The energy efficiency of Artificial Intelligence (AI) systems is a crucial and actual issue that may have an important impact on an ecological, economic and technological level. Spiking Neural Networks (SNNs) are strongly suggested as valid candidates able to overcome Artificial Neural Networks (ANNs) in this specific contest. In this study, the proposal involves the review and comparison of energy consumption of the popular Artificial Neural Network architectures implemented on the CPU and GPU hardware compared with Spiking Neural Networks implemented in specialized memristive hardware and biological neural network human brain. As a result, the energy efficiency of Spiking Neural Networks can be indicated from 5 to 8 orders of magnitude. Some Spiking Neural Networks solutions are proposed including continuous feedback‐driven self‐learning approaches inspired by biological Spiking Neural Networks as well as pure memristive solutions for Spiking Neural Networks.

Annotated medical imaging data remain scarce, and labels are often weak and noisy: in mammography, an examination comprises several high-resolution views, yet the diagnostic outcome is recorded only at the breast level. Such problems are naturally cast as Multiple Instance Learning (MIL), where the model must infer instance-level structure from bag-level labels alone. Although contemporary foundation encoders supply strong general-purpose embeddings, augmenting MIL data in this representation space remains an open problem as established techniques act on one instance at a time and ignore the statistical dependencies binding a bag together. We address this with SetFlow, a generative model that learns the distribution of complete MIL bags directly in a frozen encoder’s embedding space. SetFlow couples flow-matching training with a Set Transformer-inspired backbone, making it invariant to instance ordering while modeling intra-bag relationships. Generation is conditioned jointly on class label and per-instance scale, yielding coherent, semantically faithful bags rather than isolated vectors. Evaluating on two large public mammography datasets and two encoders, we assess distributional fidelity, nearest-neighbor behavior, and downstream augmentation utility. We show that generated bags reproduce real-data statistics and improve classification in certain configuration, with performance gains varying on the amount of synthetic data. An architecture ablation confirms each design choice contributes to performance.

Machine learning is widely proposed as a basis for smart wastewater treatment plant (WWTP) management, yet reported performance is rarely benchmarked against trivial alternatives. This study uses six years of monitoring from a municipal WWTP in Serbia serving 18,000 population equivalents, comprising 1017 fully measured days. Total nitrogen removal falls from 79.4% outside summer to 67.4% in June to August (p = 4 × 10⁻¹⁸), in five of six years, while influent load, hydraulic flow and phosphorus removal are unchanged, which points to oxygen limitation. Four model families are benchmarked against persistence and a ten-measurement rolling mean on one chronological hold-out period. Predictive skill depends decisively on forecast horizon: at one step ahead nothing beats both baselines, whereas at five steps nitrogen reaches R² of 0.516 against 0.388 and 0.320. Chemical oxygen demand (COD) never beats a rolling mean. A random forest identifies nitrogen exceedance of 15 mg/L with an area under the curve of 0.90. Explanation of that model shows it recognises a breach already under way rather than one beginning, which sets the operating conditions for its use. Evaluating on interpolated rather than measured days raises apparent accuracy from R² of 0.34 to 0.82.