Researchers are tackling the important challenge of understanding how neural networks arrive at decisions, especially in sensitive applications. Jiayi Dai and Randy Goebel from the University of Alberta, along with their collaborators, introduce a new approach to evidence extraction, a technique designed to create inherently interpretable artificial intelligence systems. Their study introduces REKD (evidence extraction through knowledge distillation). This is how the “student” network learns not only from the predictions of the task, but also from the rationale provided by the more competent “teacher” network, which is the main feature driving its predictions. This innovative knowledge transfer significantly improves the performance of less powerful models, provides a path to deploying interpretable AI even with limited computational resources, and represents a step towards AI systems that learn more like humans.
This research addresses a key challenge in the field of explainable artificial intelligence: improving the reliability of deep neural networks in high-stakes applications such as medicine and finance.
Researchers tackled the challenge of training feature selectors (generators) and predictors simultaneously within evidence extraction. This process becomes complicated when the base neural network does not have sufficient capacity. This breakthrough reveals a new approach inspired by human learning. In this approach, the “student” RE model learns not only through its own exploration, but also by leveraging the rationale and predictions of the more powerful “teacher” network.
The team accomplished this by employing knowledge distillation to transfer interpretable knowledge from teachers to students, thereby overcoming the chicken-and-egg problem inherent in traditional evidence extraction methods. This structural adjustment is consistent with how humans effectively learn from verifiable knowledge, allowing students to benefit from the teacher’s pre-verified selection of features.
This research established a method that does not rely on neural models. This means that any black box neural network can be integrated as the foundation of the REKD framework. Experiments conducted across verbal and visual classification datasets such as IMDB movie reviews, CIFAR 10, and CIFAR 100 demonstrate that REKD significantly improves the predictive performance of student RE models.
Specifically, the team utilized variants of the BERT and Vision Transformer (ViT) models to validate the feasibility of their approach. Furthermore, this study reveals a curriculum of progressively increasing complexity by synchronizing the Gumbel-Softmax annealing scheduler and the knowledge distillation temperature. This allows students to initially absorb a wide range of knowledge from the teacher and gradually refine their feature selection as the temperature decreases, enhancing accuracy during the discretization phase. This work paves the way to building more transparent and reliable AI systems, increasing confidence in decision-making processes and enabling widespread adoption in critical areas.
Teacher and student learning with evidence and prediction improves evidence extraction performance
Scientists developed Evidence Extraction with Knowledge Distillation (REKD) to improve the performance of evidence extraction models, especially when using less powerful neural networks. In this study, we tackled the challenge of simultaneously training a generator responsible for feature selection and a predictor, a complex process, with limited supervision signals.
Taking inspiration from human learning, the researchers hypothesized that a teacher model with an established and testable rationale could effectively guide the learning process of student models. To implement REKD, the team designed a system in which student RE models learn not only by optimizing their own rationale extraction, but also by leveraging the rationales and predictions of pre-trained teachers, called “rationalists.”
This approach avoids the chicken-and-egg problem inherent in standard RE training, where generators and predictors are interdependent. In our experiments, we used the Straight-Through Gumbel-Softmax estimator to enable the selection of differentiable features, which is a key step in gradient-based optimization.
This work pioneered a progressive complexity curriculum by synchronizing the knowledge distillation temperature with an annealing scheduler. This technique initially allows students to absorb a wide range of knowledge from the teacher’s soft targets, and gradually focuses on selecting precise features.
The team leveraged variants of both the BERT and Vision Transformer (ViT) models as backbone architectures to validate the viability of REKD across diverse datasets including IMDB movie reviews, CIFAR 10, and CIFAR 100. Results demonstrated that REKD significantly improved the predictive performance of student RE models across verbal and visual classification tasks, confirming the effectiveness of knowledge distillation in this context.
Significantly improve evidence extraction performance by distilling knowledge and optimizing evidence length
Scientists have achieved significant improvements in the predictive performance of evidence extraction (RE) models using a new technique called REKD (Evidence Extraction with Knowledge Distillation). The research focused on leveraging both theoretical foundations and predictions to enable less powerful neural networks, called “students,” to learn effectively from more powerful “teacher” networks.
Experiments show that REKD significantly improves accuracy across verbal and visual classification tasks, including IMDB movie reviews, CIFAR 10, and CIFAR 100 datasets. As observed for the ViT model in CIFAR 10, the results show a clear positive correlation between predictive performance and rationale length.
The team measured that changing the theoretical percentage target (ptarget) had a direct impact on accuracy, with smaller ptargets imposing stronger constraints and generally decreasing performance. Importantly, REKD significantly improved students’ RE models and reduced variance, as detailed in Tables 1, 2, and 3.
For example, in CIFAR 10, REKD improved the accuracy of ViT Small from 0.889 to 0.968 and ViT Tiny from 0.797 to 0.936 (both with a theoretical ratio of 15%). Measurements confirm that the drop in accuracy from classification to evidence extraction becomes more pronounced when the underlying neural model has limited capabilities.
Knowledge distillation improves learning of rationale through synchronous annealing of teacher and student models
Scientists have developed Evidence Extraction by Knowledge Distillation (REKD), a new technique to enhance the performance of evidence extraction models aimed at creating more interpretable artificial neural networks. REKD improves predictive performance by allowing “student” neural networks to not only explore their own rationales, but also learn from the rationales and predictions of more competent “teacher” networks.
This approach reflects how humans effectively learn from interpretable and verifiable knowledge and addresses common challenges when training small or less powerful networks. This study demonstrates that leveraging datasets such as IMDB movie reviews and CIFAR 10/100, as well as various BERT and visual transformation models, significantly improves prediction accuracy for both verbal and visual tasks.
By synchronizing the knowledge distillation temperature with Gumbel softmax annealing, REKD accelerates the curriculum of student networks and allows them to learn more efficiently. The authors acknowledge the limitations of current experiments that focus on distillation between networks of the same architecture, and suggest that future research may consider distillation between different architectures, such as ViT and ResNet.
Furthermore, we propose dynamic weight scheduling for rationale extraction and knowledge distillation loss conditions, which may further optimize the performance. Future applications may include deployment in resource-constrained environments such as mobile healthcare devices, and extensions of the method to extract other discrete latent structures such as relational graphs.
