第1章コンピューターサイエンス入門

Do you agree with the social determination of technology? Is it possible for technology—before it is used by people to solve certain problems—to encode social values? Try to come up with an example that would support this belief. What about an example that refutes this belief? What are the social implications of agreeing or disagreeing with the social determination of technology?

Today’s neural networks are designed to identify patterns and reproduce existing data. It is widely accepted that many big datasets can encode social preferences and values, particularly when the data is collected from users on the Internet. A social determination of technology accepts this explanation of AI bias and leaves the design of AI algorithms and techniques as neutral: the bias in an AI system is attributed to the social values of the data rather than the design of the AI algorithms. Critics of social determination point out that the way AI algorithms learn from big data represents a social value, one that encodes a default preference for reproducing the biases inherent in big data. This applies whether the AI application is about fair housing, medical imaging, ad targeting, drone strikes, or another topic. This is an issue that computer scientists must consider as they practice responsible computing and strive to ensure that data is gathered and handled as ethically as possible.

Today’s AI technologies work by reproducing existing patterns rather than imagining radically different futures. As much as neural networks are inspired by the human brain, it would be a stretch to suggest that AI systems have any semblance of general intelligence. Though these systems might be quite effective at identifying lettuce plants from weed plants in an image, their capacity for humanlike intelligence is limited by design. A neural network learns to recognize similar patterns that appear across millions or billions of sample images and represent these patterns with millions or billions of numbers. Mathematical optimization methods are used to choose the numeric values that best encode correlations across the sample images. However, current approaches lack a deeper, conceptual representation of objects. One criticism of very large neural networks is that there are often more numeric values than there are sample images—the network can effectively memorize the details of a million sample images by encoding them in a billion numbers. Many of today’s neural networks recognize objects in images not by relying on some intrinsic idea or concept of objects but by memorizing every single configuration of edges as they appear in the sample images.

This limitation can lead to peculiar outcomes for image recognition systems. Often, neural network approaches for image recognition have certain examples of images where objects are misidentified in unusual ways: a person’s face might be recognized in a piece of toast or in a bunch of clouds in the sky. In these examples, the pattern of edges might coincidentally trigger the neural network values so that it misidentifies objects. These are among the more human-understandable examples; there are many other odd situations that are less explainable. An adversarial attack is a sample input (e.g., an image) that is designed to cause a system to behave problematically. Researchers have found that even tweaking the color of just a single point in an image can cause a chain reaction in the neural network, leading it to severely misidentify objects. The adversary can choose the color of the point in such a way as to almost entirely control the output of some neural networks: changing a single specific point in an image of a dog might cause the system to recognize the object as a car, airplane, human, or almost anything that the adversary so desires. Moreover, these adversarial attacks can often be engineered to cause the neural network to report extremely high confidence in its wrong answers. Self-driving cars that use neural networks for image recognition can be at risk of real-world adversarial attacks when specially designed stickers are placed on signs that cause the system to recognize a red light as a green light (Figure 1.10). By studying adversarial attacks, researchers can design neural networks that are more robust and resilient to these attacks.

In general, research is an important part of computer science. Through research, computer scientists analyze ways that technology can be used and gain insight and answers to address issues and improve various aspects of society. Research enables computer scientists to make advancements like the design of new algorithms, development of new hardware and software, and applications for emerging technologies such as AI.

One important use of research is to investigate adversarial attacks to gather answers needed for computer scientists to improve foundational technologies by evaluating the negative consequences of technologies. Computer technologies offer a unique medium for learning things (not just learning computer science), connecting with each other, and enhancing the lives of people all around the world. Yet, in each of these examples, we also raised concerns about how these technologies unfolded and affected people’s lives in both positive and negative ways. While we can rarely, if ever, paint any one technology as purely “good” or “bad,” computer scientists are interested in studying questions around how technologies are designed to center social values. Social scientists are not solely responsible for answering questions about technology, but computer scientists can also contribute important knowledge and methods toward understanding computer technologies.