Explore the vast range of titles available.

Modern large language models (LLMs) achieve state-of-the-art performance through architectural advancements but require high computational costs for inference. Post-training quantization is a widely adopted approach to reduce these costs by quantizing weights and activations to lower precision, such as INT8. However, we identify a critical challenge in activation quantization for GLU (Gated Linear Unit) variants, which are commonly used in the feed-forward networks of modern LLMs like the LLaMA family. Specifically, severe local quantization errors arise due to excessively large activation magnitudes, which we refer to as activation spikes, leading to significant degradation in model performance. Our analysis reveals a systematic pattern of these spikes: they predominantly occur in the FFN (feed-forward network) layers at the early and late layers of the model and are concentrated on a small subset of tokens rather than being uniformly distributed across a token sequence. To mitigate this issue, we propose two empirical methods: Quantization-free Module (QFeM) and Quantization-free Prefix (QFeP), which isolate activation spikes during quantization. Extensive experiments demonstrated that our methods effectively improve activation quantization, particularly in coarse-grained quantization schemes, enhancing the performance of LLMs with GLU variants and addressing the limitations of existing quantization techniques. The code for implementing our methods and reproducing the experiments is publicly available our GitHub repository.

Noradrenergic neurons in the locus coeruleus (LC) release noradrenaline that acts via volume transmission to activate extrasynaptic G-protein coupled receptors (GPCRs) in target cells throughout the brain. As the closest projection, the dorsal LC laterally adjoins the mesencephalic trigeminal nucleus (MTN), in which proprioceptive primary sensory neurons innervating muscle spindles of jaw-closing muscles are exceptionally located. MTN neurons express α2-adrenergic receptors (α2-ARs) and display hyperpolarization-activated cyclic nucleotide-gated (HCN) current (Ih), which is downregulated by α2-AR activation. To quantify the activity-dependent outcome of volume transmission of noradrenaline (NA) from LC to MTN, we investigated how direct LC activation inhibits Ih in MTN neurons by performing dual whole-cell recordings from LC and MTN neurons. Repetition of 20 Hz spike-train evoked with 1-sec current-pulse in LC neurons every 30 sec resulted in a gradual decrease in Ih evoked every 30 sec, revealing a Hill-type relationship between the number of spike-trains in LC neurons and the degree of Ih inhibition in MTN neurons. On the other hand, when microstimulation was applied in LC every 30 sec, an LC neuron repeatedly displayed a transient higher-frequency firing followed by a tonic firing at 5–10 Hz for 30 sec. This subsequently caused a similar Hill-type inhibition of Ih in the simultaneously recorded MTN neuron but with a smaller Hill coefficient, suggesting a lower signal transduction efficacy. In contrast, 20 Hz activity induced by a 1-sec pulse applied every 5–10 sec caused only a transient facilitation of Ih inhibition followed by a forced termination of Ih inhibition. Thus, the three modes of LC activities modulated the volume transmission to activate α2-adrenergic GPCR to differentially inhibit Ih in MTN neurons.

Adaptive and extreme changes in shape and configuration are the functional and morphological uniqueness of soft robots, but existing design approaches still rely on the predefined coordination of their “muscle” and “nerve” functions to produce such behaviors. Herein, a strategy is introduced for building modular soft machines that can be innervated in ways that conform to their body extension or shape changes, based on modular soft electronics. The development of soft electronic adhesive interlocking (SEAL) technology allows for instant, robust, and repeatable integration of soft electronic modules that can “innervate” and activate modular soft actuators and machines in a reconfigurable manner. Demonstrations of soft robotic tentacles and their grasping capability show that the robot function can be adapted to or reconfigured within the body with a length extended more than 10 times. The modular strategy presented herein can offer a unique promise to build up future robots with dynamic, reconfigurable functions. Soft electronic adhesive interlocking facilitates reconfigurable electronic innervation of modular soft machines by providing robust Van der Waals adhesion, mechanical interlocking, and high conductivity. This mechanism allows instant, repeatable integration of soft electronic modules demonstrated through modularized soft robots that adaptively and selectively innervate electronic modules for effective body shape changes.

