated. The example response for this question details taking diverse fruit pictures and seeking peer feedback, while also committing to asking permission for any private information used.