LogiNet – “Logic-Driven Efficient Computing and Analysis of Deep Neural Networks on Hardware” – deals with deep neural networks (DNNs). DNNs are the algorithms used to realise Artificial Intelligence. They are inspired by the structure of the human brain and consist of many layers, each of which processes data in a different way and then sends the processed data to subsequent layers. DNNs have achieved major advances in many fields and are the drivers for various applications such as image recognition and text processing. However, they require enormous computing power: the massive multiplications and additions that repeatedly exist inside DNNs, known as multiply-accumulate (MAC) operations, are complex and consume enormous energy on today's computer hardware, for instance, graphics processing units.

Until now, research has focused on simply accelerating MAC operations rather than examining their actual logical functions after training. The blind execution of massive MAC operations not only consumes a lot of energy but also makes it difficult to understand the decision-making insides of DNNs. The internal decision-making processes of DNNs are almost impossible to examine and verify.

Professor Grace Li Zhang (Hardware for Artificial Intelligence Lab) is addressing these issues as part of LogiNet. She is investigating a new way of executing DNNs that focuses on the logical structure and highly abstract function of neural networks. Not only improving computing power, the project also examines the logic reasoning inside neural networks – in other words, what the networks actually do logically. The idea is to convert data that has already been learned (‘weights’) directly into simple logical circuits of MAC operations and wire them together, so to speak. This reduces complexity, computing power, data movement and energy consumption. At the same time, it also provides a better understanding of how the neural networks make decisions. LogiNet analyses the properties of DNNs, such as hardware reliability and fairness in decision-making.

The research aims to lay the foundation for DNNs to consume less energy and thus move closer to the goal of “green AI” on a wide variety of hardware platforms, from small sensors to large data centres. In addition, the project will ensure that neural networks can be better analysed in terms of transparency, fairness and reliability.