Joint Program and Layout Transformations to Enable Convolutional Operators on Specialized Hardware Based on Constraint Programming

The success of Deep Artificial Neural Networks (DNNs) in many domains created a rich body of research concerned with hardware accelerators for compute-intensive DNN operators. However, implementing such operators efficiently with complex hardware intrinsics such as matrix multiply is a task not yet automated gracefully. Solving this task often requires joint program and data layout transformations. First solutions to this problem have been proposed, such as TVM, UNIT, or ISAMIR, which work on a loop-level representation of operators and specify data layout and possible program transformations before the embedding into the operator is performed. This top-down approach creates a tension between exploration range and search space complexity, especially when also exploring data layout transformations such as im2col, channel packing, or padding. In this work, we propose a new approach to this problem. We created a bottom-up method that allows the joint transformation of both computation and data layout based on the found embedding. By formulating the embedding as a constraint satisfaction problem over the scalar dataflow, every possible embedding solution is contained in the search space. Adding additional constraints and optimization targets to the solver generates the subset of preferable solutions. An evaluation using the VTA hardware accelerator with the Baidu DeepBench inference benchmark shows that our approach can automatically generate code competitive to reference implementations. Further, we show that dynamically determining the data layout based on intrinsic and workload is beneficial for hardware utilization and performance. In cases where the reference implementation has low hardware utilization due to its fixed deployment strategy, we achieve a geomean speedup of up to × 2.813, while individual operators can improve as much as × 170.

An Architecture and Reference Implementation for WSN-Based IoT Systems

The Internet of Things and its technologies have evolved quickly in recent years. It became an umbrella term for various technologies, embedded devices, smart objects, and web services. Although it has gained maturity, there is still no clear or common definition of references for creating WSN-based IoT systems. In the awareness that creating an omniscient and ideal architecture that can suit all design requirements is not feasible, modular and scalable architecture that supports adding or subtracting components to fit a lot of requirements of various use cases should be provided as a starting point. This chapter discusses such an architecture and reference implementation. The architecture should cover multiple layers, including the cloud, the gateway, and the edges of the target system, which allows monitoring the environment, managing the data, programming the edge nodes and networking model to establish communication between horizontal and vertical embedded devices. In order to exemplify the proposed architecture and reference implementation, a smart irrigation case study is used.

Studying the Evolution of Neural Activation Patterns During Training of Feed-Forward ReLU Networks

The ability of deep neural networks to form powerful emergent representations of complex statistical patterns in data is as remarkable as imperfectly understood. For deep ReLU networks, these are encoded in the mixed discrete–continuous structure of linear weight matrices and non-linear binary activations. Our article develops a new technique for instrumenting such networks to efficiently record activation statistics, such as information content (entropy) and similarity of patterns, in real-world training runs. We then study the evolution of activation patterns during training for networks of different architecture using different training and initialization strategies. As a result, we see characteristic- and general-related as well as architecture-related behavioral patterns: in particular, most architectures form bottom-up structure, with the exception of highly tuned state-of-the-art architectures and methods (PyramidNet and FixUp), where layers appear to converge more simultaneously. We also observe intermediate dips in entropy in conventional CNNs that are not visible in residual networks. A reference implementation is provided under a free license1.

Design and Implementation of a Gesture-Aided E-Learning Platform

In gesture-aided learning (GAL), learners perform specific body gestures while rehearsing the associated learning content. Although this form of embodiment has been shown to benefit learning outcomes, it has not yet been incorporated into e-learning. This work presents a generic system design for an online GAL platform. It is comprised of five modules for planning, administering, and monitoring remote GAL lessons. To validate the proposed design, a reference implementation for word learning was demonstrated in a field test. 19 participants independently took a predefined online GAL lesson and rated their experience on the System Usability Scale and a supplemental questionnaire. To monitor the correct gesture execution, the reference implementation recorded the participants’ webcam feeds and uploaded them to the instructor for review. The results from the field test show that the reference implementation is capable of delivering an e-learning experience with GAL elements. Designers of e-learning platforms may use the proposed design to include GAL in their applications. Beyond its original purpose in education, the platform is also useful to collect and annotate gesture data.

A Protocol for a Market of Machine Learning Models

This paper describes a protocol for a market of machine learning models. The economic interaction involves two types of agents: data providers- agents that have some data and want to use it to get a predictive model, and model providers- agents able to use the data to generate predictive models. First, we will show that the process is informationally asymmetric, therefore a standard direct market can not function. Then, we design a protocol with the aim of creating a viable and efficient market mechanism for these particular services, under the specific challenges of information asymmetries. The protocol is theoretically analysed, to establish it’s correctness and computational complexity. We also propose a simple reference implementation based on a HTTP API. The implementation is then used in a few case studies, and analysed empirically.

