External links

AppNeCo: Approximate Neurocomputing

GA22-02067S [Registered results] [WWW] 2022 - 2024

Nowadays, modern AI technologies based on deep neural networks, whose computation is demanding on energy consumption, are implemented in devices with limited resources (e.g. battery powered cellphones). In error-tolerant applications (e.g. image classification), the use of approximate computing methods can save enormous amount of energy at the cost of only a small loss in accuracy. AppNeCo is a basic research project of approximate neurocomputing, whose ambition is an original synergy of approximation and complexity theory of neural networks and empirical experience with the top design of high-performance approximate implementations of hardware circuits. Its goal is to develop complexity-theoretic foundations of approximate computation by convolutional neural networks (CNN) of bounded energy complexity for application domains specified by input space distributions. This knowledge will be used in designing new strategies for approximating components and learning algorithms of low-energy high-precision CNNs. The new methods will be tested on image processing tasks.