90 likes | 277 Views
Engineering Next-generation Self-healing And Self-optimizing Neural Network Based Medical Platforms. Zhanpeng Jin Allen C. Cheng zhj6@pitt.edu acc33@pitt.edu. ASPLOS 2010, The Wild and Crazy Session VIII. Artificial Neural Network.
E N D
Engineering Next-generation Self-healing And Self-optimizing Neural Network Based Medical Platforms Zhanpeng Jin Allen C. Cheng zhj6@pitt.eduacc33@pitt.edu ASPLOS 2010, The Wild and Crazy Session VIII
Artificial Neural Network (Source: "Anatomy and Physiology" by the US National Cancer Institute's Surveillance, Epidemiology and End Results (SEER) Program.) (Source: CapitalISM BI)
Emerging Neural Hardware Neural chip (384 neurons + 100,000 synapses) @ FACET Project Microchip of a neural network @ RIEC, Tohoku University The China-Brain Project (10,000 – 15,000 neural nets) @ Hugo de Garis 3-D neural chip for better visualization @ CalTech
Bio-inspired Autonomously Reconfigurable Mechanism Autonomous Reconfigurability Topology Adaptation Redundancy (Picture Source: “Brain Injury: Recovery” in Psychology Wiki)
Autonomously Reconfigurable ANN µController Run-Time Reconfiguration Controller Flash Memory Coarse-grained Reconfiguration Bitstream Database Error Scale and Location Mask-Based Topology Adaption New topology Error Detection Request Coarse-/Fine-Grained Hybrid Reconfiguration Request Current ANN Configuration Fine-grained Reconfiguration Bitstream Coarse-grained Reconfiguration Bitstream Data Traffic Select/Merge Bitstreams New Configuration Inputs Sensors/ Database Autonomously Reconfigurable ANN (ARANN) Based Medical Processing Platforms Outputs Diagnosis/ Controllers
Virtual-physical Neuron Mapping Neural Topology Physical Neurons Virtual-to-Physical (V2P) Neuron Mapping Connectionism Evolvement (Fine-grained) = Topology Adaptation (Coarse-grained) + Autonomous ANN
WACI Conclusion • Systems are increasingly vulnerable to unexpected faults and defects, especially for emerging biomedical systems. • Non-invasive autonomous reconfigurability is promising, particularly for ANN-based biomedical platforms. • Autonomously adapting ANN’s behaviors and structures, both algorithmically and microarchitecturally. • Neuron Virtualizationhelps to decouple the fault scale and reduce the reconfiguration latency (idle time). • Mask-based Topology Adaptation can achieve significant reduction of design complexity and spatial overhead.
Thanks for Listening Questions? (This work is supported by NSF No. 0832990)