Impact of On-chip Interconnect on In-memory Acceleration of Deep Neural Networks

With the widespread use of Deep Neural Networks (DNNs), machine learning algorithms have evolved in two diverse directions—one with ever-increasing connection density for better accuracy and the other with more compact sizing for energy efficiency. The increase in connection density increases on-chip data movement, which makes efficient on-chip communication a critical function of the DNN accelerator. The contribution of this work is threefold. First, we illustrate that the point-to-point (P2P)-based interconnect is incapable of handling a high volume of on-chip data movement for DNNs. Second, we evaluate P2P and network-on-chip (NoC) interconnect (with a regular topology such as a mesh) for SRAM- and ReRAM-based in-memory computing (IMC) architectures for a range of DNNs. This analysis shows the necessity for the optimal interconnect choice for an IMC DNN accelerator. Finally, we perform an experimental evaluation for different DNNs to empirically obtain the performance of the IMC architecture with both NoC-tree and NoC-mesh. We conclude that, at the tile level, NoC-tree is appropriate for compact DNNs employed at the edge, and NoC-mesh is necessary to accelerate DNNs with high connection density. Furthermore, we propose a technique to determine the optimal choice of interconnect for any given DNN. In this technique, we use analytical models of NoC to evaluate end-to-end communication latency of any given DNN. We demonstrate that the interconnect optimization in the IMC architecture results in up to 6 × improvement in energy-delay-area product for VGG-19 inference compared to the state-of-the-art ReRAM-based IMC architectures.

Methods of Comparison of the Economic Order Quantity and Just-In-Time Restocking Technologies. The Case Study

Estimation of the optimal size of order is one of the key tasks in determining the parameters of the urban freight restocking system. The existing analytical models and methods are considering each technology separately and they do not compare the Economic Order Quantity (EOQ) and Justin-tme (JIT) restocking technologies. The purpose of this research was to evaluate efficiency of the JIT and EOQ restocking technologies. The research would help in selecting the delivery model, analyzing functioning of existing JIT and EOQ models. The article presents an approach to determining the comparison in organizing supplies to the retailer. For this, the two supply models were compared. The Just-in-Time model is characterised by costs that are spend on transportation. The Economic Order Quantity model includes costs of transportation and storage in a warehouse. After calculations, application of the Just-in-Time model in the given conditions was determined.

Quality Prediction and Control in Wire Arc Additive Manufacturing via Novel Machine Learning Framework

Wire arc additive manufacturing (WAAM) is capable of rapidly depositing metal materials thus facilitating the fabrication of large-shape metal components. However, due to the multi-process-variability in the WAAM process, the deposited shape (bead width, height, depth of penetration) is difficult to predict and control within the desired level. Ultimately, the overall build will not achieve a near-net shape and will further hinder the part from performing its functionality without post-processing. Previous research primarily utilizes data analytical models (e.g., regression model, artificial neural network (ANN)) to forwardly predict the deposition width and height variation based on single or cross-linked process variables. However, these methods cannot effectively determine the optimal printable zone based on the desired deposition shape due to the inability to inversely deduce from these data analytical models. Additionally, the process variables are intercorrelated, and the bead width, height, and depth of penetration are highly codependent. Therefore, existing analysis cannot grant a reliable prediction model that allows the deposition (bead width, height, and penetration height) to remain within the desired level. This paper presents a novel machine learning framework for quantitatively analyzing the correlated relationship between the process parameters and deposition shape, thus providing an optimal process parameter selection to control the final deposition geometry. The proposed machine learning framework can systematically and quantitatively predict the deposition shape rather than just qualitatively as with other existing machine learning methods. The prediction model can also present the complex process-quality relations, and the determination of the deposition quality can guide the WAAM to be more prognostic and reliable. The correctness and effectiveness of the proposed quantitative process-quality analysis will be validated through experiments.

Evolutionary Inverse Material Identification: Bespoke Characterization of Soft Materials Using a Metaheuristic Algorithm

The growing interest in soft robotics has resulted in an increased demand for accurate and reliable material modelling. As soft robots experience high deformations, highly nonlinear behavior is possible. Several analytical models that are able to capture this nonlinear behavior have been proposed, however, accurately calibrating them for specific materials and applications can be challenging. Multiple experimental testbeds may be required for material characterization which can be expensive and cumbersome. In this work, we propose an alternative framework for parameter fitting established hyperelastic material models, with the aim of improving their utility in the modelling of soft continuum robots. We define a minimization problem to reduce fitting errors between a soft continuum robot deformed experimentally and its equivalent finite element simulation. The soft material is characterized using four commonly employed hyperelastic material models (Neo Hookean; Mooney–Rivlin; Yeoh; and Ogden). To meet the complexity of the defined problem, we use an evolutionary algorithm to navigate the search space and determine optimal parameters for a selected material model and a specific actuation method, naming this approach as Evolutionary Inverse Material Identification (EIMI). We test the proposed approach with a magnetically actuated soft robot by characterizing two polymers often employed in the field: Dragon Skin™ 10 MEDIUM and Ecoflex™ 00-50. To determine the goodness of the FEM simulation for a specific set of model parameters, we define a function that measures the distance between the mesh of the FEM simulation and the experimental data. Our characterization framework showed an improvement greater than 6% compared to conventional model fitting approaches at different strain ranges based on the benchmark defined. Furthermore, the low variability across the different models obtained using our approach demonstrates reduced dependence on model and strain-range selection, making it well suited to application-specific soft robot modelling.

Manipulation and control of droplets on surfaces in a homogeneous electric field

A method to manipulate and control droplets on a surface is presented. The method is based on inducing electric dipoles inside the droplets using a homogeneous external electric field. It is shown that the repulsive dipole force efficiently suppresses the coalescence of droplets moving on a liquid-infused surface (LIS). Using a combination of experiments, numerical computations and semi-analytical models, the dependence of the repulsion force on the droplet volumes, the distance between the droplets and the electric field strength is revealed. The method allows to suppress coalescence in complex multi-droplet flows and is real-time adaptive. When the electric field strength exceeds a critical value, tip streaming from the droplets sets in. Based on that, it becomes possible to withdraw minute samples from an array of droplets in a parallel process.

Simplified analytical method for the robustness assessment of precast reinforced concrete structural systems

The article presents the simplified implementation of alternative load path method based on the energy balance approach. This method should be used to check the global resistance of a damaged structural system after the occurrence of an accidental event. Basic assumptions of simplified analytical models for modelling resistance of horizontal ties in a damaged structural system, taking into account the membrane (chain) effects, were presented. An approach to modelling the dynamic resistance of a damaged structural system based on the energy balance method is described. Calculated dependencies for checking the robustness of a prefabricated multi-storey building with hollow-core slabs after the loss of the central column are proposed and considered using an example. On the considered example, a comparison of the required tie sections area with the dynamic resistance designed using the energy balance method (EBM) and according to the current standards, and a statistical assessment of the reliability of the load-bearing capacity models are carried out. In the end, a brief algorithm for the simplified calculation of the dynamic resistance of a damaged structural system is proposed.