An official website of the United States government
Hardware accelerators based on emerging device technologies are gaining traction for inference workloads, but effective methods for their training remain an open area of research. We propose an efficient hardware-aware methodology for training neural networks with ternary weights that are mappable to emerging memory device arrays. We study device-network interactions across a variety of scenarios using simulated and experimentally measured datasets from ferroelectric field-effect transistor (FeFET) devices with varying characteristics. We quantify the impact of device non-idealities on network training by investigating device-level metrics, network-level metrics, loss landscapes, as well as parameter optimization trajectories. We validate our approach by mapping a hardware-aware solution to an emulated system with parameters calibrated to experimental measurements, highlighting several trade-offs. Hardware-aware training results on FeFET-based multi-layer perceptron networks, long short-term memory networks, and deep convolutional networks demonstrate competitive performance at lower overheads compared to existing schemes, indicating architectural and computational scalability. It is found that devices with low variability, non-linearity, and high dynamic range exhibit training characteristics closest to a software baseline. We provide evidence that device non-idealities inject noise during backpropagation, leading to sharper loss landscapes and higher-dimensional optimization trajectories, which make device networks more difficult to train than software counterparts. We also identify optimal operating voltages for investigated devices by utilizing our hardware-aware training and inference methodologies.
more »
« less
B-spline Parameterized Joint Optimization of Reconstruction and K-space Sampling Patterns (BJORK) for Accelerated 2D Acquisition
The proposed approach, BJORK, provides a robust and generalizable workflow to jointly optimize non-Cartesian sampling patters and a physics-informed reconstruction. Several approaches, including re-parameterization of trajectories, multi-level optimization, and non-Cartesian unrolled neural networks, are introduced to improve training effect and avoid sub-optimal local minima. The invivo experiments show that the networks and trajectories learned on simulation dataset are transferable to the real acquisition even with different parameter-weighted MRI contrasts and noise-levels, and demonstrate improved image quality compared with previous learning-based and model-based trajectory optimization methods.
more »
« less
- Award ID(s):
- 1838179
- PAR ID:
- 10309641
- Date Published:
- Journal Name:
- International Society Magnetic Resonance in Medicine
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
We present a virtual element method (VEM)-based topology optimization framework using polyhedral elements, which allows for convenient handling of non-Cartesian design domains in three dimensions. We take full advantage of the VEM properties by creating a unified approach in which the VEM is employed in both the structural and the optimization phases. In the structural problem, the VEM is adopted to solve the three-dimensional elasticity equation. Compared to the finite element method, the VEM does not require numerical integration (when linear elements are used) and is less sensitive to degenerated elements (e.g., ones with skinny faces or small edges). In the optimization problem, we introduce a continuous approximation of material densities using the VEM basis functions. When compared to the standard element-wise constant approximation, the continuous approximation enriches the geometrical representation of structural topologies. Through two numerical examples with exact solutions, we verify the convergence and accuracy of both the VEM approximations of the displacement and material density fields. We also present several design examples involving non-Cartesian domains, demonstrating the main features of the proposed VEM-based topology optimization framework. The source code for a MATLAB implementation of the proposed work, named PolyTop3D, is available in the (electronic) Supplementary Material accompanying this publication.more » « less
-
We present a discrete-optimization technique for finding feasible robot arm trajectories that pass through provided 6-DOF Cartesian-space end-effector paths with high accuracy, a problem called pathwise-inverse kinematics. The output from our method consists of a path function of joint-angles that best follows the provided end-effector path function, given some definition of ``best''. Our method, called Stampede, casts the robot motion translation problem as a discrete-space graph-search problem where the nodes in the graph are individually solved for using non-linear optimization; framing the problem in such a way gives rise to a well-structured graph that affords an effective best path calculation using an efficient dynamic-programming algorithm. We present techniques for sampling configuration space, such as diversity sampling and adaptive sampling, to construct the search-space in the graph. Through an evaluation, we show that our approach performs well in finding smooth, feasible, collision-free robot motions that match the input end-effector trace with very high accuracy, while alternative approaches, such as a state-of-the-art per-frame inverse kinematics solver and a global non-linear trajectory-optimization approach, performed unfavorably.more » « less
-
Hsia, KJ; Rogers, JA; Suo, Z; Zhao, X (Ed.)Topology optimization algorithms often employ a smooth density function to implicitly represent geometries in a discretized domain. While this implicit representation offers great flexibility to parametrize the optimized geometry, it also leads to a transition region. Previous approaches, such as the Solid Isotropic Material Penalty (SIMP) method, have been proposed to modify the objective function aiming to converge toward integer density values and eliminate this non-physical transition region. However, the iterative nature of topology optimization renders this process computationally demanding, emphasizing the importance of achieving fast convergence. Accelerating convergence without significantly compromising the final solution can be challenging. In this work, we introduce a machine learning approach that leverages the message-passing Graph Neural Network (GNN) to eliminate the non-physical transition zone for the topology optimization problems. By representing the optimized structures as weighted graphs, we introduce a generalized filtering algorithm based on the topology of the spatial discretization. As such, the resultant algorithm can be applied to two- and three-dimensional space for both Cartesian (structured grid) and non-Cartesian discretizations (e.g. polygon finite element). The numerical experiments indicate that applying this filter throughout the optimization process may avoid excessive iterations and enable a more efficient optimization procedure.more » « less
-
This paper introduces LeTO, a method for learning constrained visuomotor policy with differentiable trajectory optimization. Our approach integrates a differentiable optimization layer into the neural network. By formulating the optimization layer as a trajectory optimization problem, we enable the model to end-to-end generate actions in a safe and constraint-controlled fashion without extra modules. Our method allows for the introduction of constraint information during the training process, thereby balancing the training objectives of satisfying constraints, smoothing the trajectories, and minimizing errors with demonstrations. This “gray box” method marries optimization-based safety and interpretability with powerful representational abilities of neural networks. We quantitatively evaluate LeTO in simulation and in the real robot. The results demonstrate that LeTO performs well in both simulated and real-world tasks. In addition, it is capable of generating trajectories that are less uncertain, higher quality, and smoother compared to existing imitation learning methods. Therefore, it is shown that LeTO provides a practical example of how to achieve the integration of neural networks with trajectory optimization. We release our code at https://github.com/ZhengtongXu/LeTO.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript
- Conference Paper:
- The DOI is not currently available.
