Quantization optimizes machine learning inference for resource constrained environments by reducing the precision of its computation. In the extreme, even single-bit computations can produce acceptable results at dramatically lower cost. But this ultra-low-precision quantization is difficult to exploit because extracting optimal performance requires hand-tuning both high-level scheduling decisions and low-level implementations. As a result, practitioners settle for a few predefined quantized kernels, sacrificing optimality and restricting their ability to adapt to new hardware.

As a result of the increasing demand for deep neural network (DNN)-based services, efforts to develop dedicated hardware accelerators for DNNs are growing rapidly. However, while accelerators with high performance and efficiency on convolutional deep neural networks (Conv-DNNs) have been developed, less progress has been made with regards to fully-connected DNNs (FC-DNNs). In this paper, we propose MATIC (Memory Adaptive Training with In-situ Canaries), a methodology that enables aggressive voltage scaling of accelerator weight memories to improve the energy-efficiency of DNN accelerators. To enable accurate operation with voltage overscaling, MATIC combines the characteristics of destructive SRAM reads with the error resilience of neural networks in a memory-adaptive training process. Furthermore, PVT-related voltage margins are eliminated using bit-cells from synaptic weights as in-situ canaries to track runtime environmental variation. Demonstrated on a low-power DNN accelerator that we fabricate in 65 nm CMOS, MATIC enables up to 60-80 mV of voltage overscaling (3.3× total energy reduction versus the nominal voltage), or 18.6× application error reduction.

We introduce a learning-based framework to optimize tensor programs for deep learning workloads. Efficient implementations of tensor operators, such as matrix multiplication and high dimensional convolution are key enablers of effective deep learning systems. However, existing systems rely on manually optimized libraries such as cuDNN where only a narrow range of server class GPUs are well-supported. The reliance on hardware specific operator libraries limits the applicability of high-level graph optimizations and incurs significant engineering costs when deploying to new hardware targets. We use learning to remove this engineering burden. We learn domain specific statistical cost models to guide the search of tensor operator implementations over billions of possible program variants. We further accelerate the search by effective model transfer across workloads. Experimental results show that our framework delivers performance competitive with state-of-the-art hand-tuned libraries for low-power CPU, mobile GPU, and server-class GPU.

Many applications that can take advantage of accelerators are amenable to approximate execution. Past work has shown that neural acceleration is a viable way to accelerate approximate code. In light of the growing availability of on-chip field-programmable gate arrays (FPGAs), this paper explores neural acceleration on off-the-shelf programmable SoCs. We describe the design and implementation of SNNAP, a flexible FPGA-based neural accelerator for approximate programs. SNNAP is designed to work with a compiler workflow that configures the neural network's topology and weights instead of the programmable logic of the FPGA itself. This approach enables effective use of neural acceleration in commercially available devices and accelerates different applications without costly FPGA reconfigurations. No hardware expertise is required to accelerate software with SNNAP, so the effort required can be substantially lower than custom hardware design for an FPGA fabric and possibly even lower than current “C-to-gates” high-level synthesis (HLS) tools. Our measurements on a Xilinx Zynq FPGA show that SNNAP yields a geometric mean of 3.8× speedup (as high as 38.1×) and 2.8× energy savings (as high as 28 x) with less than 10% quality loss across all applications but one. We also compare SNNAP with designs generated by commercial HLS tools and show that SNNAP has similar performance overall, with better resource-normalized throughput on 4 out of 7 benchmarks.

BACKGROUND: With the growing number of public repositories for high-throughput genomic data, it is of great interest to combine the results produced by independent research groups. Such a combination allows the identification of common genomic factors across multiple cancer types and provides new insights into the disease process. In the framework of the proportional hazards model, classical procedures, which consist of ranking genes according to the estimated hazard ratio or the p-value obtained from a test statistic of no association between survival and gene expression level, are not suitable for gene selection across multiple genomic datasets with different sample sizes. We propose a novel index for identifying genes with a common effect across heterogeneous genomic studies designed to remain stable whatever the sample size and which has a straightforward interpretation in terms of the percentage of separability between patients according to their survival times and gene expression measurements. RESULTS: The simulations results show that the proposed index is not substantially affected by the sample size of the study and the censoring. They also show that its separability performance is higher than indices of predictive accuracy relying on the likelihood function. A simulated example illustrates the good operating characteristics of our index. In addition, we demonstrate that it is linked to the score statistic and possesses a biologically relevant interpretation.The practical use of the index is illustrated for identifying genes with common effects across eight independent genomic cancer studies of different sample sizes. The meta-selection allows the identification of four genes (ESPL1, KIF4A, HJURP, LRIG1) that are biologically relevant to the carcinogenesis process and have a prognostic impact on survival outcome across various solid tumors. CONCLUSION: The proposed index is a promising tool for identifying factors having a prognostic impact across a collection of heterogeneous genomic datasets of various sizes.

MOTIVATION: The most used criterion in microarray data analysis is nowadays the false discovery rate (FDR). In the framework of estimating procedures based on the marginal distribution of the P-values without any assumption on gene expression changes, estimators of the FDR are necessarily conservatively biased. Indeed, only an upper bound estimate can be obtained for the key quantity pi0, which is the probability for a gene to be unmodified. In this paper, we propose a novel family of estimators for pi0 that allows the calculation of FDR. RESULTS: The very simple method for estimating pi0 called LBE (Location Based Estimator) is presented together with results on its variability. Simulation results indicate that the proposed estimator performs well in finite sample and has the best mean square error in most of the cases as compared with the procedures QVALUE, BUM and SPLOSH. The different procedures are then applied to real datasets. AVAILABILITY: The R function LBE is available at http://ifr69.vjf.inserm.fr/lbe CONTACT: [email protected].