As in most systems, AI is only as strong as its weakest link. If you have good data but no way to process it you have nothing, and likewise even if you have the most advanced architecture but you feed it low quality data the results will be underwhelming.
In this chapter we will discuss the importance of providing high quality data when training an AI model in addition to evaluating what criteria data needs to meet to be considered high quality.
There are systems that use simulated learning where an AI interacts with a simulated environment and learns by trial and error through reinforcement (this will be explained in chapter 4). For other AI models training data is the only source of understanding for the AI. All of the patterns observed by the model are based on the data it is being fed. This is because the AI, unlike humans, has no body or personal experience to draw knowledge from and is reliant on the input to define its truth. It has no further way to test its conclusions unless it is later told that the information is incorrect. This is why high quality data is necessary for AI models to function. That being said, what constitutes high quality data?
The first consideration is accuracy. While there are both labeled and unlabeled sets of data, accuracy is critical to both. When using labeled data in combination with supervised learning (more on this in chapter 4) the accuracy of the data relates to how detailed and correct the labels are relative to the data. For example you can train an AI to identify objects in an image.
Example input image https://www.rspb.org.uk/birds-and-wildlife/robin
If we want to give an example for accuracy in the context of unlabeled data used in unsupervised training, we can imagine a large language model (LLM). LLMs such as ChatGPT can be trained on unlabeled data such as articles, tweets, reddit posts, science journals, etc. This is unlabeled data and the AI uses it to identify patterns in the sentence structure and forms relationships between words. By training it on enough of this data the AI starts to produce sentences that a human might write. If, however, you train it on school homework written by children of age 6, you may get undesired outputs from the AI. This is what quality means in the context of unlabeled data.
Shortly put, if you teach it things that aren’t true the AI has no way to correct this. Make sure the data being used to train an AI is high quality or it might identify cats as dogs. If you are training it on text, make sure it is written in coherent and grammatically correct text otherwise the outputs may be nonsensical. Even if a small subset of the data is low quality it may contaminate the model since AI works on patterns and it will make incorrect connections. The quality of the input data is directly correlated to the quality of the outputs.
Another consideration for high quality data is built in biases. Even if there is technically nothing wrong with the data you can still have massive problems if you overlook biases. In one PR scandal an image recognition AI seems to have been trained mostly on caucasian human faces which lead to it identifying people from other races as primates. This can be solved by taking into consideration the 2 next criteria for high quality data – volume and diversity.
The larger the brain of the AI – measured in number of parameters – the more data is needed to accurately and effectively encode values into its nodes since each input will only slightly modify the node values. Think of nodes as mathematical functions – they take some input and give some output after doing some math. For this reason, it is important to have sufficient volume when training an AI. There are various suggestions online for how much data is enough, but the only real way to tell is by testing the AI. This of course tests the whole system and not only the amount of input data so it is difficult to identify the weak point, but by testing the model you will get a feel for the results you can expect.
The final important consideration for training data is the diversity. If you give the AI too much of one type of input it will overfit and not recognize general patterns, but only be optimised for that one question. Of course this depends on the purpose of the AI but for large language models it is critical that they have a broad understanding and try to reason on their own. Let’s see as some examples: if you only show it adding questions then it will not know how to multiply. If you show it too many articles then it will sound like a reporter in all contexts so you may want to include reddit threads or tweets. In an image recognition AI you may feed it too many similar photos – if all photos with a bird have a blue sky background then it may assume that the blue is contributing to the image being a bird.