Search for: All records

Transformer models have been widely investigated in different domains by providing long-range dependency handling and global contextual awareness, driving the development of popular AI applications such as ChatGPT, Gemini, and Alexa. State Space Models (SSMs) have emerged as strong contenders in the field of sequential modeling, challenging the dominance of Transformers. SSMs incorporate a selective mechanism that allows for dynamic parameter adjustment based on input data, enhancing their performance. However, this mechanism also comes with increasing computational complexity and bandwidth demands, posing challenges for deployment on resource-constraint mobile devices. To address these challenges without sacrificing the accuracy of the selective mechanism, we propose a sparse learning framework that integrates architecture-aware compiler optimizations. We introduce an end-to-end solution–C 4 n kernel sparsity, which prunes n elements from every four contiguous weights, and develop a compiler-based acceleration solution to ensure execution efficiency for this sparsity on mobile devices. Based on the kernel sparsity, our framework generates optimized sparse models targeting specific sparsity or latency requirements for various model sizes. We further leverage pruned weights to compensate for the remaining weights, enhancing downstream task performance. For practical hardware acceleration, we propose C 4 n -specific optimizations combined with a layout transformation elimination strategy. This approach mitigates inefficiencies arising from fine-grained pruning in linear layers and improves performance across other operations. Experimental results demonstrate that our method achieves superior task performance compared to other semi-structured pruning methods and achieves up-to 7→ speedup compared to llama.cpp framework on mobile devices. 

We propose a novel VQA dataset, BloomVQA, to facilitate comprehensive evaluation of large vision-language models on comprehension tasks. Unlike current benchmarks that often focus on fact-based memorization and simple reasoning tasks without theoretical grounding, we collect multiple-choice samples based on picture stories that reflect different levels of comprehension, as laid out in Bloom's Taxonomy, a classic framework for learning assessment widely adopted in education research. Our data maps to a novel hierarchical graph representation which enables automatic data augmentation and novel measures characterizing model consistency. We perform graded evaluation and reliability analysis on recent multi-modal models. In comparison to low-level tasks, we observe decreased performance on tasks requiring advanced comprehension and cognitive skills with up to 38.0\% drop in VQA accuracy. In comparison to earlier models, GPT-4V demonstrates improved accuracy over all comprehension levels and shows a tendency of bypassing visual inputs especially for higher-level tasks. Current models also show consistency patterns misaligned with human comprehension in various scenarios, demonstrating the need for improvement based on theoretically-grounded criteria. 

We study current injection tensors of Quantum interference control (QuIC) processes by measuring their dependences on single-polarization rotation. We show that the direction of QuIC photocurrents is determined by the polarization of odd-order optical absorptions. 

Creep strength in polycrystalline Ni-based superalloys is influenced by the formation of a rich variety of planar faults forming within the strengthening γ' phase. The lengthening and thickening rate of these faults – and therefore the creep rate – depends on an intriguing combination of dislocation interactions at the γ/γ' interface and diffusional processes of the alloying elements at the core of the fault tip. The effect of alloy composition on this process is not fully understood. In this work we use correlative high resolution transmission electron microscopy and energy-dispersive X-ray spectroscopy to study the deformation faults in two different Ni-based superalloys with carefully designed ratios of disordering-to-ordering-promoting elements (Co-Cr against Nb-Ta-Ti). The results show that the additions of ordering-promoting elements reduce the diffusional processes required for the faults to lengthen and thicken thus reducing the creep rates found for the higher Nb-Ta-Ti alloy. These insights provide a path to follow in the design of improved grades of creep-resistant polycrystalline alloys beyond 700 °C. 

Through direction numerical simulation (DNS) of a model manta ray body, pectoral fin scaled pitching effect on hydrodynamic performance and wake is investigated. The manta ray model is derived from high-speed video of manta ray swimming with motion of the model prescribed to match the actual manta ray. Rotation angles of the model skeletal joints is altered to scale the pitching. This results in four manta ray models with different pectoral fin pitching ratios. The models are simulated using an in-house developed immersed boundary method-based numerical solver. Notable discrepancies in thrust production during the downstroke are observed, with the θ =1.0 case producing instantaneous thrust peak that is 19% higher than the θ =0.72 model. Cycle averaged thrust is highest for the θ =0.72 model case, however, which can be attributed to extended reverse thrust for the θ =1.0 model. Through analysis of the near-body wake structures produced during the downstroke, late leading-edge vortex (LEV) formation is discovered to be primarily responsible for the detrimental reverse thrust seen for the θR =1.0 model. Surface pressure contours confirm this finding. Meanwhile the upstroke possesses less pronounced force production. 

