ABSTRACT
The constant drive to achieve higher performance in deep neural networks (DNNs) has led to the proliferation of very large models. Model training, however, requires intensive computation time and energy. Memristor-based compute-in-memory (CIM) modules can perform vector-matrix multiplication (VMM) in place and in parallel, and have shown great promises in DNN inference applications. However, CIM-based model training faces challenges due to non-linear weight updates, device variations, and low-precision. In this work, a mixed-precision training scheme is experimentally implemented to mitigate these effects using a bulk-switching memristor-based CIM module. Low-precision CIM modules are used to accelerate the expensive VMM operations, with high-precision weight updates accumulated in digital units. Memristor devices are only changed when the accumulated weight update value exceeds a pre-defined threshold. The proposed scheme is implemented with a system-onchip of fully integrated analog CIM modules and digital sub-systems, showing fast convergence of LeNet training to 97.73%. The efficacy of training larger models is evaluated using realistic hardware parameters and verifies that CIM modules can enable efficient mix-precision DNN training with accuracy comparable to full-precision software-trained models. Additionally, models trained on chip are inherently robust to hardware variations, allowing direct mapping to CIM inference chips without additional re-training.
ABSTRACT
BACKGROUND: The low incidence of cytomegalovirus (CMV) infection in neonates decreases the risk of viral transmission with cord blood transplantation. Cord blood donors are screened by testing the maternal sample for total antibodies to CMV. Some cord blood banks also screen cord blood for CMV-DNA. The aim of this study was to develop and validate a multiplex real-time polymerase chain reaction assay to measure CMV viral load in cord blood from asymptomatic infants with congenital CMV infection and to assess the impact of CMV infection on cord blood hematopoietic progenitor cell concentrations and colony-forming unit functionality. STUDY DESIGN AND METHODS: CMV infection was evaluated in two groups of cord blood donors: 1) 30,308 neonates prospectively screened by saliva culture, including 41 positive cases (0.14%), all from mothers with total antibodies to CMV; and 2) 4712 newborns from mothers with total antibodies to CMV who were screened retrospectively by polymerase chain reaction, including 18 positive cases (0.38%). All 59 infants with CMV were asymptomatic at birth. RESULTS: Among the 59 positive cases, the average CMV viral load in cord blood was 20.6 × 104 viral copies (vc)/mL; seven of 59 mothers (12%) had CMV-DNA detected, however, with no association to their newborns' CMV viral load. Levels of colony-forming units, CD34+ /CD45+ cells, and total nucleated cells measured in a cohort of CMV-positive cord blood samples were higher than those in the matched control group. CONCLUSION: We developed and validated a multiplex real-time polymerase chain reaction assay to detect CMV-DNA in cord blood. In our study, maternal total antibodies to CMV or CMV-DNA at birth were poor predictors of infection in cord blood donors. Furthermore, our results suggest that CMV congenital infection impacts CD34+ /CD45+ cells and some hematopoietic progenitor cells toward higher proliferation.
