PROJECTS
Additive Manufacturing
X-ray computed tomography (XCT) plays a critical role in non-destructive evaluation (NDE) of complex parts in metal additive manufacturing (AM), where characterization of metals in 3D with high spatial resolution is critical for qualification/certification of manufactured parts. However, for metallic objects, beam hardening and metal artifacts pose significant challenges for the analysis of reconstructed images from XCT scanners. This could be further exacerbated due to complex geometry of the part to be scanned as well as the noise and scattering effects in the measurements. Current methods to mitigate the noise and artifacts typically involve very long scan times, making several measurements at localized regions of interests (ROIs) with smaller field of views (FOVs), and development of new algorithm. Still, standard approaches produce artifacts for high quality reconstruction of the complex parts being scanned, which make tasks such as detecting pores and defects in the parts very challenging. In this work, we are developing AI-CT, a framework that uses CAD (computer-aided designs) models of the AM parts, along with physics-based parameters to simulate XCT data with noise/artifacts, and leverages a 2.5D convolutional neural network (CNN) to learn to suppress noise/artifacts in the synthetically generated XCT reconstructions. Once the network is trained on the synthetic data, we apply it to experimental data sets by leveraging generative adversarial networks and domain adaptation techniques.
Resources: abstract
Electronics Test
Owing to the inherent fault tolerance of deep neural network (DNN) models used for classification, many structural faults in the processing elements (PEs) of a systolic array- based AI accelerator are functionally benign. Brute-force fault simulation for determining fault criticality is computationally expensive due to many potential fault sites in the accelerator array and the dependence of criticality characterization of PEs on the functional input data. Supervised learning techniques can be used to accurately estimate fault criticality but it requires ground truth for model training. The ground-truth collection involves extensive and computationally expensive fault simulations. We present a framework for analyzing fault criticality with a negligible amount of ground-truth data. We incorporate the gate-level structural and functional information of the PEs in their “neural twins”, referred to as “PE-Nets”. The PE netlist is translated into a trainable PE-Net, where the standard-cell instances are substituted by their corresponding “Cell-Nets” and the wires translate to neural connections. Each Cell-Net is a pre-trained DNN that models the Boolean-logic behavior of the corresponding standard cell. In the PE-Net, every neural connection is associated with a bias that represents a perturbation in the signal propagated by that connection. We utilize a recently proposed misclassification-driven training algorithm to sensitize and identify biases that are critical to the functioning of the accelerator for a given application workload. The proposed framework achieves up to 100% accuracy in fault-criticality classification in 16-bit and 32-bit PEs by using the criticality knowledge of only 2% of the total faults in a PE.
Resources: video
Complex Network
In this paper, we develop an analytical framework which explains the emergence of superpeer networks on execution of the commercial peer-to-peer bootstrapping protocols by incoming nodes. Bootstrapping protocols exploit physical properties of the online peers like resource content, processing power, storage space, connectivity etc as well as take the finiteness of bandwidth of each online peer into consideration. With the help of rate equations, we show that execution of these protocols results in the emergence of superpeer nodes in the network - the exact degree distribution is evaluated. We validate the framework developed in this paper through extensive simulation. The analysis of the results shows that the amount of superpeers produced in the network depends on the protocol as well as the properties of the joining nodes. Interestingly, our analysis reveals that increase in the amount of resource and the number of resourceful nodes do not always help to increase the fraction of superpeer nodes in the network. The impact of the frequent leaving of the peers on the topology of the emerging network is also evaluated. As an application study, we show that our framework can explain the topological configuration of commercial Gnutella networks. The developed model can almost perfectly match the degree distribution of Gnutella network.
Resources: paper
NoSQL Database
Considerable work has been done on the topic of schema design for relational databases. However, it has not been formally studied for NoSQL databases. We consider the schema design for NoSQL databases in the following two use cases. In the first case, an application administrator has an application running on a relational database and wants to migrate it to a NoSQL system. In the second case, an application designer is developing a new application on a NoSQL database. To recommend the schema in these two cases, we have developed transformation rules, a cost model, and a search algorithm as part of a recommendation utility that analyses the observed or the user-provided application access patterns and offers a recommendation.