Advanced intelligent systems (Weinheim an der Bergstrasse, Germany)最新文献

Deep-Learning-Guided Surface-Enhanced Raman Spectroscopy

In article number 2400587, Muhammad Usman, Liang Wang, and co-workers present a novel approach combining deep-learning-guided surface-enhanced Raman spectroscopy (SERS) and a variational autoencoder (VAE) with a long short-term memory (LSTM) neural network to classify vaginal cleanliness levels rapidly and accurately. Enhanced spectral quality and an optimized VAE–LSTM model yielded an 85% accuracy on blind test data. This method, which improves signal-to-noise ratios and diagnostic efficiency, shows strong potential for clinical applications in assessing vaginal cleanliness through SERS analysis of vaginal secretions.

Decoupling Implantation Prediction and Embryo Ranking in Machine Learning

Itay Erlich, Assaf Zaritsky, and co-workers establish that optimizing a machine learning model to predict in vitro fertilization embryo implantation success by inclusion of clinical properties is not an optimal strategy for the task of embryo ranking (see article number 2400048). The reason for this is “shortcut learning”, the model relies on the clinical factor as a proxy for implantation – hampering its ability to approximate the embryo quality. The authors’ practical recommendation is to exclusively focus on the embryo intrinsic features for ranking.

Interpretable Machine Learning

Interpretable machine learning is essential for building trustworthy AI systems. Automated Semantically Interpretable AI (AS-XAI) extracts the common semantic feature space of diverse data samples and combines this feature space with a sensitivity analysis of neural networks in each semantic space to understand the networks’ decision-making processes. AS-XAI leverages the model’s understanding of common semantics in existing data to enable a wide range of fine-grained and scalable real-world applications. This approach allows for comprehensive semantic conceptual interpretations of out-of-distribution hybrids as well as species that are difficult for humans to recognize. See article number 2400359 by Changqi Sun, Hao Xu, Yuntian Chen, and Dongxiao Zhang.

Autonomous, cloud-based laboratories (CBLs) are transforming scientific research by democratizing access to advanced instruments that accelerate high-throughput discovery. As artificial intelligences (AIs) become integrated or “embodied” with CBLs and gain independence from human oversight, efforts to identify novel pharmaceuticals, renewable energies, and agricultural biotechnologies will accelerate. AI-driven CBLs can perform tasks more efficiently and accurately than human scientists at lower costs, achieving results in weeks rather than years. However, as AI systems approach or exceed human intelligence, their decision-making abilities could outpace the need for human input, raising ethical, economic, and safety concerns. Aligning AI goals with human values is critical, as unregulated systems could pose existential risks, including global health hazards or the distortion of knowledge-generating systems. AI-driven misinformation in research highlights the need for transparency and data integrity, which may be achieved by aligning incentivizes and engineered fail-safes to promote long-term human flourishing. To mitigate risks, strict compartmentalization of AI systems and CBLs with third-party supervision at fine temporal resolutions will be necessary. While current CBLs are piloted by humans, future AI systems may relegate humans to the role of co-pilot. Anticipating increased AI-CBL integration, policies must balance innovation with caution to maximize benefits and avoid unintended harm.

Touchscreens are essential parts of many electronics in daily lives of sighted people in the digital information era. On the other hand, visually impaired users rely on tactile displays as one of the key communication devices to interact with the digital world. However, due to their working mechanism and the uneven surface of tactile displays, one of the key features of screens for sighted users is surprisingly challenging to implement: precision touch input. To overcome this, a hand gesture recognition system is developed using a frequency-modulated continuous wave millimeter-wave radar. A multifeature encoder method is used to obtain the range and velocity information from the radar to translate the data into spectrogram images. Gesture recognition is implemented for common input gestures: single/double-click, swipe-right/left, scroll-up/down, zoom-in/out, and rotate-anticlockwise/clockwise. The gesture recognition and classification are based on machine learning, support vector machines, deep learning, and convolutional neural network approaches. The chosen model You-Only-Look-Once (YOLOv8) shows a high accuracy of 97.1% by iterating only 30 epochs with only 500 collected data samples per gesture. This research paves the way toward using radar sensors not only for tactile displays but also for other digital devices in human–computer interaction.

This study presents a novel rolled dielectric elastomer actuator, based on Poly(vinylidene fluoride-trifluoroethylene-chlorotrifluoroethylene), i.e., P(VDF-TrFE-CTFE), nanofibers that are fabricated by means of an electrospinning process. The soft actuator is realized by rolling a mat of two active layers, interleaved with two electrodes. The active layers are mats of P(VDF-TrFE-CTFE)-based nanofibers, embedded in an elastomeric matrix of polydimethylsiloxane (PDMS). The electrodes are made of a mixture of carbon black and PDMS, bonded to the active layers by electrostatic adhesion force. The effects of the nanofibers in the soft actuators are investigated. Specimens of the soft actuator are realized that differ in the thickness of the nanofibrous mats used for the active layers, that is, 60 and 120 μm. An electromechanical characterization is performed to analyze and measure the axial force and axial displacements of the soft actuators when different electric fields are applied to the specimens in the transversal direction. The experimental results show that the presence of P(VDF-TrFE-CTFE)-based nanofibers enhances the force-to-weight ratio of the soft actuators by up to 43.5%, and the energy density by up to 12.9%, compared to a control specimen with the active layer made of PDMS only.

