Vojtěch Mrázek delivered a public habilitation lecture

On Friday, February 27, 2026, Ing. Vojtěch Mrázek, Ph.D., delivered his public habilitation lecture entitled „Energy-efficient neural networks for embedded systems“.

Evolutionary optimization of neural network accelerators – the topic of Mrázek's habilitation thesis summarizes well the author's scientific journey to date. Neural networks (hereinafter NN) are the stars of mathematical models used in the field of information technology, a concept we encounter every day. Mrázek does not deal with the algorithmic design of these networks, but with how to calculate (implement) the resulting networks as efficiently as possible, given their significant energy consumption and the type of hardware on which they are to "run." And that is a global topic. After all, today we are able to run a limited neural network even on portable devices, in mobile phones, on various IoT components, etc. These limited versions of NN are often sufficient for us – we expect a thermostat to only recognize our voice commands; we want our smart watches to simply detect the type of activity we are doing, etc. However, even this limited performance has its requirements for computing capacity and energy resources. Hand in hand with the development of artificial intelligence, hardware must also evolve so that it is capable of implementing NN even in small portable devices. Today, this is often aided by a so-called neural network accelerator. And this is precisely where the main scientific challenge lies, one that Vojtěch Mrázek has been successfully tackling for several years: how to assemble and configure the hardware itself, specifically the accelerator, so that the resulting neural network with which it is to work functions effectively.

Identifying and designing the parameters of the NN accelerator (memory organization, setting the size and structure of computing units, or the method of calculation) is far from a straightforward task. "Of course, we are able to identify the ideal combination of neural network and hardware solution. But the real goal is always a compromise between energy consumption and computing power," says Mrázek, describing the desired final state. The topic of his habilitation lecture focuses precisely on ways to get neural network inference computing into the embedded system. "There are 'tricks' we have to do. Various network compressions such as quantization, pruning, methods such as introducing targeted errors into the calculation, and so on. I will also show how hardware can be optimized for a specific type of neural network."

From dissertation to major GA ČR project

Vojtěch Mrázek approaches the design of energy-efficient hardware accelerators through evolutionary algorithms.

In an earlier interview for the faculty website, he defined his research position using an analogy of three layers: "We have our own chip software architecture, e.g., for processing internet traffic or video, which are topics that other research groups at the faculty are working on. Then there is the level of chip technology, the 'construction' of hardware, which microelectronics engineers and physicists focus on. And between them there is a layer called EDA, Electronic Design Automation. That's us. We are in the middle, straddling, figuratively speaking, between software and hardware. You can think of us as a translator between hardware and chip software architecture. I should ensure the most efficient and energy-efficient communication between both levels." The experimental results, which he arrived at in collaboration with his colleagues and which he also describes in his habilitation thesis "Evolutionary Optimization of Neural Network Accelerators," demonstrate significant energy savings in all the architectures studied. This includes the most elementary solution, which is printed electronics, which Mrázek addresses in his habilitation thesis as an example demonstrating the breadth of applicability of the methods he proposes.

The habilitation thesis itself nicely summarizes the gradual development of Mrázek's research. His dissertation, which he defended in 2018, focused on the advanced use of evolutionary algorithms for working with hardware circuits: Together with his colleagues, he attempted to design circuits at various levels of description, from small transistors to gates to larger blocks. In his dissertation, he demonstrated that evolutionary algorithmization in the field of approximate computing works. From this topic, he then moved on to the problem of optimizing entire systems, i.e., embedding components designed during his dissertation into applications, in a postdoctoral position at TU Wien, where he spent a year. And what about Vojtěch Mrázek's present and near future in research? It is best defined by the recently awarded prestigious GA ČR Junior Star grant, which, among other things, allows him to put together a new research team at the faculty: "Evolutionary algorithms alone are no longer enough for us; we have to help ourselves with machine learning. However, machine learning and evolutionary algorithms influence each other and, as a result, improve each other." As part of the project, the team aims to combine the advantages of evolutionary electronic circuit design and advanced machine learning methods to improve chip design. At the same time, they want to take advantage of the fact that the influence of evolutionary algorithms and AI algorithms is mutual—evolutionary design algorithms are also improved in this interaction. Mrázek adds that they try to keep up with machine learning trends, even though this is a big challenge due to its dynamic nature. "We try to reflect trends. We are not concerned with the orthodoxy of the process, i.e., strictly adhering to evolutionary algorithms alone. We want to achieve a good result. And the combination with machine learning is currently a clear promise."

Habilitation? A natural step

And how does Vojtěch Mrázek view habilitation itself? A major turning point, or perhaps a necessity? "It's part of a natural progression, and I'm glad for that. I appreciate being able to build a team, work with doctoral students, and participate more intensively in teaching. The opportunity to conduct joint research with foreign universities is a huge benefit for me, something I never even dreamed of," comments Mrázek. "Academic creative freedom is a huge advantage. For example, in doctoral studies, you are given a free choice of topic and it is up to you which path you choose. And, of course, it can also be a dead end in research. That happened to me too," he adds with a smile. "At moments like that, the support of your family is especially important," concludes Mrázek, looking back on his scientific career so far. If you did not attend the lecture in person, please use the online recording, which is available HERE.