Czech Science Foundation President’s Commendation Awarded to Jiří Šíma for Research on Energy-Efficient Neurocomputing (AppNeCo)

Czech Science Foundation Presidenťs Commendation Res. Prof. Jiří Šíma from the Institute of Computer Science of the Czech Academy of Sciences (ICS CAS) has been awarded the Honorary Certificate of the President of the Czech Science Foundation (GACR) for the outstanding outcomes of the project AppNeCo: Approximate Neurocomputing. Conducted in collaboration with researchers at the Faculty of Information Technology, Brno University of Technology (FIT BUT), the project established a novel theoretical model of the energy complexity of neural networks and derived fundamental energy consumption limits independent of any specific hardware implementation.

[ [list] ]

Approximate Computing as a Pathway to Energy Savings

With the widespread deployment of deep neural networks in domains such as object recognition, text translation, and autonomous driving, their computational and energy demands have grown rapidly. Every evaluation of a network requires hundreds of millions to billions of arithmetic operations. In many practical tasks, however, absolute numerical precision is not strictly necessary. The paradigm of approximate computing relies on intentionally simplifying a portion of intermediate calculations, significantly lowering power consumption while preserving virtually identical result quality.

Bridging Mathematical Theory and Hardware Engineering

Hailo

The project AppNeCo: Approximate Neurocomputing (GA22-02067S), funded by the Czech Science Foundation, united Dr. Jiří Šíma’s team at the Institute of Computer Science of the CAS—focusing on mathematical neural network theory, statistical robustness, and computational complexity—with Prof. Lukáš Sekanina’s team at FIT BUT in Brno, which specializes in hardware accelerators and digital circuit design. This collaboration confronted theoretical and engineering perspectives, allowing mathematical models to inspire novel hardware architectures while empirical implementation insights highlighted the most crucial theoretical questions for real-world applications.

Key Discovery

Fundamental Energy Consumption Limits of Neural Networks One of the project's significant outcomes is the development of a novel machine-independent model of the energy complexity of convolutional neural networks that accounts for the energy costs of data transfers between memory and the processor. The proposed model makes it possible to calculate and compare the fundamental energy limits of neural networks independently of the specific hardware, even before their chips are physically fabricated and tested.

Publication Success and Ongoing Research

Hailo slide

The project's findings were published in leading international venues, including Neural Networks in the first decile of the field. The fruitful scientific collaboration between both research teams continues in the follow-up project LEDNeCo: Low Energy Deep Neurocomputing (GA25-15490S), which focuses on energy-efficient approximations of deep architectures, including transformers. Res. Prof. Jiří Šíma has been working at the Institute of Computer Science of the CAS since 1991; he co-authored the first Czech textbook on neural networks in 1996 and has dedicated his scientific career to elucidating the theoretical limits of computational power, learning, and energy complexity in artificial intelligence.