Search for: All records

Existing machine learning inference-serving systems largely rely on hardware scaling by adding more devices or using more powerful accelerators to handle increasing query demands. However, hardware scaling might not be feasible for fixed-size edge clusters or private clouds due to their limited hardware resources. A viable alternate solution is accuracy scaling, which adapts the accuracy of ML models instead of hardware resources to handle varying query demands. This work studies the design of a high-throughput inferenceserving system with accuracy scaling that can meet throughput requirements while maximizing accuracy. To achieve the goal, this work proposes to identify the right amount of accuracy scaling by jointly optimizing three sub-problems: how to select model variants, how to place them on heterogeneous devices, and how to assign query workloads to each device. It also proposes a new adaptive batching algorithm to handle variations in query arrival times and minimize SLO violations. Based on the proposed techniques, we build an inference-serving system called Proteus and empirically evaluate it on real-world and synthetic traces. We show that Proteus reduces accuracy drop by up to 3× and latency timeouts by 2-10× with respect to baseline schemes, while meeting throughput requirements. 

Existing machine learning inference-serving systems largely rely on hardware scaling by adding more devices or using more powerful accelerators to handle increasing query demands. However, hardware scaling might not be feasible for fixed-size edge clusters or private clouds due to their limited hardware resources. A viable alternate solution is accuracy scaling, which adapts the accuracy of ML models instead of hardware resources to handle varying query demands. This work studies the design of a high-throughput inferenceserving system with accuracy scaling that can meet throughput requirements while maximizing accuracy. To achieve the goal, this work proposes to identify the right amount of accuracy scaling by jointly optimizing three sub-problems: how to select model variants, how to place them on heterogeneous devices, and how to assign query workloads to each device. It also proposes a new adaptive batching algorithm to handle variations in query arrival times and minimize SLO violations. Based on the proposed techniques, we build an inference-serving system called Proteus and empirically evaluate it on real-world and synthetic traces. We show that Proteus reduces accuracy drop by up to 3× and latency timeouts by 2-10× with respect to baseline schemes, while meeting throughput requirements. 

ABSTRACT The nucleus accumbens (NAcc) and ventral pallidum (VP) are key nodes in the mesolimbic reward pathway that facilitate stimulus salience, including the regulation of social motivation and attachment. Primate species display variation in social behaviors, including different levels of impulsivity, bonding, and aggression. Previous research has implicated neuromodulation of the reward pathway in the differential expression of various social behaviors, suggesting that differences in neurotransmitter innervation may play a role in species‐specific patterns. To explore this, we examined serotonergic innervation in the NAcc and VP among primates. We used stereology to quantify serotonin transporter‐immunoreactive (SERT‐ir) axon length density in the NAcc and VP of 13 primate species, including humans, great apes, and cercopithecid and platyrrhine monkeys. Our data show that serotonergic innervation density within both the NAcc and VP is highly conserved among species. This finding contrasts with our previous findings of higher levels of SERT‐ir axons in the dorsal striatum of humans and great apes relative to monkeys, a human‐specific increase in dopaminergic innervation within the NAcc and VP, and a human‐specific increase of neuropeptide Y in the NAcc, highlighting the mosaic nature of innervation patterns among species. 

The extracellular matrix (ECM) regulates carcinogenesis by interacting with cancer cells via cell surface receptors. Discoidin Domain Receptor 2 (DDR2) is a collagen‐activated receptor implicated in cell survival, growth, and differentiation. Dysregulated DDR2 expression has been identified in various cancer types, making it as a promising therapeutic target. Additionally, cancer cells exhibit mechanosensing abilities, detecting changes in ECM stiffness, which is particularly important for carcinogenesis given the observed ECM stiffening in numerous cancer types. Despite these, whether collagen‐activated DDR2 signaling and ECM stiffness‐induced mechanosensing exert similar effects on cancer cell behavior and whether they operate through analogous mechanisms remain elusive. To address these questions, we performed bulk RNA sequencing (RNA‐seq) on human SH‐SY5Y neuroblastoma cells cultured on collagen‐coated substrates. Our results show that DDR2 downregulation induces significant changes in the cell transcriptome, with changes in expression of 15% of the genome, specifically affecting the genes associated with cell division and differentiation. We validated the RNA‐seq results by showing that DDR2 knockdown redirects the cell fate from proliferation to senescence. Like DDR2 knockdown, increasing substrate stiffness diminishes cell proliferation. Surprisingly, RNA‐seq indicates that substrate stiffness has no detectable effect on the transcriptome. Furthermore, DDR2 knockdown influences cellular responses to substrate stiffness changes, highlighting a crosstalk between these two ECM‐induced signaling pathways. Based on our results, we propose that the ECM could activate DDR2 signaling and mechanosensing in cancer cells to orchestrate their cell fate through distinct mechanisms, with or without involving gene expression, thus providing novel mechanistic insights into cancer progression. 

Abstract Astrocytes are the main homeostatic cell of the brain involved in many processes related to cognition, immune response, and energy expenditure. It has been suggested that the distribution of astrocytes is associated with brain size, and that they are specialized in humans. To evaluate these, we quantified astrocyte density, soma volume, and total glia density in layer I and white matter in Brodmann's area 9 of humans, chimpanzees, baboons, and macaques. We found that layer I astrocyte density, soma volume, and ratio of astrocytes to total glia cells were highest in humans and increased with brain size. Overall glia density in layer I and white matter were relatively invariant across brain sizes, potentially due to their important metabolic functions on a per volume basis. We also quantified two transporters involved in metabolism through the astrocyte‐neuron lactate shuttle, excitatory amino acid transporter 2 (EAAT2) and glucose transporter 1 (GLUT1). We expected these transporters would be increased in human brains due to their high rate of metabolic consumption and associated gene activity. While humans have higher EAAT2 cell density, GLUT1 vessel volume, and GLUT1 area fraction compared to baboons and chimpanzees, they did not differ from macaques. Therefore, EAAT2 and GLUT1 are not related to increased energetic demands of the human brain. Taken together, these data provide evidence that astrocytes play a unique role in both brain expansion and evolution among primates, with an emphasis on layer I astrocytes having a potentially significant role in human‐specific metabolic processing and cognition. 

Abstract The genusMacacais an ideal model for investigating the biological basis of primate social behavior from an evolutionary perspective. A significant amount of behavioral diversity has been reported among the macaque species, but little is known about the neural substrates that support this variation. The present study compared neural cell density and serotonergic innervation of the amygdala among four macaque species using histological and immunohistochemical methods. The species examined included rhesus (Macaca mulatta), Japanese (M. fuscata), pigtailed (M. nemestrina), and moor macaques (M. maura). We anticipated that the more aggressive rhesus and Japanese macaques would have lower serotonergic innervation within the amygdala compared to the more affiliative pigtailed and moor macaques. In contrast to our prediction, pigtailed macaques had higher serotonergic innervation than Japanese and moor macaques in the basal and central amygdala nuclei when controlling for neuron density. Our analysis of neural cell populations revealed that Japanese macaques possess significantly higher neuron and glia densities relative to the other three species, however we observed no glia‐to‐neuron ratio differences among species. The results of this study revealed serotonergic innervation and cell density differences among closely related macaque species, which may play a role in modulating subtle differences in emotional processing and species‐typical social styles.