04534ntm a22005537i 4500 000722959 CZ-PrVSE 20260201122909.0 m d cr n|||||||||| 260201s2025 xr fsbm 000 0 eng d NEZPRACOVANÝ IMPORT ABA006 cze ABA006 ABA006 rda Liubchenko, Vladislav ISIS:170733 dis Analýza emergentních schopností v umělých a biologických neuronových sítích eng Analysis of Emergent Properties in Biological and Artificial Neural Networks / Vladislav Liubchenko 2025 ?? stran : digital, PDF soubor Vedoucí práce: Miroslav Vacura Diplomová práce (Ing.)—Vysoká škola ekonomická v Praze. Fakulta informatiky a statistiky, 2026 Obsahuje bibliografii Textový (vysokoškolská kvalifikační práce) Rok obhajoby 2026 Emergent behaviors—system-level complex properties that appear suddenly when neural networks surpass critical scales[2]—are a major problem for both AI and neuroscience. We focus here on the mechanisms, boundary conditions and functional analogies of emergent phenomena in artificial and biological neural networks. Author first presents a literature review on new capabilities discovered in large-scale language models (e.g., GPT-3, few-shot learning, scaling law thresholds, zero-shot sentiment analysis) and in biological systems (e.g., neuronal avalanches, self-organized criticality, spontaneous order). I proceed to build a conceptual framework, based on Anokhin's concept of the hypersystem and criticality models, which allows me to formalize how distributed interactions and feedback loops help produce qualitatively novel order. In practice, I experimentally study 3 versions of Meta’s OPT language models (350M, 1.3B and 2.7B parameters) with a replicable experimental protocol for the few-shot, syntax-parsing, and generative tasks. While I do see slight accuracy gains in the 2.7 B model over the 350 M baseline, it shows that emergent-like improvements come with much higher parameter counts. Without our own experiments using BNNs, I make use of comparative studies —gleaning from the literature in in vitro neural cultures and hybrid “DishBrain” systems —to find analogies between scale-dependent thresholds in ANNs and the transitions to the critical state in BNNs. Theoretical--experimental synthesis points out that artificial and biological networks operate through a combination of bottom-up parameter scaling (such as network size, excitation--inhibition balance) and top-down feedback (e.g., attention mechanisms, active inference), resulting in emergent function. I compare these to cortical structure and function—e.g., attention heads as cortical-column analogues—and discuss implications for understanding intelligence. Lastly, I point out open challenges on the mechanistic grounding, cross-modal birth, and the responsible regulation of unpredictable capabilities. Our results pave the way towards unifying theories of emergence and can guide the design of new neuro-inspired architectures. Způsob přístupu: Internet kognitivní informatika [obor dipl. práce] diplomové práce fd132022 czenas master's theses eczenas Artificial neural networks Biological neural networks Philosophy Emergent abilities Vacura, Miroslav mzk2006343245 ths Zamazal, Ondřej ISIS:7282 opn Vysoká škola ekonomická v Praze. Fakulta informatiky a statistiky kn20010709399 dgg https://insis.vse.cz/zp/88461/podrobnosti VŠKP v InSIS https://insis.vse.cz/zp/88461 Hlavní práce https://insis.vse.cz/zp/88461/posudek/vedouci Hodnocení vedoucího https://insis.vse.cz/zp/88461/posudek/oponent/86556 Oponentura https://insis.vse.cz/zp/88461/priloha/32881 Přiloha k práci https://insis.vse.cz/zp/88461/priloha/32882 Přiloha k práci https://insis.vse.cz/zp/88461/priloha/32883 Přiloha k práci https://insis.vse.cz/zp/88461/podrobnosti dc:identifier NEPOSILAT VSKP vse88461 260124 88461