Current projects

Sparse Adaptive Local Learning for Sensing and Analytics (SALLSA)

State-of-the-art inference models (IMs) learn their deep network structure directly from their sensory inputs, yielding sparse approximations that enable high-performance target detection and tracking. In the Sparse Adaptive Local Learning for Sensing and Analytics (SALLSA) project, we propose to apply a locally competitive algorithm based a network of sparse spiking neurons as compute elements that self-organize and self-adapt to minimize an energy function whose ground states correspond to optimal sparse representations. An imager is directly integrated with the neural network to accomplish end-to-end sensing and analytics. We will create a shared bus architecture to exploit the sparsity of neuron firing and pixel encoding for a scalable ultra low- power and high-performance implementation. The emerging memristor device will be used as the associative memory in neurons to perform lookup, storage, and compute-in-memory. Altogether, the proposed system will deliver more than three orders of magnitude improvement in performance by taking advantage of the local algorithm and sparsity, and the scalable shared bus architecture and memristor-based neuron implementations will enable sub-10mW IM implementations that represent at least four orders of magnitude improvement in power efficiency over the current state-of-the-art.
The envisioned image processing pipeline (IPP) for target detection and tracking is based on a collection of IMs that are tasked to perform feature extraction and detection. The IM is an entirely self-contained single-chip module that senses the environment through an imager, translates the pixels to sparse pulse trains, and learns and detects the features using a fully- interconnected sparse-spiking neural network. We propose two neuron implementations: one based on mixed-CMOS and another based on nanoscale memristor crossbar that performs compute-in-memory for the ultimate energy efficiency. A close analogy can be drawn between the proposed IPP and the human visual system. We summarize the key innovations in this research that span algorithm and architecture, to circuits and devices. Powerful and sophisticated simulation tools underpin the entire investigation.

Funding:

• DARPA, Unconventional Processing of Signals for Intelligent Data Exploitation (UPSIDE)
• Contract: # HR0011-13-2-0015
• Duration: May 1, 2013 – Aug 30, 2017

Team:

• Wei Lu, UMICH (PI)
• Zhenya Zhang, UMICH (co-PI)
• Michael Flynn, UMICH (co-PI)
• Garrett Kenyon, LANL/New Mexico Consortium (co-PI)
• Christof Teuscher, PSU (co-PI)

Inference at the Nanoscale

The goal of this project is to develop a new inference-based information processing structure that performs probabilistic computing using radically new nanoscale devices.  Our approach exploits the analog, time-dependent properties of such devices, and their massive parallelism.  By doing so, such a computing structure will be more efficient and scalable than by using more traditional digital hardware.  This approach is the first that we are aware of to include time-dependent circuit elements to build analog associative memories that approximate Bayesian inference, and which are, in turn, assembled into complex networks that capture higher order structure in streams of data.  Our ultimate goal is to use these circuits to develop hybrid CMOS / molecular scale implementations of a Field Adaptable Bayesian Array (FABA), which, we believe, has the potential to be a key component for Cyber-Enabled discovery.
Two key developments then are a design exploration methodology for such devices, and a massively parallel architecture for data capture and inference.  In this work we explore a new paradigm for using nanoscale electronics for emerging applications by starting with the “top-down” system requirements rather than by finding applications for new device concepts (“bottom-up”).
We address Cyber-Enabled discovery in two ways.  The first concerns the design of analog circuits based on complex nano and molecular scale devices with time-varying properties. Designing analog nano-electronic circuits that perform inference through space and time and which consist of dynamic components (such as mem-resistance and mem-capacitance) is extraordinarily difficult. This is particularly true when one considers the wide range of complex devices that are being developed in laboratories around the world for nano and molecular scale electronics.  For this effort we have defined an Exploration Methodology that combines multiple levels of abstraction and evolvable computation.
As the semiconductor industry struggles with where to go next, the work proposed here may provide insight into radical new approaches to architecture, circuits and devices.  This research will ultimately benefit society by enhancing human cognition and generating new knowledge from the wealth of heterogeneous digital data society has to deal with.

