In this paper we describe a new mobile architecture, MobileNetV2, that improves the state of the art performance of mobile models on multiple tasks and benchmarks as well as across a spectrum of different model sizes. We also describe efficient ways of applying these mobile models to object detection in a novel framework we call SSDLite. Additionally, we demonstrate how to build mobile semantic...

Recent advances in computing allow taking new look at matrix multiplication, where the key ideas are: decreasing interest in recursion, development of processors with thousands (potentially millions) of processing units, and influences from the Algebraic Path Problems. In this context, we propose a generalized matrix-matrix multiply-add (MMA) operation and illustrate its usability. Furthermore,...

The multiply-accumulate (MAC) unit is an addition to the ColdFire core that provides a new level of support for signal processing algorithms. This paper details the associated support functions within the instruction set. Signal processing features are discussed for potential applications in embedded systems, and examples are presented for implementing some useful algorithms.

Usually a Riemannian geometry is considered to be the most general geometry, which could be used as a space-time geometry. In fact, any Riemannian geometry is a result of some deformation of the Euclidean geometry. Class of these Riemannian deformations is restricted by a series of unfounded constraints. Eliminating these constraints, one obtains a more wide class of possible space-time geometr...

Is there reason to believe that our brains have evolved to make efficient decisions so that the details of the internal process are irrelevant? I develop a model which illustrates a limitation of adaptive processes: improvements tend to come in the form of kludges. A kludge is a marginal adaptation that compensates for, but does not eliminate fundamental design inefficiencies. When kludges accu...

In this paper, a fast radix-4 complex FFT implementation using 4-parallel SIMD instructions is presented. Four radix-4 butterflies are calculated in parallel at all stages by loading consecutive 4 elements into a register. At the last stage, every 4 elements is packed into a register and calculated in parallel. This regular data flow enables higher parallelism and an overhead reduction in data ...

We classify all path-delay faults of a combinational circuit into three categories: singly-testable (ST), multiply-testable (MT), and singly-testable dependent (ST-dependent). The classification uses any unaltered single stuck-at fault test generation tool. Only two runs of this tool on a model network derived from the original network are performed. As a by-product of this process, we generate...

We provide a detailed analysis of Kahan’s algorithm for the accurate computation of the determinant of a 2 × 2 matrix. This algorithm requires the availability of a fused multiply-add instruction. Assuming radix-β, precision-p floating-point arithmetic with β even, p ≥ 2, and barring overflow or underflow we show that the absolute error of Kahan’s algorithm is bounded by (β+1)/2 ulps of the exa...

An existing rule-based assertion classifier is ported from English to Swedish: pyConTextSwe. Evaluation on Swedish clinical texts shows that the English lexical resources are useful, but that there are assertion cues not obtainable in existing resources. Iterative error correction of cue lexicons improves results for the ported classifier. Overall final results are 82% F-score on a development ...

Stack filters are a class of non-linear, digital filters whose behaviour is defined by Positive Boolean Functions rather than the multiply-accumulate, arithmetic operators that are widely used in conventional digital filters. The implementation of stack filters that can be adapted in real-time is feasible using dynamically reconfigurable FPGAs. This paper reports on efforts to map adaptive stac...