Clinically, there is an urgent need to identify new therapeutic strategies for selectively treating cancer cells. One of the directions in this research is the development of biocompatible therapeutics that selectively target cancer cells. Here, we show that novel aminated graphene oxide (haGO-NH₂) nanoparticles demonstrate increased toxicity towards human hepatocellular cancer cells compared to pristine graphene oxide(GO). The applied novel strategy for amination leads to a decrease in the size of haGO-NH₂ and their zeta potential, thus, assuring easier penetration through the cell membrane. After characterization of the biological activities of pristine and aminated GO, we have demonstrated strong cytotoxicity of haGO-NH₂ toward hepatic cancer cells—HepG2 cell line, in a dose-dependent manner. We have presented evidence that the cytotoxic effects of haGO-NH₂ on hepatic cancer cells were due to cell membrane damage, mitochondrial dysfunction and increased reactive oxygen species (ROS) production. Intrinsically, our current study provides new rationale for exploiting aminated graphene oxide as an anticancer therapeutic.

The process of training of Artificial Neural Networks essentially is optimization of the values of the weights $ w_{pq} $ associated with the arcs, connecting the nodes of the layers. This is a process of minimization of the Loss function (maximization of Accuracy function). During the training, the training data set recursively is utilized at subsequent stages, called \textit{Epochs}. The training continues until a satisfactory values of the Loss, Accuracy etc. parameters are reached. The matrices $ W^{UV} $ comprising the weights of the arcs connecting the layers $ U $ and $ V $, can be regarded as gray-scale images of a surface. Starting as random matrices, processed by recursive procedures, they gradually become fractal structures, characterized with respective fractal dimension $ D_f $. In the presented article we have made an attempt to utilize the correspondence of $ D_f $ with the Loss/Accuracy values, in order to forecast the optimal ending point of the NN training process. Similar conclusions were made for the correspondence between the number of layer’s nodes and $ D_f $. An attempt to apply statistically more rigorous approach in the determination of the slope of the regression line in Richardson-Mandelbrot plot, was made.

We present an information-theoretic framework that models creative reasoning as structured, task-directed motion within hierarchically organized epistemic structures. Creative problem solving is described as a progressive reconfiguration of an initial epistemic state toward a desired target configuration through a sequence of intermediate representations. Each representation is modeled by an empirical structural distribution over extracted features, enabling two complementary quantitative diagnostics: (i) a divergence measure—the Jensen–Shannon (JS) metric—capturing structural departure, novelty, and analogical proximity; and (ii) energy-based plausibility measures expressing conformity to dominant structural regularities and agent-relative constraints. Their interaction induces a geometry in which exploration balances novelty against structural admissibility, and cross-domain transfer is enabled through alignment of compatible representations. We introduce algebraic and probabilistic principles governing the generation, evaluation, and selection of candidate representations, including neighborhood-restricted exploration, history-sensitive evaluation, and non-redundant comparison under progressively refined interpretive conditions. The framework is operationalized at the level of the epistemic structures accessible to an individual reasoning agent, while large language models are interpreted as mechanisms that facilitate access to broader reservoirs of structured knowledge. Although a musical case study (J. S. Bach’s The Art of Fugue) is used for illustration, the proposed framework is domain-general and applies to any setting involving structured representations and lawful transformations. The resulting formalism supports principled approaches to task-oriented creative search, analogical reasoning, and autonomous knowledge exploration, with potential implications for machine-assisted discovery and structurally grounded communication across intelligent systems.

Large language models (LLMs) now form a regular part of scientific research practice, where they are used to assist with hypothesis formulation, literature synthesis, and various forms of formal reasoning. Their use builds on earlier ideas of \emph{delegated cognition} and brings into sharper focus questions about how epistemic agency and moral responsibility are distributed across human--machine arrangements. This paper develops a conceptual and formal framework for examining these hybrid modes of reasoning, drawing on an analogy with familiar academic hierarchies in which a principal investigator (PI) coordinates and supervises junior collaborators. Within this framework, three related operators are distinguished: \emph{verification} $V(g)$, which concerns logical consistency and empirical adequacy; \emph{responsibility mapping} $R(g)$, which assigns epistemic and moral accountability to human agents; and \emph{epistemic value} $E(g)$, which characterizes the justificatory status and cognitive standing of a result, regardless of whether it is produced by a human or an artificial system. Verification and moral authorship are treated as closely connected aspects of epistemic responsibility, in the sense that verifying a claim amounts to accepting responsibility for its truth. On this view, the ethical boundary in scientific research is not drawn between human and machine reasoning, but between \emph{responsible} and \emph{negligent} forms of delegation within a distributed cognitive system. The paper also introduces the notion of an \emph{epistemic audit} as an institutional mechanism, comparable to established quality-assurance practices, for documenting transparency, reproducibility, and coherence in AI-assisted research. The analysis contributes to ongoing discussions in cognitive epistemology and the philosophy of AI concerning authorship, verification, and responsibility in extended systems of scientific reasoning.

We present cEntMax, an information-theoretic framework for training set optimization that selects classwise informative samples via cross-entropy divergence to prototype pivots. Under a noisy-channel generative view and local linearity of deep networks, the method connects predictive entropy, Fisher information, and G-optimal coverage. Experiments on EMNIST and KMNIST show faster convergence, lower validation loss, and greater stability than random sampling, especially for moderate sampling fractions.