When most people think about AI or Machine Learning, they probably think of AI models. These could include generative models like Chat-GPT, object detection models such as YOLO or Mask-RCNN that detect human faces or specific objects in images, predictive models for anomaly detection and forecasting, or clustering models that classify customers. Many people might assume that AI developers or Data scientists primarily focus on creating these models. Of course, to build AI models, you need a complex and sophisticated model structure, but you also need an indispensable resource — big data.

According to [How Much Data Is Generated Every Day in 2024?], a staggering 3.5 quintillion bytes of data are created every day. For instance, X (formerly Twitter) generates 456,000 tweets every minute, and 5 million posts are uploaded to Instagram. This enormous amount of data can be utilized to create AI models. However, you can’t directly use the raw data to train the models. Since raw data often contains a lot of garbage data, filtering (preprocessing) is crucial.

Let me give you an example. Imagine you are developing a classification model to categorize different types of snacks. You choose EfficientNet, a well-known classification model, and collect 10,000 snack images. These 10,000 images represent 250 types of snacks, with 40 images for each type. After selecting the model and collecting the dataset, what’s next? You can’t train the model using the raw data that you just collected. Each image needs to be labeled with the corresponding product name, and this labeling process is often done manually. The person responsible for this task is sometimes referred to as an “oracle.”

Now, imagine someone sitting in a chair, labeling over 10,000 images one by one. Think about the amount of time and energy this process requires. Furthermore, not all the collected data is intact; there may be garbage data, such as out-of-focus images or images where the products are cut off. In this situation, you might wonder: what if you could prioritize labeling the most important images first? What if you could filter out the garbage data and select only the images that significantly influence the model? By labeling these important images first, you could quickly build a prototype.

Press enter or click to view image in full size

Anyone who has done data labeling might have thought about this at some point. There is a solution to this problem: Active Learning. Active Learning is a technique that can effectively reduce the time required for data labeling. As mentioned earlier, it involves selecting data that is likely to have a significant impact on model training. Depending on the data format, Active Learning can be categorized into Pool-based and Stream-based scenarios. There are also various strategies for selecting the data, such as Uncertainty Sampling, Query-By-Committee, and Expected Model Change.

Today, I introduced the concept of Active Learning. In the next discussion, I will delve deeper into the scenarios and strategies involved.