Augmented reality (AR) has received much attention in the cultural heritage domain as an interactive medium for requesting and accessing information regarding heritage sites. In this study, we developed a mobile AR system based on Semantic Web technology to provide contextual information about cultural heritage sites. Most location-based AR systems are designed to present simple information about a point of interest (POI), but the proposed system offers information related to various aspects of cultural heritage, both tangible and intangible, linked to the POI. This is achieved via an information modeling framework where a cultural heritage ontology is used to aggregate heterogeneous data and semantically connect them with each other. We extracted cultural heritage data from five web databases and modeled contextual information for a target heritage site (Injeongjeon Hall and its vicinity in Changdeokgung Palace in South Korea) using the selected ontology. We then implemented a mobile AR application and conducted a user study to assess the learning and engagement impacts of the proposed system. We found that the application provides an agreeable user experience in terms of its affective, cognitive, and operative features. The results of our analysis showed that specific usage patterns were significant with regard to learning outcomes. Finally, we explored how the study’s key findings can provide practical design guidance for system designers to enhance mobile AR information systems for heritage sites, and to show system designers how to support particular usage patterns in order to accommodate specific user experiences better.

Application of cold or cold-mimicking chemicals to injury has long been recognized as an effective means of pain relief and is widely utilized in daily life. However, underlying neural mechanisms remain elusive. Here, we identified a cold-responsive neuronal subset within lateral parabrachial nucleus (lPBN), the thermosensory relay region in the hindbrain, that mediates cold-induced analgesia in mice. Selective activation of these neurons and their projection to ventrolateral periaqueductal gray (vlPAG) increased nociceptive threshold via opioid receptor signaling in vlPAG. Conversely, ablation of these neurons attenuated analgesia induced by cold-mimicking chemicals. We further identified that these neurons express precursor gene of enkephalin, which is released into vlPAG for pain relief. Activation of cold-responsive neurons in descending pain modulation circuitry reduced spinal cord responses to noxious stimuli, suggesting the involvement of top-down pain modulation pathway. These findings propose a central mechanism underlying cold-induced pain relief, which could be a novel therapeutic target. Cold-sensitive lPBN neurons release enkephalin into vlPAG for top-down analgesic action during cold-induced pain relief

Modern large language models (LLMs) have established state-of-the-art performance through architectural improvements, but still require significant computational cost for inference. In an effort to reduce the inference cost, post-training quantization (PTQ) has become a popular approach, quantizing weights and activations to lower precision, such as INT8. In this paper, we reveal the challenges of activation quantization in GLU variants, which are widely used in feed-forward network (FFN) of modern LLMs, such as LLaMA family. The problem is that severe local quantization errors, caused by excessive magnitudes of activation in GLU variants, significantly degrade the performance of the quantized LLM. We denote these activations as activation spikes. Our further observations provide a systematic pattern of activation spikes: 1) The activation spikes occur in the FFN of specific layers, particularly in the early and late layers, 2) The activation spikes are dedicated to a couple of tokens, rather than being shared across a sequence. Based on our observations, we propose two empirical methods, Quantization-free Module (QFeM) and Quantization-free Prefix (QFeP), to isolate the activation spikes during quantization. Our extensive experiments validate the effectiveness of the proposed methods for the activation quantization, especially with coarse-grained scheme, of latest LLMs with GLU variants, including LLaMA-2/3, Mistral, Mixtral, SOLAR, and Gemma. In particular, our methods enhance the current alleviation techniques (e.g., SmoothQuant) that fail to control the activation spikes. Code is available at https://github.com/onnoo/activation-spikes.

A clock distribution network (CDN) is a ubiquitous on-chip element that provides synchronized clock signals to all different circuit blocks in the chip. To maximize the chip performance, today’s CDN demands lower jitter, skew, and heat dissipation. Conventionally, on-chip clock signals have been distributed in the electric voltage domain, resulting in increased jitter, skew, and heat dissipation due to clock drivers. While low-jitter optical pulses have been locally injected in the chip, research on effective distribution of such high-quality clock signals has been relatively sparse. Here, we demonstrate femtosecond-precision distribution of electronic clocks using driver-less CDNs injected by photocurrent pulses extracted from an optical frequency comb source. Femtosecond-level on-chip jitter and skew can be achieved for gigahertz-rate clocking of CMOS chips by combining ultralow comb-jitter, multiple driver-less metal-meshes, and active skew control. This work shows the potential of optical frequency combs for distributing high-quality clock signals inside high-performance integrated circuits, including 3D integrated circuits. A clock distribution network (CDN) is a ubiquitous element that provides synchronized clocks to all circuit blocks in the chip. Here the authors demonstrate femtosecond-precision distribution of clocks in CMOS chips using optical frequency combs and driverless CDNs.

