You only need one click to get the most informative items from your dataset.
Let me put the following scenario out there… You have a dataset that contains 1M images, and you have a deep learning supervised task that requires your data to be annotated in order to train it. The big question then becomes, do you send all 1M images for labeling? This is extremely time-consuming and not to mention very costly. And here’s another curveball, what if your budget can only handle 10% of the data? How do you choose which items to send for annotating?
A Little Help Understanding Your Unlabeled Dataset
In most cases, you don’t really need your whole dataset to be labeled. As a Data Scientist, you should have the ability to understand what your dataset consists of and how you can sample it wisely according to time and budget limitations. For these reasons, we developed a new tool- the Information Extraction Tool. The main goal of this tool is to find out what matters in a dataset and send it for annotation. Enabling an intelligent coverage of all of the variance in the dataset, according to the selected percentage, which is dependent on your time and budget.
For example, let’s say you have a dataset of several classes. One of the classes’ images are very similar to the other, for example – multiple images taken from the same angle, same object position, and the same background and lighting conditions. We don’t need to feed our model with all of these similar items, we could simply use a sample of it and send it for annotation.
Our tool aims to receive an unlabelled dataset and through a self-supervised process determine which items in the data are considered anomalies and which are the most informative. In the selection process, the tool makes sure to cover the dataset variance and extract only what is sufficient to “describe” the whole dataset.
For example, if you have an anomaly in your dataset that you should be paying attention to, or some unrelated images that were inserted by mistake, this tool will help you find them, and then you’ll be able to decide whether to label them or discard them.
The most important value you can get out of this tool is the ability to gain information from your dataset without any required labels. We use self-supervised and unsupervised learning methods in order to analyze the data, which we primarily represent in a high-dimensional embedded feature representation.
Our method investigates how every item is represented in comparison to the other items in the same data collection, revealing hidden patterns, and by using anomaly detection, cluster analysis, and smart sampling, it extracts the dataset’s most informative representation. This tool is also a learnable one, meaning it could be used to train your unlabelled datasets in order to improve its ability.
Summed Up
The days of manually going over your dataset and trying to find the most informative parts are over. With our new Information Extraction Tool, you can quickly get a “tailor-made” dataset that covers exactly what you need to focus on, in order that you can handle your time and annotation budget smoothly and efficiently. Let us know if you’d like to try it!
