The use of adaptive resonance theory for creating artificial neural networks can help in solving the stability-plasticity dilemma.
Suppose you grew up in one town and moved to another for several years, and now you have to move back. The streets and avenues of that town would have changed over the years due to progress and construction. To navigate through this town, you have to combine your learning of new streets with your old knowledge of navigation. Therefore, you will have to incorporate new pieces of knowledge and simultaneously not forget your existing knowledge. You cannot choose one over the other. And that’s the kind of problems the stability-plasticity dilemma raises.
The stability-plasticity dilemma is a significant constraint in artificial and biological neural systems. The underlying idea is that learning in neural networks requires plasticity so that it can integrate new knowledge, and it also requires stability so that it does not lose previous knowledge. Adaptive resonance theory (ART), developed by Stephen Grossberg and Gail Carpenter, is a theory developed to address the stability-plasticity dilemma. The terms “adaptive” and “resonance” means that it can adapt to new learning (adaptive) without losing previous information (resonance).
The basic ART architecture consists of three layers. F1 layer that processes all the input data. F2 layer has several cluster units, and the unit with the most significant input data becomes the first unit for learning, and rest units are ignored. The third layer is the module layer, where the module unit decides whether the cluster unit should be allowed to learn new things based on how similar or different is the input data of previous knowledge and new information. Adaptive resonance theory is an architecture used to create different types of neural networks that provide several applications across all the industries.
Applications of Adaptive Resonance Theory
As ART architecture can help create neural networks that can quickly adapt to changes, most of its applications are where real-time analysis is required.
For self-moving humanoid robots, real-time motion control is of utmost importance as they have to move in a dynamic environment. And if they are cannot make real-time motion control, then they might keep on colliding with the obstacles on their way. For instance, if a person suddenly moves a chair in front of the robot, then the robot might collide with it and stop. The surroundings of a robot can change suddenly, and hence motion control in the dynamic environment becomes challenging for developers. ART can help adapt to those sudden changes for real-time motion control. For instance, a mode-adaptive neural network is an adaptive neural network that can provide real-time quadruped control.
Face recognition systems work by matching the face of a person with images provided in training data. But, emotions can create slight changes in the overall facial structure of a person, and face recognition systems might not be able to detect them accurately. ART-based neural networks can adapt to the changing emotive faces of a person for accurate recognition.
If researchers solve the dilemma issues with the help of adaptive resonance theory, then they can create adaptive neural networks that can adapt to changes, for all AI applications. And this might provide a chance to create a single neural network that can do the functions of multiple networks. For instance, an adaptive neural network for facial recognition will also be able to detect emotions.