Hardware-Efficient Stochastic Binary CNN Architectures for Near-Sensor Computing

With recent advances in the field of artificial intelligence (AI) such as binarized neural networks (BNNs), a wide variety of vision applications with energy-optimized implementations have become possible at the edge. Such networks have the first layer implemented with high precision, which poses a challenge in deploying a uniform hardware mapping for the network implementation. Stochastic computing can allow conversion of such high-precision computations to a sequence of binarized operations while maintaining equivalent accuracy. In this work, we propose a fully binarized hardware-friendly computation engine based on stochastic computing as a proof of concept for vision applications involving multi-channel inputs. Stochastic sampling is performed by sampling from a non-uniform (normal) distribution based on analog hardware sources. We first validate the benefits of the proposed pipeline on the CIFAR-10 dataset. To further demonstrate its application for real-world scenarios, we present a case-study of microscopy image diagnostics for pathogen detection. We then evaluate benefits of implementing such a pipeline using OxRAM-based circuits for stochastic sampling as well as in-memory computing-based binarized multiplication. The proposed implementation is about 1,000 times more energy efficient compared to conventional floating-precision-based digital implementations, with memory savings of a factor of 45.

Interface Strengthening of α-Mg/C14–Mg2Ca Eutectic Alloy

This study investigates the effect of the α/C14 interface on the creep strength of α-Mg/C14–Mg2Ca eutectic alloy at 473 K under a stress of 40 MPa. The α/C14 interface is composed of terraces and steps, with terraces parallel to the (1101)α pyramidal plane of the α-Mg lamellae and to the (1120)C14 columnar plane of the C14–Mg2Ca lamellae. The creep curves of the alloy exhibit three stages: a normal transient creep stage, a minimum creep rate stage, and an accelerating stage. The minimum creep rate is proportional to the lamellar spacing, indicating that the α/C14 lamellar interface plays a creep-strengthening role. In the high-resolution transmission electron microscopy image captured of the specimen after the creep test, <a> dislocations can be mainly seen within the soft α-Mg lamellae, and they are randomly distributed at the α/C14 interface. In contrast, dislocations are rarely introduced in the hard C14–Mg2Ca lamellae. It is deduced that the α/C14 interface presents a barrier to dislocation gliding within the α-Mg lamellae and does not help rearrange the dislocations.

Image Processing Filters for Grids of Cells Analogous to Filters Processing Grids of Pixels

Intra- and extra-cellular processes shape tissues together. For understanding how neighborhood relationships between cells play a role in this process, having image processing filters based on these relationships would be beneficial. Those operations are known and their application to microscopy image data typically requires programming skills. User-friendly general purpose tools for pursuing image processing on a level of neighboring cells were yet missing. In this manuscript I demonstrate image processing filters which process grids of cells on tissue level and the analogy to their better known counter parts processing grids of pixels. The tools are available as part of free and open source software in the ImageJ/Fiji and napari ecosystems and their application does not require any programming experience.

Revealing the Systematic Nature of Coral Attachment to Reef Substrates

Reef-building coral colonies propagate by periodic sexual reproduction and continuous asexual fragmentation. The latter depends on successful attachment to the reef substrate through modification of soft tissues and skeletal growth. Despite decades of research examining coral sexual and asexual propagation, the contact response, tissue motion, and cellular reorganisation responsible for attaching to the substrate via a newly formed skeleton have not been documented. Here, we correlated fluorescence and electron microscopy image data with ‘live’ microscopic time-lapse of the coral tissue biomechanics and developed a multiscale imaging approach to establish the first “coral attachment model” (CAM) - identifying three distinct phases that determine the timing and success of attachment during asexual propagation: (i) an initial immune response, followed by (ii) fragment stabilisation through anchoring by the soft tissue and (iii) formation of a “lappet appendage” structure leading to substrate bonding of the tissue for encrustation through the onset of skeletal calcification. In developing CAM, we provide new frameworks and metrics that enable reef researchers, managers and coral restoration practitioners to evaluate attachment effectiveness needed to optimise species-substrate compatibility.

Labkit: Labeling and Segmentation Toolkit for Big Image Data

We present Labkit, a user-friendly Fiji plugin for the segmentation of microscopy image data. It offers easy to use manual and automated image segmentation routines that can be rapidly applied to single- and multi-channel images as well as to timelapse movies in 2D or 3D. Labkit is specifically designed to work efficiently on big image data and enables users of consumer laptops to conveniently work with multiple-terabyte images. This efficiency is achieved by using ImgLib2 and BigDataViewer as the foundation of our software. Furthermore, memory efficient and fast random forest based pixel classification inspired by the Waikato Environment for Knowledge Analysis (Weka) is implemented. Optionally we harness the power of graphics processing units (GPU) to gain additional runtime performance. Labkit is easy to install on virtually all laptops and workstations. Additionally, Labkit is compatible with high performance computing (HPC) clusters for distributed processing of big image data. The ability to use pixel classifiers trained in Labkit via the ImageJ macro language enables our users to integrate this functionality as a processing step in automated image processing workflows. Last but not least, Labkit comes with rich online resources such as tutorials and examples that will help users to familiarize themselves with available features and how to best use \Labkit in a number of practical real-world use-cases.