There has been remarkable progress in generating ultralow-noise microwaves from optical frequency combs in the last decade. While a combination of techniques has enabled tens to hundreds of attoseconds residual jitter in microwave extraction, so far most of research efforts have been focused on extracting single-tone microwaves from combs; there has been no study on the noise properties of photocurrent pulses directly extracted from the photodiode. Here, we reveal that the residual jitter between optical pulses and rising edges of photocurrent pulses can be in the tens of attoseconds regime. The rising-edge jitter is much lower than the falling-edge jitter, and further, this ultralow rising-edge jitter could be obtained by both p-i-n and (modified-)uni-travelling-carrier photodiodes. This finding can be directly used for various edge-sensitive timing applications, and further shows the potential for ultrahigh-precision timing using silicon-photonic-integrable on-chip p-i-n photodiodes. For edge-sensitive timing applications, the edge jitter of electrical pulses is important. Here, the authors report on very low rising edge jitter extracted from an optical frequency comb and explore the best condition for low jitter by minimizing the amplitude-to-timing conversion in photodiodes.

IntroductionThis systematic review and meta-analysis aimed to evaluate the efficacy and safety of traditional herbal medicine (THM) in improving opioid-induced constipation (OIC) in patients with cancer.MethodsTo identify randomized controlled trials (RCTs) evaluating orally administered THM for OIC in patients with cancer, a comprehensive search of seven databases was conducted from inception to 29 August 2024. The primary outcome was improvement in OIC, which was assessed using the total effective rate (TER). Secondary outcomes included stool form, difficulty of defecation, defecation time, and the Karnofsky performance scale (KPS). The methodological quality of the included studies was assessed using the Cochrane Risk of Bias tool, and the certainty of evidence was evaluated according to the Grading of Recommendations Assessment, Development, and Evaluation method.ResultsIn total, 21 RCTs involving 2,108 patients were included. Compared to conventional medicine, THM significantly improved OIC as measured by TER [risk ratio (RR) 1.21, 95% confidence intervals (CIs) 1.14–1.25], with high certainty. THM showed a significant improvement in stool form [mean difference (MD) −0.16, 95% CIs −0.43–0.10; very low certainty], difficulty of defecation [MD -0.31, 95% CIs −0.49 to −0.13; low certainty], defecation time [MD -0.28, 95% CIs −0.45 to −0.10; moderate certainty], and KPS measured by mean changes in scores [MD 6.76, 95% CIs 4.32–9.20; low certainty]. Adverse events were mainly gastrointestinal symptoms such as diarrhea, nausea, and abdominal pain, but such events were not serious.ConclusionThe findings of this systematic review indicate that THM may be considered a safe and potentially alternative option for improving OIC in patients with cancer. However, more robust and high-quality RCTs are required to strengthen this evidence.Systematic Review Registrationhttps://www.crd.york.ac.uk/prospero, Identifier: CRD42024557773.

This systematic review and meta-analysis evaluated the preventive efficacy and safety of orally-administered traditional herbal medicine (THM) for the management of chemotherapy-induced peripheral neuropathy (CIPN) in patients with cancer. Randomized controlled trials (RCTs) evaluating the efficacy of orally-administered THM in the prevention of CIPN published up to 30 April 2024 were retrieved from nine databases. The primary outcome was the incidence of CIPN, and the secondary outcomes included changes in neuropathic pain intensity, nerve conduction study parameters, Karnofsky Performance Scale (KPS) scores, and the incidence of adverse events. The quality of the studies and the strength of the evidence were evaluated using the Cochrane Risk of Bias Assessment Tool and Grading of Recommendations Assessment, Development, and Evaluation (GRADE) method. Key herbal combinations were identified by conducting an association rule analysis. Thirty-seven RCTs involving 2,882 patients were included. Significant differences were observed between THM and the placebo [RR 0.83, 95% CI 0.74-0.93, p < 0.05; low quality of evidence], usual care [RR 0.51, 95% CI 0.37-0.69, p < 0.05; moderate quality of evidence], and no treatment [RR 0.62, 95% CI 0.54-0.71, p < 0.05; moderate quality of evidence] in terms of in the incidence rate of CIPN. A significant reduction in the intensity of neuropathic pain [SMD -0.81, 95% CI -1.07 to -0.56, p < 0.05; high quality of evidence] and a significant improvement in KPS [MD 8.18, p < 0.05; low quality of evidence] were observed in the THM compared to no treatment. Furthermore, compared with usual care and no treatment, the use of THM yielded a significant improvement in the nerve conduction parameters with low quality of evidence. No serious adverse events were reported. The combination of Astragali Radix and Cinnamomi Ramulus as the strongest herbal combination used for the prevention of CIPN. THM may be a promising option for the prevention of CIPN in patients with cancer. Low certainty of evidence, and substantial heterogeneity and risk of bias can limit the strength of the conclusions. Further well-designed and rigorously reported randomized controlled trials are warranted to confirm these findings and clarify their clinical applicability. https://www.crd.york.ac.uk/prospero, Identifier: CRD42021270942.