SystemX

This talk will cover the use of memory technology within computing platforms, from building large memory systems, to use in neuromorphic computing. What use cases can benefit from novel use of Memory-Driven Computing techniques. How do the latest industry moves creating open memory fabrics (including Gen-Z and Compute Express Link) impact system design? The use of high bandwidth memories and non-volatile memories - where do these technologies play relative to each other? How can they impact the way we build systems to deal with the challenge of processing and gaining knowledge/insights from all the data we are collecting at exponentially growing rates.

This talk will give an overview of how publicly announced products that have the Edge TPU make use of the product. The talk will then focus on what the Edge TPU architecture philosophy is the approach it takes to building custom silicon for ML workloads.

Join mailing list: Additional questions: Jon Candelaria, SystemX Seminar Instructor (jjcandel@stanford.edu)

Creating realistic virtual humans has traditionally been considered a research problem in Computer Animation primarily for entertainment applications. With the recent breakthrough in collaborative robots and deep reinforcement learning, accurately modeling human movements and behaviors has become a common challenge also faced by researchers in robotics and artificial intelligence. In this talk, I will first discuss our recent work on developing efficient computational tools for simulating and controlling human movements. By learning a differentiable kinematic constraints from the real world human motion data, we enable existing multi-body physics engines to simulate more humanlike motion. In a similar vein, we learn task-agnostic boundary conditions and energy functions from anatomically realistic neuromuscular models, effectively defining a new action space better reflecting the physiological constraints of the human body. The second part of the talk will focus on two different yet highly relevant problems: how to teach robots to move like humans and how to teach robots to interact with humans. While Computer Animation research has shown that it is possible to teach a virtual human to mimic real athletes' movements, the current techniques still struggle to reliably transfer a basic locomotion control policy to robot hardware in the real world. We developed a series of sim-to-real transfer methods to address the intertwined issue of system identification and policy learning for challenging locomotion tasks. Finally, I will showcase our effort on teaching robot to physically interact with humans in the scenarios of robot-assisted dressing and walking assistance.

Particle accelerators represent an indispensable tool in science, healthcare, and industry. However, the size and cost of conventional radio-frequency accelerators limits the utility and reach of this technology. Dielectric laser accelerators (DLAs) provide a compact and cost-effective solution to this problem by driving accelerator nanostructures with visible or near-infrared (NIR) pulsed lasers, resulting in a factor of 10,000x reduction in scale. Current implementations of DLAs rely on free-space lasers directly incident on the accelerating structures, limiting the scalability of this technology due to the need of bulky optics and precise mechanical alignment. Therefore, integration with an inherently scalable architecture, such as photonic integrated circuits, is paramount to the development of an MeV-scale DLA for applications.

In this talk, I will present the demonstration of a waveguide-integrated DLA, designed using a photonic inverse design approach. I will first review the operation of DLAs and describe how one can formulate a figure-of-merit for the optimization of these structures. I will then briefly introduce the inverse design framework that allows for efficient free-form optimization of these structures, enabling search of a design-space that goes far beyond that of the tuning of a few geometric parameters. I will discuss the approaches we take to couple light to these devices before presenting the results of our single-stage on-chip integrated accelerator. I will conclude with the directions we are taking to reach higher on-chip acceleration gradients and energy-gain, including utilizing foundry fabrication for multi-stage accelerators.

Understanding the neural basis of brain function and dysfunction requires developing multimodal methods to record and stimulate neural activity in the brain with high spatiotemporal resolution. We have been designing high-density opto-electrical devices to enable bi-directional (read/write) interfacing with the brain for long-term chronic studies.

One of the challenges of optical techniques for structural and functional recording and imaging is the scattering and absorption of light, limiting light-based methods to superficial layers of tissue. To overcome this challenge, implantable photonic waveguides such as optical fibers or graded-index (GRIN) lenses have been used. The prohibitive size and rigidity of these optical implants cause damage to the brain tissue and vasculature. In this talk, I will discuss our research on developing next generation optical neural interfaces that are microfabricated on flexible materials to minimize damage to the tissue.

First, I will discuss a compact flexible photonic platform based on biocompatible polymers, Parylene C and PDMS, for high-resolution light delivery into the tissue in a minimally-invasive way. This photonic platform can be monolithically integrated with implantable electrical neural interfaces.

I will also discuss our recent work on developing a novel complementary method for confining and steering light in the tissue using ultrasound. I will show that ultrasound waves can sculpt virtual optical waveguides in the tissue to define and steer the trajectory of light, thus obviating the need for implanting invasive physical devices in the brain.

These novel neurophotonic techniques will enable a whole gamut of applications from fundamental science studies to designing next generation neural prostheses.

No question in cognitive science is more challenging, or more fascinating, than the nature of consciousness. Until recently, the study of consciousness was usually considered a matter for philosophy rather than science, but there has been an explosion in the past decade of theories and experiments concerning how consciousness works. This talk will examine the nature of human consciousness from the interdisciplinary perspective of cognitive science. It will cover topics from the philosophy of mind, cognitive linguistics, neuroscience, psychology, and whether it may ever be possible for computers to be conscious, a topic studied in artificial intelligence.

This presentation will provide an overview of one of the new generation of 7nm devices from Xilinx, the Versal ACAP multi-core vector processor. The overall device architecture together with a detailed description of the Vector processor datapath will be presented. The programming model for the device will be described along with some application use cases for machine learning inference and 4G/5G wireless radio will be covered.

The Astrobee robots began operations on the International Space Station (ISS) in April of 2019. These 32-cm-wide, 9-kg, cube-shaped free flyers are capable of autonomously traversing the US Orbital Segment of ISS and performing tasks to assist astronauts in their day-to-day activities. The Astrobee Facility also provides an on-orbit testbed for guest scientists desiring to try new technologies in zero-g. This presentation will provide an overview of the Astrobee system and some of its key enabling technologies, including propulsion, navigation, and human interfaces. Additional information and publications can be found on the Astrobee website: nasa.gov/astrobee.

LIDAR can provide detailed spatial maps of an environment, a capability that is essential for applica- tions such as autonomous driving, robotics, building modeling, or augmented reality. Bulky optical elements and moving parts in current LIDAR systems have unfortunately made these systems heavy, expensive, and vibration-sensitive. Solid-state optical devices for the functionalities required for a LIDAR system have therefore driven much integrated photonics research in recent years. For on-chip beam-steering, optical phased arrays (OPAs) have emerged as a promising solution; however, the size and efficiency specifications for an effective LIDAR-OPA design have remained challenging for classical photonics design methods. In this talk, I will present how photonic inverse design methods can improve OPA design. Unlike classical photonic design, inverse design allows for arbitrary device geometries through the use of advanced elec- tromagnetic optimization methods. The extra degrees of freedom these methods provide has resulted in numerous new photonic devices with improved efficiency, reduced on-chip footprint or novel functionali- ties. Leveraging these new capabilities, we demonstrate an OPA design with a larger scanning cone and a smaller operating bandwidth as compared to the classical photonic design solutions.

Quantum computers powered by hundreds of gate based, superconducting qubits are just over the horizon. Several groups have already announced activities on quantum processors (QP) of 50 qubits or more. The systems containing these processors will be of fundamental importance in quantum algorithm development, on our path toward quantum advantage in the noisy intermediate scale quantum (NISQ) era. Quantum advantage is the point at which quantum computers can solve a problem faster, cheaper, or more accurately than their classical counterparts, which is likely to be accomplished before fully fault tolerant machines are available. This talk will address the trade-offs in the design of hybrid quantum-classical computing systems based on gate based superconducting processors approaching 100 qubits.