Sketching and sampling approaches for fast and accurate long read classification

In modern sequencing experiments, identifying the sources of the reads is a crucial need. In metagenomics, where each read comes from one of potentially many members of a community, it can be important to identify the exact species the read is from. In other settings, it is important to distinguish which reads are from the targeted sample and which are from potential contaminants. In both cases, identification of the correct source of a read enables further investigation of relevant reads, while minimizing wasted work. This task is particularly challenging for long reads, which can have a substantial error rate that obscures the origins of each read. Existing tools for the read classification problem are often alignment or index-based, but such methods can have large time and/or space overheads. In this work, we investigate the effectiveness of several sampling and sketching-based approaches for read classification. In these approaches, a chosen sampling or sketching algorithm is used to generate a reduced representation (a "screen") of potential source genomes for a query readset before reads are streamed in and compared against this screen. Using a query read's similarity to the elements of the screen, the methods predict the source of the read. Such an approach requires limited pre-processing, stores and works with only a subset of the input data, and is able to perform classification with a high degree of accuracy. The sampling and sketching approaches investigated include uniform sampling, methods based on MinHash and its weighted and order variants, a minimizer-based technique, and a novel clustering-based sketching approach. We demonstrate the effectiveness of these techniques both in identifying the source microbial genomes for reads from a metagenomic long read sequencing experiment, and in distinguishing between long reads from organisms of interest and potential contaminant reads. We then compare these approaches to existing alignment, index and sketching-based tools for read classification, and demonstrate how such a method is a viable alternative for determining the source of query reads. Finally, we present a reference implementation of these approaches at https://github.com/arun96/sketching.

Automated Deployment of Data Lake

A Data Lake is a central location that can store all your structured and unstructured data, no matter the source or format. Automated deployment for data lake solution is an automated reference implementation that deploys a highly available, cost-effective data lake architecture on the AWS Cloud along with a user-friendly console for searching and requesting datasets. The solution automatically configures the core AWS services necessary to easily tag, search, share, transform, analyse, and govern specific subsets of data across a company or with other external users. The solution deploys a console that users can access to search and browse available datasets for their business needs. Keywords: Data Lake, Cloud Computing, Aws, Ec2, S3, Athena, Glue, Cloud formation.

A general framework for modeling and dynamic simulation of multibody systems using factor graphs

In this paper, we present a novel general framework grounded in the factor graph theory to solve kinematic and dynamic problems for multibody systems. Although the motion of multibody systems is considered to be a well-studied problem and various methods have been proposed for its solution, a unified approach providing an intuitive interpretation is still pursued. We describe how to build factor graphs to model and simulate multibody systems using both, independent and dependent coordinates. Then, batch optimization or a fixed lag smoother can be applied to solve the underlying optimization problem that results in a highly sparse nonlinear minimization problem. The proposed framework has been tested in extensive simulations and validated against a commercial multibody software. We release a reference implementation as an open-source C++ library, based on the GTSAM framework, a well-known estimation library. Simulations of forward and inverse dynamics are presented, showing comparable accuracy with classical approaches. The proposed factor graph-based framework has the potential to be integrated into applications related with motion estimation and parameter identification of complex mechanical systems, ranging from mechanisms to vehicles, or robot manipulators.

A Canonical Set of Operations for Editing Dashboard Layouts in Virtual Reality

Virtual reality (VR) is a powerful technological framework that can be considered as comprising any kind of device that allows for 3D environments to be simulated and interacted with via a digital interface. Depending on the specific technologies used, VR can allow users to experience a virtual world through their different senses, i.e., most often sight, but also through touch, hearing, and smell. In this paper, it is argued that a key impediment to the widespread adoption of VR technology today is the lack of interoperability between users’’ existing digital life (including 2D documents, videos, the Web, and even mobile applications) and the 3D spaces. Without such interoperability, 3D spaces offered by current VR platforms seem empty and lacking in functionality. In order to improve this situation, it is suggested that users could benefit from being able to create dashboard layouts (comprising 2D displays) for themselves in the 3D spaces, allowing them to arrange, view and interact with their existing 2D content alongside the 3D objects. Therefore, the objective of this research is to help users organize and arrange 2D content in 3D spaces depending on their needs. To this end, following a discussion on why this is a challenging problem—both from a scientific and from a practical perspective—a set of operations are proposed that are meant to be minimal and canonical and enable the creation of dashboard layouts in 3D. Based on a reference implementation on the MaxWhere VR platform, a set of experiments were carried out to measure how much time users needed to recreate existing layouts inside an empty version of the corresponding 3D spaces, and the precision with which they could do so. Results showed that users were able to carry out this task, on average, at a rate of less than 45 s per 2D display at an acceptably high precision.