Overview of AI Safety
1. Overview of AI Safety
Welcome to AI safety. This module consists of seven lessons. Today, you will learn to define safety for AI. Discover key considerations for safety. Explore techniques for AI safety, and describe which Google Cloud products can help with AI safety. Let's start with the overview of what AI safety means. AI safety involves many perspectives and considerations, and it has many difficulties, such as: Unknown action space. When machine learning is applied to problems that are difficult for humans to solve, it becomes challenging to predict all scenarios ahead of time, and especially so in the era of generative AI. Performance and safety tradeoffs. It is difficult to build systems that provide both the necessary proactive restrictions for safety, as well as the flexibility needed to generate creative solutions or adapt to unusual inputs. And attackers adapt fast to new technology. As AI technology continues to evolve, attackers will evolve too and will surely find new means to attack, resulting in new solutions that will need to be developed in tandem. AI safety is especially difficult when it comes to generative AI with classic discriminative AI models, such as those trained for classification and regression tasks. The output space is fairly well known. Classification models can only predict the classes they were trained on, so we can simply exclude classes we deem to be harmful. On the other hand, regression models can only predict a number and we know what that number is supposed to represent. That means when we build the models, we have a good idea of what the model can at least predict. This doesn't mean that the models are entirely safe and can't cause harm, but it does make it easier to find problems. However, with generative AI, we're creating a system that allows users to explore the breadth and depth of their creativity. This means that the models are trained on large amounts of data, and they are designed to learn to generate outputs that are very different from the data they were trained on. The creativity of the user combined with the emergent creativity of the model means that it is difficult to know the breadth and depth of a model's responses in advance. To approach safety, there are roughly two different approaches which are complimentary to each other. First, is a technical approach that seeks technical solutions through changes to the model or the system we build or through engineering practices. On the other hand, a non-technical or institutional approach is also very important as it seeks formal and informal institutional solutions to AI safety at the lab. Usually, this involves industry-wide, national and international scale discussions and commitments. This approach is also known as a topic of AI governance. And while this is a technical course for developers focusing on the technical aspects of AI safety, note that non-technical aspects often promote technical approaches, and they are complimentary to each other. How can we approach AI safety from a technical perspective, especially in generative AI? There are multiple measures we can take. This diagram shows an overview of the generative AI system. The system receives prompt inputs, passes it to the AI model, and returns outputs. Between the prompt input and the final output, you can see input safeguards and output safeguards. These safeguards take care of input-output filtering based on defined safety criteria. By using safeguards properly, we can harness the system input-output to avoid unintended harmful outcomes. In addition to safeguards. You can also try to intervene in the model training and tuning process in order to embed safety concepts into the model and align its behavior toward safety. It is also critical to understand how to evaluate AI systems throughout the entire process. Evaluation involves exhaustive testing to find an unintended behavior of the system. This is called adversarial testing. We'll look at each item as we progress through this course.2. Let's practice!
Create Your Free Account
or
By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.