Ground Truth in Machine Learning: Importance and 7 Key Challenges

What is Ground Truth?

“Ground truth” is a term commonly used in statistics and machine learning. It refers to the correct or “true” answer to a specific problem or question. It is a “gold standard” that can be used to compare and evaluate model results. 

For example, in an image classification system, the algorithm learns to classify each image into a set of classes (e.g., cat, dog, sheep). The algorithm trains using training data with ground truth labels indicating the true class of each object. When the model is ready, it is applied to unseen images. 

An ideal model will generate a prediction identical to the ground truth label for every source image. Or in simpler language—it will know that an image of a cat is a cat, and an image of a sheep is a sheep.

Why is Ground Truth Important in Machine Learning?

In supervised learning algorithms, ground truth data is critical to training new algorithms. The more annotated data is available, and the higher its quality, the better algorithms will perform.

In many cases, ground truth labels must be prepared by human evaluators or annotators. This is costly and time consuming, especially if the dataset contains thousands or millions of entries. Because assembling large datasets with ground truth labels can be a huge task, many researchers have taken upon themselves to create a high quality dataset which can serve as a benchmark or initial training ground for new algorithms.

These datasets are a basic requirement for many types of machine learning research. To see examples of large-scale datasets in the computer vision field, see our detailed listing of face datasets and image datasets.

How Supervised and Unsupervised ML Models Use Ground Truth Labels

Supervised Ground Truth Labels

Supervised methods learn a target domain by training on large amounts of data with corresponding ground truth labels. Initially, the model trains on data with ground truth labels. When an initial version of the model is ready, it is evaluated on evaluation and testing datasets where the model does not know the ground truth label. Predictions are compared to the actual ground truth label to determine model performance.

Semi-Supervised Ground Truth Labels

In a semi-supervised approach, ground truth labels are needed only for some data points. The problem with this approach is that its success depends largely on the available examples. It is suitable for problems where you need to identify individual structures in your data, but is less suitable for complex problem spaces with only a small difference between classes.

Read our benchmark report Hands-Up: Leveraging Synthetic Data for Hands-on-Wheel Detection

A Process for Creating Ground Truth Datasets

Here is a general process for creating a large-scale dataset with ground truth labels:

  1. Planning—in a new project, the first step is to determine the requirements of the algorithms that will train on the data. You’ll need to identify the volume of data needed, the type and format of data, and what level of variability exists in the real-world population being modeled. The dataset must take into account all relevant edge cases. 
  2. Pilot project—most dataset projects begin with a small pilot in which researchers collect a small number of relevant data samples. The objective at this stage is to identify challenges in data collection, the time and skills required to source and annotate the data, and build the relevant project team. 
  3. Data privacy and compliance—before beginning the project, it is important to involve legal or compliance functions at the organization to determine what are the compliance implications of collecting the data. In today’s legal environment there are many restrictions on collecting information that can be used to identify living humans.
  4. Project design—based on the pilot project, researchers design the full scale project—data sources, how many people will participate in data collection, and methods for validating the data and ensuring high quality. In some cases, automated mechanisms or existing data sources can be used to reduce the volume of annotation work required.
  5. Annotation—the team hires a team of annotators, who may be in-house, freelancers, or crowdsourced, who review data samples and annotate them according to project specifications.
  6. Reviewing data quality—once datasets are ready, the team evaluates the quality of annotations and possible biases that datasets are subject to. The trained model will only be as good as its training data, so this step is critical to ensure sufficient model performance.

 

7 Challenges of Collecting and Managing Labeled Data

Here are some of the challenges you might encounter when setting out to collect a large-scale ground truth dataset:

    • Collecting enough data—is it difficult to know in advance how much data will be needed to train an algorithm. 
    • Ensuring sufficient variability—the dataset must have enough data about all relevant classes and edge cases that must be handled by the algorithm.
    • Data privacy issues—it is becoming increasingly difficult to gather ground truth data that includes images or data concerning living humans. 
    • Cost of data collection—the cost and labor involved in creating any large dataset is immense, and involves a large logistical effort.
    • Time to market—large-scale ground truth dataset can take months or years to assemble, delaying machine learning research and applications. 
    • Data quality—there is no guarantee that the data collected will be of sufficiently high quality to train the algorithms on all necessary tasks and use cases.
    • Generality—some datasets may be highly suitable for specific algorithms or use cases, but may not generalize well to other use cases. This can limit the usability and return on investment of a data collection project.

Read our benchmark report Hands-Up: Leveraging Synthetic Data for Hands-on-Wheel Detection