Michal Piňos from the Department of Computer Systems will defend his dissertation in June

We invite you to the defense of the dissertation by Ing. Michal Piňos from the Department of Computer Systems, FIT BUT, which will take place on Monday, June 29, 2026, at 2:00 p.m. in conference room G108. The advisor for the dissertation, titled “Automated Design of Approximate Convolutional Neural Network Architectures for Energy-Efficient Inference,” is Prof. Lukáš Sekanina.

Michal Piňos focuses on the automated design of neural network architectures (NNs) that should not only be accurate but also energy-efficient when running on specific hardware. This is important, for example, for devices with limited power—typically sensors, embedded systems, and the like. Specifically, he focuses on convolutional neural networks, a type of NN most commonly used for processing image data (e.g., object recognition in images). “Neural networks consist of layers. The challenge is to choose which layers to use, how many there will be, and how to connect them to each other—and each layer also has its own parameters,” says Piňos, explaining the starting point of his work. A fundamental premise of his approach is the fact that there is no single universal neural network suitable for all purposes. “For every task and every type of hardware, a neural network must be designed to be optimal.” The term “optimal” encapsulates the core of the problem: The network must be as accurate as possible for the given task, but at the same time, it must consume as little energy as possible.

In convolutional neural networks, inference — that is, the application of a pre-trained model — is particularly computationally intensive. During training, the network is repeatedly presented with input data and the correct answers, and its internal parameters are gradually fine-tuned using the backpropagation method. “Inference means that we already have a trained model available and are putting it into practice. We provide it with input, the data passes through the network once, and we receive the output,” says Michal Piňos, explaining the context of the problem. The main question of his research is: Is it possible, already during the design of a convolutional neural network, to estimate how much energy it will consume on a given hardware platform? In other words, how can we ensure energy efficiency during inference in advance? “The ideal scenario is to find an algorithm to which we specify, as input, what we want to do, where we want to do it—that is, on what hardware—and with what parameters, for example, regarding energy consumption. And the algorithm will design the parameters of such a convolutional neural network,” says Piňos, specifying the research goal.

At the core of the author’s dissertation is the use of approximate computations. In layman’s terms: We’re talking about a situation where the hardware doesn’t always compute with absolute precision, but with sufficient precision. The most energy-intensive arithmetic operation in convolutional neural networks is multiplication—it accounts for the majority of the energy consumed during inference. The idea, then, is to use approximate multiplication. The benefit? Reduced energy consumption. Neural networks are robust computational models capable of handling errors. “Moreover, with convolutional neural networks, we’re working—to put it simply—with image data, with pixels, so minor inaccuracies in the approximated multipliers won’t degrade the result beyond an acceptable level,” adds Michal Piňos, providing further context. Significantly, in his work, Piňos is already designing networks to compute with approximated weights from the very beginning — in contrast to the previous approach, which added the approximation ad hoc to an existing neural network. To achieve this, the author uses a method based on the NAS algorithm, specifically its hardware-oriented variant (HW-NAS) — that is, automated design of neural network architectures that take the target hardware into account—which Piňos further extends to account for approximated multiplication as well. “Within this algorithm, I can use an evolutionary approach, a differentiable approach—there are many options. But I always design neural networks to account for the use of approximate multipliers.” The resulting networks differ in structure from conventional, manually designed architectures—the algorithm “discovers” their form on its own, aiming to best balance accuracy and energy consumption. To design networks in this way within a reasonable amount of time, Piňos also had to develop fast methods for estimating energy consumption: The design algorithm compares thousands of candidate networks, and accurately measuring each one on hardware would be unrealistically slow. Furthermore, experiments show that NS networks designed in this way offer a better trade-off between accuracy and energy consumption than networks approximated only after the fact. Piňos considers this to be the key finding of his dissertation.

When asked how he views the completion of his doctoral degree, Piňos is quite pragmatic. The topic of neural networks and evolutionary algorithms appealed to him when he was considering thesis topics — and when the offer came to continue his research at the doctoral level, he accepted it because, as he himself notes, he would have had a hard time finding another opportunity to work in such a specific area of IT. “Right now, though, I think that completing my dissertation also marks the end of my academic journey. I’ve been working full-time at NXP for several years now, and that’s the direction I’d like to continue in.” He is grateful for having completed his Ph.D.; he never found himself in a situation that would have forced him to consider giving it up. He particularly highlights the opportunity to participate in major international conferences on neural networks, such as the IJCNN conference held as part of the IEEE WCCI 2024 Congress in Yokohama, Japan. He also repeatedly mentions the people who supported him at FIT. Specifically, he names the long-standing guidance and assistance of Prof. Lukáš Sekanina. “I would like to express my sincere thanks to him and his family.”

We wish Michal Piňos a successful defense and the best of luck in his future career.