SYMBIOSIS OF HETEROGENEOUS PARALLELISMS

本研究具有双重目的：设计一个实现异构并行共生的仿真器，并提出新的连接学习算法。在仿真器的实现上，使用了两个工作站。一个是SIMD机制，其中模拟了细粒度的并行性。另一种是控制大量并行部分的粗粒度并行。KL1用于这种控制机制。在该共生系统上运行了多重下降代价竞争学习算法。发现由并行性引起的不确定性对于从坏的局部极小值中退出是很有价值的。对于新的学习算法的发展，首席研究员提出了两种主要的新方法。在监督学习上，提出了带附加惩罚的反向传播算法。该算法在权重和输出上包含熵/散度惩罚。实现了网络的修剪、误差的改善和泛化。在无监督情况下，首席调查员创造了和谐竞争学习。该算法能够借助自组织解决多准则优化问题。对数竞争偏差和对数权值突变解决了数据压缩时的局部最优性问题。因此，这个研究项目完成了所声称的结果。

This study has a dual purpose : Designing an emulator which realizes symbiosis of heterogeneous parallelisms and presenting new connectionst learning algorithms. On the realization of the emulator, two workstations are used. One is for an SIMD mechanism where a finegrained parallelism is emulated. The other is for a coarse-grained parallelsm which controls the massive parallel part. KL1 was used for this control mechanism. The multiply descent cost competitive learning algorithm was run on this symbiotic system. The nondeterminism caused by the parallelsm was found to be rather meritorious for the exit from bad local minima.For the developement of new learning algorithms, the head investigator presented two major new methods. On the supervised learning, the backpropagation with additional penalties was presented. This algorithm includes entropy/divergence penalties on the weithts and outputs. Pruning of the network and improvement of errors and generalization were acheived.On the unsupervised case, the head investigator created the harmonic competitive learning. This algorithm enables to solve multiple criteria optimization with the aid of self-organization. The logarithmic competition bias and the logarithmic weight mutation solved the local optimality in the case of data compression.Thus, this research project was completed by accomplishing the claimed results.