In Fei-Fei Li's TED Talk, “How We Can Teach Computers to Understand Pictures,” she delves into cutting-edge advances in computer vision, specifically how the team she led trained computers to understand the content of images by constructing a database containing 15 million photos.
She noted that while computers have made progress on perceptual tasks (e.g., image categorization), they still perform poorly on cognitive tasks (e.g., image description and question and answer).
To that end, her team created a dataset of more than 100,000 images, each containing an average of 21 objects, 18 attributes, and 18 pairs of relationships between objects, to help the model understand the interactions and relationships in the images.
Her research has undoubtedly advanced AI, but from a critical perspective, we should also reflect on its limitations and potential problems.
1. Data limitations and biases
Her research relies on large-scale datasets such as ImageNet, but data is not always neutral. The source of the dataset, the way it is labeled, and even the biases that may have been present during the collection process can all affect the computer's “understanding” of the world. For example, some datasets may lack diversity, leading to poor performance of the AI in a particular population or cultural context.
Are we inadvertently allowing AI to inherit the biases of human society?
2. Understanding vs. recognizing
The title of the talk was “How to Teach Computers to Understand Pictures,” but in reality, computers are more about recognition than understanding; AI can recognize and classify objects in an image, but it doesn't really “understand” what those objects mean.
In the development of AI, the ability to recognize should not simply be equated with the ability to understand.
3. The relationship between humans and AI: replacement or collaboration?
Fifi Lee's research is undoubtedly a technological breakthrough, but does it mean that humans should let AI replace certain decision-making processes? Advances in computer vision have allowed machines to automate more and more tasks, such as autonomous driving and medical image analysis. However, if AI takes over these tasks completely, will we lose control of our lives as a result? Will humans become overly dependent on AI and weaken their own cognitive abilities?
I did a game of rock-paper-scissors with hand signals.