Hierarchical Network Dissection (HND) is a technique used to interpret face-centric inference models. This method pairs units of the model with concepts in a "Face Dictionary" to understand the internal representation of the model. HND is inspired by Network Dissection, which is used to interpret object-centric and scene-centric models.
Understanding HND
Convolution is a widely used technique in deep learning models. A convolutional layer in a deep learning model contains multiple filters, and each filter is trained to recognize a different feature of an image. HND identifies interpretable units in a convolutional layer that correspond to facial concepts. By pairing these units with facial concepts from a Face Dictionary, researchers can understand what these units are learning.
The Face Dictionary contains a collection of facial concepts with corresponding sample images. Examples of these concepts include nose, mouth, eyes, and forehead. By pairing these concepts with interpretable units, researchers have a better understanding of how the deep learning model is recognizing and interpreting facial features.
Benefits of HND
HND provides a way for researchers to better understand the internal workings of deep learning models. This is important because deep learning models are often referred to as "black boxes" due to their complex nature. HND provides a way to open up these boxes and gain insight into what the model is truly learning.
Understanding how deep learning models interpret facial features can have many real-life applications. For example, it can improve the accuracy of facial recognition systems used in security and surveillance. It can also be used in medical imaging to help identify and diagnose facial abnormalities.
How HND Works
HND works by using a probabilistic formulation to pair units of the model with facial concepts. This is done by comparing the responses of each unit in the convolutional layer to the sample images in the Face Dictionary. The units that have the highest responses to a particular concept are paired with that concept. This process is repeated for all concepts in the Face Dictionary to determine which units are most strongly associated with each concept.
Once the units are paired with facial concepts, researchers can analyze them to determine what features they are recognizing. For example, a unit that is strongly associated with the concept of a nose may be recognizing the shape or size of a nose in an image.
Pitfalls of HND
Despite its benefits, HND is not a perfect technique. One of the main challenges of HND is that it requires a Face Dictionary with a large number of concepts and sample images. Creating a comprehensive Face Dictionary can be time-consuming and difficult.
Another challenge of HND is selecting the appropriate threshold for unit-concept affinity. If the threshold is set too high, important associations may be missed. If the threshold is set too low, meaningless associations may be included.
Future of HND
HND has the potential to be used in a wide variety of applications. The technique can be adapted for use in other types of models besides face-centric ones. It can also be used to interpret other types of data, such as text or audio.
As deep learning models continue to be used in more applications, the need for interpretability techniques like HND will only increase. HND and other techniques like it are an important step towards creating deep learning models that are truly transparent and understandable.
