Introduction
Knowledge distillation is a popular ML technique that deals with the transfer of knowledge between two models. Here, the knowledge travels from a larger model to a smaller one. Knowledge distillation sees great use in the field of deep learning, where it is used for purposes like compressing models and transferring knowledge. And the best part? There is no validity loss during this transfer.
Mathworks stats show that when a smaller neural network gets training using KD, its accuracy increases a lot when compared to typical cross-entropy loss. So, it is clear that knowledge distillation not merely transfers knowledge but also improves the smaller neural network.
History
In 2015, Geoffrey Hinton, along with Oriol Vinyals and Jeff Dean, proposed a method of training shallow models with the aid of already-trained ensembles. They termed it ‘Knowledge Distillation’ as you distill the knowledge from an existing model into a new one. This appears to be like a teacher teaching a student kind of thing. That is why this method is also referred to as ‘Teacher-Student Learning.’ In Knowledge Distillation, the authors employed the output probability of the pre-trained model and used that as the label for the new shallow model.
Also Read: What is Dense Layer in Neural Network?
While not explicitly named ‘Knowledge Distillation,’ this idea came from a paper published in 2006 by Caruana et al named ‘Model Compression.’ The authors utilized a huge ensemble model consisting of hundreds of base-level classifiers that were one of the most sophisticated classification models at that time. They used it to label a lot of data and trained thereafter a single neural network on that labeled data set in the intended way of supervised learning. This compact model was a thousand times tinier and faster. Plus, it was capable of achieving the same results as that of the ensemble.
What is Knowledge Distillation?
Knowledge distillation is a machine learning paradigm that concerns itself with training smaller models (the student model) to leverage large pre-trained models (the teacher model). It is helpful in deep learning for model compression and knowledge transference. That is true, especially for large deep neural networks.
The purpose of knowledge distillation is to teach a smaller model how to act like a larger, more sophisticated model. In traditional deep learning for example, the goal is to teach the artificial neural network in such a way that the predictions of the artificial neural network trained on the training data set get as close as possible to the output examples of the training data set. In contrast, the main attention when distilling knowledge is towards enabling the student network to produce predictions that match the ones of the teacher network.
Also Read: What is AI Energy Consumption?
There are three distinct types of knowledge in knowledge distillation:
Response-Based Knowledge
Response-based knowledge focuses on the last layer of the target network. The concept is that the student network attempts to reproduce the output of the teacher network. For this, a loss function called distillation loss is used. It quantifies the difference between the student and teacher model logits. As this loss diminishes in the course of training, the student model improves in predicting the output of the teacher model.
In computer vision tasks such as image recognition, response-based knowledge includes soft targets. The soft targets depict the probability distribution over the output classes with the help of the softmax function. A parameter referred to as ‘temperature’ helps control the contribution of each soft target to the knowledge. Generally speaking, response-based knowledge distillation that utilizes soft targets is a typical technique in supervised learning.
Feature-Based Knowledge
Feature-based knowledge deals with the information that stays concealed within the intermediate layers. In case you were unaware, a teacher model that is already trained contains useful information that resides in its intermediate layers. This is also important for deep neural networks. These layers distinguish certain aspects, and that can help the process of student model training.
The goal is for the student model to reproduce the activations of the features of the teacher model. The distillation loss function allows one to achieve this by minimizing the distance between the feature activations of the two models.
Relation-Based Knowledge
Both response-based as well as feature-based knowledge are known for concentrating on particular model layer outputs. On the contrary, relation-based knowledge’s focus stays on the tie-ups between various layers or feature maps that depict the activations at various locations/layers.
Basically, relation-based knowledge is a solid way to train the student network to mimic how the teacher model thinks. You can model these relationships and connections in a bunch of different ways, like looking at correlations between feature maps, using matrices to show how similar different layers are, or working with feature embeddings and the probability distributions of those feature representations.
Source: arxiv
How Does Knowledge Distillation Work?
Knowledge distillation is a technique in deep learning where a complex, well-trained model (referred to as the “teacher”) shares its knowledge with a simpler, more lightweight model (aka the “student”). Here’s how it works:
1. Teacher Model Training
- The teacher model is first trained using labeled data for pattern and interrelationship discovery purposes.Â
- The teacher model has a big capacity that helps it pick up on tiny details. This leads to better performance on the task at hand.
- The teacher then makes predictions based on the training data provided. These predictions act as a standard that the student model needs to mimic. Â
2. Knowledge Transfer to the Student Model
- The Student learns using the same data as the teacher model. But there is a little dissimilarity.Â
- The student model is trained using soft labels instead of the usual hard labels. Soft labels provide a much more detailed representation of the data. This is because they are probability distributions over the classes given by the teacher model.
- The use of soft labels is essential in making the student model capable of not just copying the decisions of the teacher but also understanding why the teacher made those decisions.Â
- This way, the student model gets the power to grasp and emulate the knowledge from the teacher model. This leads to a more concise data representation.
Some popular knowledge distillation algorithms are:Â
- Adversarial Distillation: Adversarial distillation enhances student models through the aid of adversarial training. Here, students are trained with outputs from the teachers and, simultaneously, with difficult artificial data. It consists of training the teacher on the ground truth labels, whereas the student is trained with synthetic and real data. An adversarial network generates difficult samples for the teacher, and the student learns to classify those samples. This approach contributes to increasing the robustness against attacks and enriching the student’s generalization abilities by giving them hard examples.
- Multi-Teacher Distillation: Multi-teacher distillation is the process by which a whole bunch of teacher models are used to train one student model – this gives the student model variety and a wide range of features to perform better. There are two separate phases in the training process. In the first phase, each teacher model is trained individually before the student model is trained with all the teacher models in the second phase. This has the effect of reducing overfitting and allows the model to learn from different perspectives, thus, it makes the model stronger.
- Cross-Modal Distillation: Cross-modal distillation is about sharing knowledge between different modalities of data, which is super handy when you have info in one but not in another. For example, you might have text descriptions for images but no labels. The way it works is you first train a teacher model using the data from the source modality, and then you train a student model on the target modality, using the teacher’s outputs as a guide. This helps the student model get better at understanding the target modality by tapping into what the teacher knows.
Also Read: Top AI Story Generators for Crafting Immersive Interactive Tales in 2025
Definition With an Example
Amazon has shown well how knowledge distillation benefits from contrastive decoding and counterfactual reasoning. There are some risks with knowledge distillation, like irrelevant reasoning popping up. Student models might churn out answers that don’t really connect logically.
With contrastive decoding on the teacher side, the company ensured that the reasons behind true claims are quite different from those behind false ones. Counterfactual reasoning trains the student model to handle both true and false rationales effectively.
This strategy beats the usual methods in reasoning tasks, and varying the training helps clear up any confusion. Also, contrastive decoding leads to more convincing rationales, improving outcomes in subsequent tasks. This led to Amazon researchers winning the Outstanding Paper award in 2023’s ACL.
Source: rmoklesur.medium
Conclusion
Knowledge distillation improves small models by transferring knowledge from large models. It enhances efficiency and performance in AI applications like computer vision, NLP, and speech recognition. Besides, it makes ML models more competent. Research is still ongoing, and it will only transform KD for the better.
For more information on Tech, click on the links given below:
