Leveraging Active Learning For Visual Inspection
In the rapidly evolving field of AI, the quest for more accurate and efficient models is unending. One promising approach that has gained significant traction is active learning. In this article, Miron Shtiglitz, director of product management at AI for visual inspection specialist QualiSense, explores how active learning can be a game changer in improving the performance of AI models in the field of quality inspection.
Active learning is a machine learning paradigm that involves the model actively selecting the most informative data points for training, rather than relying solely on randomly chosen samples, with the goal of achieving higher accuracy with fewer labelled examples. In traditional supervised learning, a model is trained on a fixed dataset with pre-labelled examples. However, labelling data can be expensive and time-consuming.
While active learning itself is not a new methodology, its application to the field of defect detection and quality inspection is pioneering. In this article, we have outlined four core benefits to active learning, although there is significant overlap between these categories. Here, we look at the four key benefits and their implications for detecting defects in production processes.
Optimising Data Efficiency: Active learning enables AI models to learn more from fewer labelled examples. Traditional machine learning models often require large amounts of labelled data to generalise well. In contrast, active learning identifies the most uncertain or ambiguous instances, actively seeking additional information where it matters most.
This results in a more efficient use of labelled data, reducing the annotation burden and lowering the overall cost of model development. Until now, the need for a higher number of images and the time and resources involved in the model training process has been the biggest obstacle to the introduction of AI models in defect detection, but active learning overcomes this challenge.
Reducing Class Imbalance: Many real-world use-cases suffer from class imbalance, where certain classes have limited representation. Active learning can help mitigate this challenge by focusing on underrepresented classes. By actively selecting samples that belong to minority classes or are particularly challenging for the model, active learning helps in creating a more balanced and robust AI model.
This is particularly important in production, where certain types of recurring defects are common. For example, imagine a machine within a production process repeatedly produces a chip in the left corner of a component. Rather than showing the model thousands of images of the same defect, active learning allows us to teach the model to detect this defect with fewer images, shifting time and resources to other less common types of defect.
Without a solution to the class imbalance problem, previous models have been generally good at detecting common defects as these classes are overrepresented in the data they are trained on. In comparison, rare types of defect are neglected and therefore slip through the net of detection. By actively seeking out outliers, active learning optimises the model to detect even the rarest of defects.
Reducing Annotation Costs: The cost of annotating large datasets can be a significant bottleneck in AI development. Active learning strategically selects data points for annotation, reducing the number of labelled examples needed. This not only saves time and resources but also facilitates the development of AI models in scenarios where obtaining labelled data is a resource-intensive process.
Continuous Model Improvement: Active learning allows models to adapt and improve continuously by actively seeking out samples that challenge their current understanding. This adaptability is crucial in dynamic environments, such as production lines, where the model needs to evolve with the incoming data.
As production continually adapts, models need to be updated with new data. The ability to do this with less data is therefore key. This applies even with samples of okay or non-defective products. In a dynamic production environment, what is considered okay today may not be the same in future.
Enhanced Model Generation: Active learning focuses on uncertainty, selecting samples that the model finds challenging. By actively addressing regions of the data distribution where the model lacks confidence, active learning helps improve generalisation performance. This is particularly valuable in scenarios where the model needs to perform well on diverse and unseen data.
QualiSense was founded in 2022 in partnership by Cortica and Johnson Electric to develop a groundbreaking augmented AI platform that processes non-labelled production data to train itself to detect defective parts with minimal user guidance.
For more information: www.qualisense.ai
