Hardware architecture of the Protein Processing Associative Memory and the effects of dimensionality and quantisation on performance

TY - JOUR

T1 - Hardware architecture of the Protein Processing Associative Memory and the effects of dimensionality and quantisation on performance

AU - Qadir, Omer

AU - Lenz, Alex

AU - Tempesti, Gianluca

AU - Timmis, Jon

AU - Pipe, Tony

AU - Tyrrell, Andy

PY - 2014/9

Y1 - 2014/9

N2 - The Protein Processor Associative Memory (PPAM) is a novel hardware architecture for a distributed, decentralised, robust and scalable, bidirectional, hetero-associative memory, that can adapt online to changes in the training data. The PPAM uses the location of data in memory to identify relationships and is therefore fundamentally different from traditional processing methods that tend to use arithmetic operations to perform computation. This paper presents the hardware architecture and details a sample digital logic implementation with an analysis of the implications of using existing techniques for such hardware architectures. It also presents the results of implementing the PPAM for a robotic application that involves learning the forward and inverse kinematics. The results show that, contrary to most other techniques, the PPAM benefits from higher dimensionality of data, and that quantisation intervals are crucial to the performance of the PPAM.

AB - The Protein Processor Associative Memory (PPAM) is a novel hardware architecture for a distributed, decentralised, robust and scalable, bidirectional, hetero-associative memory, that can adapt online to changes in the training data. The PPAM uses the location of data in memory to identify relationships and is therefore fundamentally different from traditional processing methods that tend to use arithmetic operations to perform computation. This paper presents the hardware architecture and details a sample digital logic implementation with an analysis of the implications of using existing techniques for such hardware architectures. It also presents the results of implementing the PPAM for a robotic application that involves learning the forward and inverse kinematics. The results show that, contrary to most other techniques, the PPAM benefits from higher dimensionality of data, and that quantisation intervals are crucial to the performance of the PPAM.

KW - Protein processing

KW - PPAM

KW - FGPA

KW - Associative memory

KW - BERT2

KW - Inverse kinematics

KW - Dimensionality

KW - Quantisation

KW - Non-standard computation

U2 - 10.1007/s10710-014-9217-1

DO - 10.1007/s10710-014-9217-1

M3 - Article

SN - 1389-2576

VL - 15

SP - 245

EP - 274

JO - Genetic programming and evolvable machines

JF - Genetic programming and evolvable machines

IS - 3

ER -