Flying snakes are the only snakes on Earth capable of aerial gliding, taking advantage of fluid dynamic principles to leap from point to point among the trees. During their gliding, the locomotion of aerial undulation is observed. We hypothesize that this locomotion and its associated unsteady vortex dynamics are critical to their aerodynamic performance. However, there is a lack of detailed three-dimensional flow field information around the snake body in gliding due to the difficulties in experimental flow visualizations of live animals. In this study, a computation fluid dynamics (CFD) study has been conducted to study the fluid dynamics of a snake-like gliding. A mathematical equation describing the horizontal undulation motion was applied for constructing snake-like 3D computational models and a series of flow simulations were conducted. An immersed-boundary-method (IBM)-based direct numerical simulation (DNS) flow solver along with adaptive mesh refinement (AMR) was used in the simulation. Specifically, different head positions, corresponding to different horizontal wave shapes and their effect on aerodynamic performance, flow field and wake structures behind the body will be studied. In addition, the dynamic undulating motion is introduced in the model and a CFD simulation is also conducted. Results from this study are expected to bring a step stone to understanding snake-inspired locomotion. 

Through computational fluid dynamics (CFD) simulations of a model manta ray body, the hydrodynamic role of manta-like bioinspired flapping is investigated. The manta ray model motion is reconstructed from synchronized high-resolution videos of manta ray swimming. Rotation angles of the model skeletal joints are altered to scale the pitching and bending, resulting in eight models with different pectoral fin pitching and bending ratios. Simulations are performed using an in-house developed immersed boundary method-based numerical solver. Pectoral fin pitching ratio (PR) is found to have significant implications in the thrust and efficiency of the manta model. This occurs due to more optimal vortex formation and shedding caused by the lower pitching ratio. Leading edge vortexes (LEVs) formed on the bottom of the fin, a characteristic of the higher PR cases, produced parasitic low pressure that hinders thrust force. Lowering the PR reduces the influence of this vortex while another LEV that forms on the top surface of the fin strengthens it. A moderately high bending ratio (BR) can slightly reduce power consumption. Finally, by combining a moderately high BR = 0.83 with PR = 0.67, further performance improvements can be made. This enhanced understanding of manta-inspired propulsive mechanics fills a gap in our understanding of the manta-like mobuliform locomotion. This motivates a new generation of manta-inspired robots that can mimic the high speed and efficiency of their biological counterpart. 

Context. Atomic gas in the diffuse interstellar medium (ISM) is organized in filamentary structures. These structures usually host cold and dense molecular clumps. The Galactic magnetic field is considered to play an important role in the formation of these clumps. Aims. Our goal is to explore the role of the magnetic field in the H I -H 2 transition process. Methods. We targeted a diffuse ISM filamentary cloud toward the Ursa Major cirrus where gas transitions from atomic to molecular. We probed the magnetic field properties of the cloud with optical polarization observations. We performed multiwavelength spectroscopic observations of different species in order to probe the gas phase properties of the cloud. We observed the CO ( J = 1−0) and ( J = 2−1) lines in order to probe the molecular content of the cloud. We also obtained observations of the [C ii ] 157.6 µ m emission line in order to trace the CO-dark H 2 gas and estimate the mean volume density of the cloud. Results. We identified two distinct subregions within the cloud. One of the regions is mostly atomic, while the other is dominated by molecular gas, although most of it is CO-dark. The estimated plane-of-the-sky magnetic field strength between the two regions remains constant within uncertainties and lies in the range 13–30 µG. The total magnetic field strength does not scale with density. This implies that gas is compressed along the field lines. We also found that turbulence is trans-Alfvénic, with M A ≈ 1. In the molecular region, we detected an asymmetric CO clump whose minor axis is closer, with a 24° deviation, to the mean magnetic field orientation than the angle of its major axis. The H i velocity gradients are in general perpendicular to the mean magnetic field orientation except for the region close to the CO clump, where they tend to become parallel. This phenomenon is likely related to gas undergoing gravitational infall. The magnetic field morphology of the target cloud is parallel to the H i column density structure of the cloud in the atomic region, while it tends to become perpendicular to the H i structure in the molecular region. On the other hand, the magnetic field morphology seems to form a smaller offset angle with the total column density shape (including both atomic and molecular gas) of this transition cloud. Conclusions. In the target cloud where the H i –H 2 transition takes place, turbulence is trans-Alfvénic, and hence the magnetic field plays an important role in the cloud dynamics. Atomic gas probably accumulates preferentially along the magnetic field lines and creates overdensities where molecular gas can form. The magnetic field morphology is probed better by the total column density shape of the cloud, and not its H i column density shape.