In this project proposal, we plan to build novel multipurpose deep-learning systems, as descendants of advanced digital signal processing (DSP) methods. The novelty is based on their specific structure, which ensures desired directivity in the learning process. On the one hand, the idea is inspired by physics-informed neural networks (NNs). The laws of physics in such networks are incorporated in the cost criteria. The prior knowledge of general physical laws acts in the training of NNs as a regularization agent, limiting the space of admissible solutions and increasing the correctness of the objective function approximation. On the other hand, we found our inspiration in adaptive systems designed by the authors of this proposal more than a decade ago: the fixed part of such systems provides for retaining the desired properties, while the variable parameter part provides for adaptation. Remapping the same idea into the new field of deep learning will bring new and exciting results. The structure of such deep-learning networks will ensure that the solution remains in the frame of an acceptable mathematical solution or given limitations in physics, chemistry, or biology. For reference, we call it science-informed deep learning. 

Among others, one important application will be in the field of sampling (“new sampling paradigm”). Instead of function dictionaries, deep-learned networks that best fit the measurement will be used to model the reconstructed objects. Next, it will be used to create behavioral models of dynamic systems, which is an important element of autonomous systems. In different applications, it leads to solving ordinary or partial differential equations. Furthermore, the results of this project can be extended to almost any known application of DSP. It will open a completely new branch of intelligent processing systems.

Opposing traditional programming that was based on lists of precise human-authored instructions, the deep learning mantra is that everything can be learned from data: no instructions needed. Still, well-know instruction-based algorithms do the majority of data processing jobs, especially when the predictability of the results is crucial. Our life experience tells us that some things can be learned, but some must be given in advance to enable efficient learning. Many technical systems are inspired by nature. Deep learning systems and neural networks are not an exception. On the one hand, mammal brains consist of a large mass of general purpose neurons, in which damaged tissue can be replaced by the healthy one. It corresponds to the mentioned deep learning mantra: learning from data. On the other hand, specialized groups of neurons take care of hearing, image processing, keeping balance, even some logical operations. Moreover, control of life support functions, such as body temperature, heart beats or breathing rely on hard-wired and very specialized neuron groups.

In this project, we focused our efforts to embed the knowledge into the structure of the deep learning systems, rather than the cost criterion. It will bring predictability and interpretability of the results and speed up the learning process.

prof. dr. sc. Damir Seršić project leader full professor,  Department of Electronic Systems and Information Processing e-mail: damir.sersic@fer.hr
prof. dr. sc. Davor Petrinović full professor,  Department of Electronic Systems and Information Processing e-mail: davor.petrinovic@fer.hr
izv. prof. dr. sc. Ana Sović Kržić associate professor,  Department of Electronic Systems and Information Processing e-mail: ana.sovic@fer.hr
izv. prof. dr. sc. Azra Tafro associate professor,  University of Zagreb Faculty of Forestry e-mail: azra.tafro@fer.hr
dr. sc. Ivan Ralašić external associate,  CTO of Forsight e-mail: ivan.ralasic@fer.hr
dr. sc. Tin Vlašić  external associate, Postdoctoral Research Fellow at  A*STAR Singapore e-mail: tin.vlasic@fer.hr
Tomislav Matulić, mag. ing. assistant,  Department of Electronic Systems and Information Processing e-mail: tomislav.matulic@fer.hr
Ivan Bukić, mag. ing.  assistant,  Department of Electronic Systems and Information Processing e-mail: ivan.bukic@fer.hr

Journal articles

Other