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 report the discovery of a novel form of Ruddlesden−Popper (RP) nickelate that stands as the first example of long-range, coherent polymorphism in this class of inorganic solids. Rather than the well-known, uniform stacking of perovskite blocks ubiquitously found in RP phases, this newly discovered polymorph of the bilayer RP phase La3Ni2O7 adopts a novel stacking sequence in which single-layer and trilayer blocks of NiO6 octahedra alternate in a “1313” sequence. Crystals of this new polymorph are described in space group Cmmm, although we note evidence for a competing Imam variant. Transport measurements at ambient pressure reveal metallic character with evidence of a charge density wave transition with an onset at T ≈ 134 K. The discovery of such polymorphism could reverberate to the expansive range of science and applications that rely on RP materials, particularly the recently reported signatures of superconductivity in bilayer La3Ni2O7 with Tc as high as 80 K above 14 GPa. 

Abstract The meltwater streams of the McMurdo Dry Valleys are hot spots of biological diversity in the climate-sensitive polar desert landscape. Microbial mats, largely comprised of cyanobacteria, dominate the streams which flow for a brief window of time (~10 weeks) over the austral summer. These communities, critical to nutrient and carbon cycling, display previously uncharacterized patterns of rapid destabilization and recovery upon exposure to variable and physiologically detrimental conditions. Here, we characterize changes in biodiversity, transcriptional responses and activity of microbial mats in response to hydrological disturbance over spatiotemporal gradients. While diverse metabolic strategies persist between marginal mats and main channel mats, data collected from 4 time points during the austral summer revealed a homogenization of the mat communities during the mid-season peak meltwater flow, directly influencing the biogeochemical roles of this stream ecosystem. Gene expression pattern analyses identified strong functional sensitivities of nitrogen-fixing marginal mats to changes in hydrological activities. Stress response markers detailed the environmental challenges of each microhabitat and the molecular mechanisms underpinning survival in a polar desert ecosystem at the forefront of climate change. At mid and end points in the flow cycle, mobile genetic elements were upregulated across all mat types indicating high degrees of genome evolvability and transcriptional synchronies. Additionally, we identified novel antifreeze activity in the stream microbial mats indicating the presence of ice-binding proteins (IBPs). Cumulatively, these data provide a new view of active intra-stream diversity, biotic interactions and alterations in ecosystem function over a high-flow hydrological regime. 

The recent prediction that honeycomb lattices of Co2+ (3d7) ions could host dominant Kitaev interactions provides an exciting direction for exploration of new routes to stabilizing Kitaev’s quantum spin liquid in real materials. Na3Co2SbO6 has been singled out as a potential material candidate provided that spin and orbital moments couple into a Jeff = 1/2 ground state, and that the relative strength of trigonal crystal field and spin-orbit coupling acting on Co ions can be tailored. Using x-ray linear dichroism (XLD) and x-ray magnetic circular dichroism (XMCD) experiments, alongside configuration interaction calculations, we confirm the counterintuitive positive sign of the trigonal crystal field acting on Co2+ ions and test the validity of the Jeff = 1/2 description of the electronic ground state. The results lend experimental support to recent theoretical predictions that a compression (elongation) of CoO6 octahedra along (perpendicular to) the trigonal axis would drive this cobaltate toward the Kitaev limit, assuming the Jeff = 1/2 character of the electronic ground state is preserved. 

Abstract Understanding the brain requires understanding neurons’ functional responses to the circuit architecture shaping them. Here we introduce the MICrONS functional connectomics dataset with dense calcium imaging of around 75,000 neurons in primary visual cortex (VISp) and higher visual areas (VISrl, VISal and VISlm) in an awake mouse that is viewing natural and synthetic stimuli. These data are co-registered with an electron microscopy reconstruction containing more than 200,000 cells and 0.5 billion synapses. Proofreading of a subset of neurons yielded reconstructions that include complete dendritic trees as well the local and inter-areal axonal projections that map up to thousands of cell-to-cell connections per neuron. Released as an open-access resource, this dataset includes the tools for data retrieval and analysis^1,2. Accompanying studies describe its use for comprehensive characterization of cell types^3–6, a synaptic level connectivity diagram of a cortical column⁴, and uncovering cell-type-specific inhibitory connectivity that can be linked to gene expression data^4,7. Functionally, we identify new computational principles of how information is integrated across visual space⁸, characterize novel types of neuronal invariances⁹and bring structure and function together to uncover a general principle for connectivity between excitatory neurons within and across areas^10,11. 

Minimizers are k-mer sampling schemes designed to generate sketches for large sequences that preserve sufficiently long matches between sequences. Despite their widespread application, learning an effective minimizer scheme with optimal sketch size is still an open question. Most work in this direction focuses on designing schemes that work well on expectation over random sequences, which have limited applicability to many practical tools. On the other hand, several methods have been proposed to construct minimizer schemes for a specific target sequence. These methods, however, require greedy approximations to solve an intractable discrete optimization problem on the permutation space of k-mer orderings. To address this challenge, we propose: (a) a reformulation of the combinatorial solution space using a deep neural network re-parameterization; and (b) a fully differentiable approximation of the discrete objective. We demonstrate that our framework, DEEPMINIMIZER, discovers minimizer schemes that significantly outperform state-of-the-art constructions on genomic sequences.