Print this Page

Overview

Research activities at the Adaptive Systems Laboratory (ASLab) are focusing on computer and system architectures, which are expanding beyond the usual focus on performance to have other quantitative and qualitative criteASL_research_Intro_June2016ria. The quantitative criteria mainly include energy/power consumption. The qualitative criteria include adaptability, efficiency (e.g. performance/area or performance/energy), and reliability which is important in: (1) dynamic environments, where physical context, network topologies, and workloads are always changing, and (2) harsh environments, such as an environment with high radiation (i.e. space missions). There are several enabling techniques or paradigms, such as reconfigurable fabric, ASIP, and neuro-inspired, that allow a computing system to perform such an adaptation to achieve the criteria mentioned above. This is achieved if such adaptive computing systems can monitor themselves, analyze their behavior, learn and adapt to several execution environments while keeping the system’s complexity invisible to the user.

We are engaged in research and development of adaptive computing systems for incremental learning and adaptation in dynamic and harsh environments targeted for a wide range of applications, including mobile robots, IoT, and motion analysis for activity recognition. Our current interests include the following computing, power, and communication areas:

Adaptive Neuro-inspired Computing

The biological brain implements parallel computations using a complex structure that is different from the conventional von Neumann or load/store computing style. Our brain is a low-power, fault-tolerant, and high-performance machine! It consumes only about 20W and brain circuits continue to operate as the organism needs even when the circuit (neuron, neuroglia, etc.) is perturbed or died. Computations in neural networks are naturally parallel and distributed among billion neurons. This very high degree of distributed parallelism allows us to design a low-clock frequency (low-power) and high-throughput neuro-inspired circuits and systems. Hardware implementations of neural networks are very efficient and effective methods to provide cognitive functions on a chip compared with conventional von Neumann processors. The challenges that need to be solved toward reaching such a computing system include building a small-size massively parallel architecture with scalable interconnects, low-power consumption, and reliable neuro-inspired computing schemes for implementation of learning in hardware.

Our focus in this area is to develop a novel ultra-low power massively-parallel full-custom neuro-inspired hardware and platform able to scale up to biological levels to address the full spectrum of applications from autonomous objects (i.e. IoT) to high-performance computing co-processing. Currently, we are studying the following challenges:

  • Algorithm, and architecture co-optimization for efficient machine-learning hardware design
  • Models for neurons and synapses;
  • Spiking neuro-inspired architectures building blocks;
  • Reliable communication networks for neuro-inspired chips/systems;
  • Reconfigurability and adaptability methods;
  • On-line/On-chip learning schemes
  • Deep learning models;
  • Conventional hardware (i.e. VLSI, FPGAs) and innovative hardware (i.e., memristor) implementation of Neuro-inspired systems;
  • Software simulation and programming methodologies for Neuro-inspired computing systems;
  • New applications of on-chip learning (i.e., mobile devices, IoT).

Reliable Scalable On-chip Interconnects 

Future computing systems would contain hundreds of components made of processor cores, DSPs, memory, accelerators, and I/O all integrated into a single die area of just a few square millimeters. Such complex system would be interconnected via a novel on-chip interconnect (Network-on-Chip) closer to a sophisticated network than to current bus-based solutions. This network must provide high throughput and low latency while keeping area and power consumption low.

Our research effort in this area is about solving several design challenges to enable the packet-switched and other novel switching schemes in multi and many-core systems/SoCs. In particular, we are investigating the following topics:

  • Implementation techniques for TSV based NoCs
  • 3D-IC integration
  • Fault-tolerant and reliability issues
  • New topologies and flow-control techniques
  • Photonic Interconnects

=> Current Projects [現在のプロジェクト]

Permanent link to this article: http://adaptive.u-aizu.ac.jp/?page_id=668