Funding:

• National Science Foundation (NSF), Cyber-enabled Discovery and Innovation (CDI), Type II award
• Award #: 1028378
• Duration: Sep 15, 2010 – Aug 31, 2016 (with NCE)

Team:

• Christof Teuscher (PI)
• Dan Hammerstrom, PSU
• John Carruthers, PSU
• Dmitry Strukov, UC Santa Barbara

Molecular Computing for the Real World

Molecular computing is a promising computational paradigm, in which computational functions are evaluated at the nanoscale, with potential applications in smart molecular diagnostics
and therapeutics. However, despite recent advances in the field, prospects for direct application of these techniques to solve real-world problems are limited by the lack of robust interfaces between molecular computers and biological and chemical systems. This project will address these limitations by targeting two application domains, wide-spectrum chemical sensing and
cell surface analysis using molecular logic cascades. Drawing on a combination of experimental, theoretical, and computational tools, molecular computing systems will be developed for use in these application domains. Molecular circuit architectures that process sensor inputs from chemical sensors and cell-surface analysis reactions will be designed, modeled, and implemented in the laboratory, and computational modeling will be used to predict and optimize interactions between DNA circuit components and their binding targets. Furthermore, advanced molecular circuit architectures capable of adaptive, bio-inspired behavior, such as dynamic learning and adaptation, will be designed, with a view to future experimental implementations of these features.

Funding

• National Science Foundation (NSF), CISE, SHF
• Duration: Sep 1, 2015 – Aug 31, 2020

Team:

• Darko Stefanovic, UNM
• Steven Graves, UNM
• Lakin Matthew, UNM
• Lydia E Tapia, UNM
• Sergei Rudchenko
• Milan Stojanovic, Columbia University
• Christof Teuscher, PSU

Unified English Braille through a Powerful and Responsive eLearning Platform (UEB PREP)

Funding:

• Rehabilitation Services Administration, Department of Education.
• Duration: Sep 1, 2014 – Aug 31, 2019.

Team:

• Holly Lawson, PSU (PI)
• Sam Sennott, PSU (co-PI)
• Christof Teuscher, PSU (co-PI)

Random Automata Architectures

The goal of this project is to assess and design emerging computing architectures based on unstructured physical devices. Molecular and nanoscale electronics seeks to build devices to implement computation by using collections of molecules. It is generally expected that such emerging computing devices will be built in a bottom-up and hierarchical way from vast numbers of simple, densely arranged components that exhibit high failure rates, are relatively slow, and connected in a disordered way

Developmental mechanisms for massive-scale computing assemblies

Nature has evolved multiple adaptation techniques on multiple time-scales, which help organisms to be resilient against changes in the environment. One of the basic mechanisms behind the resilience of biological organisms is cellular division, i.e., the ability of the cells to self-replicate. Self-replication in computing machines has been explore first by John von Neumann in the 1950s, with more recent research in the 1980s by Chris Langton.
The goal of this project is to propose developmental mechanisms that can be applied to future and emerging nano-scale electronics.

Adaptive Control of Self-Assembled Computing Systems

Molecular and nanoscale electronics seeks to build devices to implement computation by using collections of molecules. It is generally expected that such emerging computing devices will be built in a bottom-up and hierarchical way from vast numbers of simple, densely arranged components that exhibit high failure rates, are relatively slow, and connected in a disordered way. Such devices are the prototypical example of complex systems that show emergent behavior not obvious from considering the separate components. They are not programmable by standard means because the reductionist approach fails. The research questions we address are as following: What internal system configuration results in a desired input-output mapping? How can we adapt the system by an algorithm that acts on the control signals to reach such an internal configuration? How can we make the control scalable and robust against certain component failures?