-
Asked by anon-381472 on 23 Jan 2024.
-
Rosemary J Thomas answered on 23 Jan 2024:
– Data is like the food that AI eats to learn and grow. The more and better data we give to AI, the more it can understand and do things correctly.
– Testing is like the exercise that AI does to get stronger and smarter. The more we test AI with different situations, the more it can avoid errors and handle challenges.
– Explaining is like the teaching that AI gets to improve and adapt. The more we explain AI, how it works, the more it can be transparent and trustworthy.
– Rules and values are like the guides that AI follows to behave and cooperate. The more we follow some rules and values, the more we can make sure that AI is fair and ethical.
These are a few ways that we can make AI do what we ask it more accurately, or make AI less prone to being wrong.
-
-
Carl Peter Robinson answered on 24 Jan 2024:
“Data! Data! Data!” as quoted by Arthur Conan Doyle’s character, Sherlock Holmes. The more data you can acquire for the problem you are working on, the better. Not just quantity though, you need to get an accurate spread of the types of sample in your data, to ensure it is representative of a real-world scenario. So you need to make sure your dataset is balanced and isn’t biased.
You could also improve the evaluation phase of your process, where an AI model is trained and validated. We use things called hyperparameters to tweak a model during training and then observe what that does to its performance. Think of these hyperparamters as being a bit like the settings on your smartphone, in that when you first buy that phone you change the settings to get the phone to behave in the best way for you, e.g., screen brightness, volume, notification sounds, etc. So, in a kind of similar way, we want these model hyperparameters to be set such that the model gives us its best performance.
Additionally, you can provide feedback to the model, which is one of the ways ChatGPT has been trained, using something called Reinforcement Learning with Human Feedback. Here, humans provide feedback to the model by assessing the output it gave to a specific input by giving scores based on the quality of that output. This works in a similar way to a teacher marking the answers in your class work. The better answers you give, the better marks you get, the better rewarded you are (perhaps even with a gold star sticker!) It’s the same idea with the process where humans give feedback to the AI model. The better scores the model gets, the higher a reward it gets, so it knows the output it gave was good. That “encourages” it to produce the same/similar output the next time it receives that kind of input.
Comments