Accurately identifying user needs in terms of assist timing and magnitude presents challenges for wearable power-assist limb devices. Traditional approaches to gait perception—such as estimating joint angles and walking conditions—often rely on electronic sensors and neural networks, which can compromise wearability and impose high computational demands. Physical reservoir computing (PRC), which utilizes the inherent nonlinearity of physical systems for data processing, offers a promising alternative. This study proposes a novel self-estimated physical reservoir computing (SEPRC) model that improves traditional PRC models for gait perception using a wearable pneumatic physical reservoir. A core feature of the new model is the self-estimation structure, wherein the outputs of the physical reservoir are mutually estimated. Experimental evaluations indicate that the SEPRC model outperforms traditional PRC in clustering time-series reservoir output sequences with the same dimensionality. This enhanced clustering performance is subsequently leveraged in gait perception by incorporating Takagi–Sugeno fuzzy logic for joint angle estimation and a softmax activation function for walking condition recognition. The newly proposed time-sequence processing approach facilitates the traditional PRC model to achieve higher accuracy in gait perception and greater robustness against the user's walking pattern variations while preserving PRC's hardware simplicity.

Printed electronics have traditionally used substrates and materials derived from fuel-based or less abundant and toxic resources, raising environmental concerns. Wood as a substrate reduces processing steps and enables the integration of intelligent functionalities in wooden furniture, offering biodegradability, nontoxicity, and derivation from renewable sources. In this work, sustainably printed transistors using zinc oxide nanoparticles as the active layer and honey electrolyte on wood substrates are demonstrated as a promising approach to reduce the environmental footprint of electronics. Despite the substrate's high roughness, the transistor exhibits excellent performance for screen-printed devices, with low on-voltage of 0.32 ± 0.12 V and high I_on/I_off of (2.4 ± 0.9) × 10⁴. Further analysis of hysteresis in transfer curves under varying scan rates and sweep ranges reveals the device's ability to adjust memory windows and on-current. Notably, these devices successfully emulate synapses, exhibiting neural facilitation and plasticity, indicating a shift toward sustainable computing. The device's dynamic response to single and successive presynaptic pulses demonstrates its ability to adjust synaptic weight, transition from transient to persistent memory, and pulse width-, frequency-, voltage-, and number-dependent excitatory postsynaptic currents. The successful emulation of the learning–forgetting–relearning–forgetting process underscores the device's potential for use in sustainable high-performance neuromorphic systems.

For energy- and time-efficient artificial intelligence (AI) computing, implementing hardware-based spiking neural networks (SNNs) has become a core technology. In SNNs, synaptic devices store weights in memory, and neurons process received weighted information and generate spike signals. Upon feeding spike signals into synaptic arrays, the synaptic weights multiply the signals, which subsequently sum up to perform vector-matrix multiplication (VMM). Simultaneous access to multiple synaptic devices, however, reduces the equivalent resistance of these synaptic arrays. This reduction alters the voltage division between the pre-synaptic array and the input resistance of the neuron circuit, distorting the read voltage across synaptic devices. This phenomenon is known as the fan-in problem, which leads to non-ideal VMM operations and degrades system accuracy. To address this issue, a novel compensated current mirror (CCM) neuron circuit is proposed, which incorporates a single additional transistor into a conventional current mirror. This CCM neuron achieves exceptional current linearity (R² > 0.999) and efficiently compensates for VMM error with low complexity and energy consumption (3.33 pJ spike⁻¹). Furthermore, the CCM neuron demonstrates ≈7-%p higher inference accuracy than conventional ones when integrated with a 512 × 512 large-scale synaptic array, which is comparable to the accuracy of software-based SNNs.

In the perspective of this article, the emergence of materials and systems for magneto-mechanical actuation in the field of mechanobiology is presented, and their potential to promote and advance biomedical research is discussed. These materials, ranging from single particles to compliant 2D substrates to 3D scaffolds, enable mechanical modulation of cells in a remote, dynamic, and reversible fashion. These features represent a major advance enabling researchers to reproduce time-evolving physiological and pathological processes in vitro and transmit mechanical forces and deformations to activate cellular responses or promote directed cell migration. As smart in vitro platforms, magneto-responsive systems may accelerate the discovery of mechanically mediated cellular mechanisms as therapeutic targets. In addition, the low magnetic susceptibility of biological tissues may facilitate the translation of in vitro approaches to in vivo settings, opening new routes for biomedical applications.