CCF: Small: Paradox and Brain-inspired Computer Architecture

为了实现真正的机器智能，必须在其对未来计算机体系结构的潜在影响的背景下理解大脑的结构组织。目前在人工智能和机器学习方面的研究主要集中在创建复杂的程序上，这些程序需要大量的数据集来进行训练，并且没有提供关于大脑结构如何实现智能的见解，利用大量的并行性，权衡实时响应的准确性，以及从相对较少的经验中创建新的解决方案。大脑利用了当前计算机体系结构无法容纳的大量结构并行性。它还必须实时运行，并根据相对较少的先前经验提供最佳答案。为了利用大脑中巨大的结构并行性，最好的可能答案必须从同时持有的一组独立的可能答案中得出。但是不同的答案导致了冲突，这可能是智能的基础，但与传统的计算机体系结构不一致。大脑被组织成意识和潜意识，潜意识利用了最大的并行性。对大脑如何获得智力的洞察可能在于对潜意识中的冲突的检查——多个相互竞争的答案的产生、适应和解决。悖论是一种逻辑结果，它同时产生对和错的结果——这是计算机电路无法适应的。当计算机程序强制解决这种冲突时，大脑可能会持有多个结果而不解决它们，而是持有和发展多个相互冲突的模型，这些模型随着时间的推移不断发展，并根据需要进行选择。大脑不是解决冲突，而是根据情况进行挑选和选择。从根本上说，悖论对同一个问题产生了多个相互矛盾的答案。虽然计算机程序在大量逻辑上运行，但它们无法容纳悖论。他们不是控制冲突，而是强迫对方给出答案。通过关注悖论，大脑的结构属性在计算机架构的背景下可以被检查。因此，大脑结构组织的意义将被更好地理解，未来的计算机体系结构可以利用大脑顺序的并行性，真正的机器智能可以从新的角度来研究。通过关注大脑的一个显著特性，即大脑如何适应悖论，大脑可能像一台MISD（多指令单数据流）计算机一样运行。机器学习的结果是算法解决方案，但并没有深入了解大脑的结构如何实现真正的智能，包括创造力。与此同时，传统的计算机架构已经无法适应人类大脑中大量的并行性和复杂性。MISD计算模型被认为是毫无意义的，因为就像悖论一样，它们会导致多个相互冲突的结果。然而，由于它们产生不同的结果并且不需要收敛，MISD计算可以成为近乎完美并行的基础，因此可能掌握大脑如何适应大程度并行性以及适应冲突的秘密。大脑还必须实时运行，因此可能会对相同的输入集合持有不同的解决方案，其中一些可以适应快速的答案，但总体上不是最好的，但这是生存所必需的。这项工作将侧重于算法/处理器对，如排序和语音识别，其中不同的算法处理相同的输入集，产生不同的结果。然后选择最佳结果作为时间和质量的元组，其中质量是为了时间而牺牲的。例如，一开始可能会对口语单词进行不精确的处理，但总比没有好。然而，更多的思考可能会得到不同的答案。这两种处理模式处理的是同一组输入信息，但分别位于大脑的不同部位，它们的运作方式几乎完全独立。通过这种方式，新的受大脑启发的计算机架构可以导致潜意识包含大量的MISD并行算法，这些算法几乎完全独立地运行，而意识被视为许多可能答案的选择器。由于创造力也需要冲突，这项研究也可能揭示大脑的结构是如何产生创造力的，而创造力是真正的机器智能所必需的。该项目将进行实验，以演示具有多种可能的算法解决方案的问题，这些解决方案根据时间和资源的不同而具有不同程度的准确性。这将形成一个新的架构的高级基础，这个架构可以容纳高度的并行性、时间和质量的权衡、解决冲突，并导致对大脑如何创造的初步研究。将进行实验来说明我们的方法，并将开发架构和方法的泛化。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

In order to achieve true machine intelligence, the structural organization of the brain must be understood in the context of its potential impacts on future computer architecture. Current research in artificial intelligence and machine learning are focused on creating complex programs that requires large datasets for training and do not provide insight as to how the structure of the brain achieves intelligence, utilizes massive amounts of parallelism, trades off accuracy for real-time response, and creates new solutions from relatively few experiences. The brain utilizes large amounts of structural parallelism that current computer architecture cannot accommodate. It must also operate in real-time and provide the best possible answers based upon relatively few prior experiences. In order to utilize the vast structural parallelism in the brain, the best possible answer must be drawn from a set of independent possible answers, held simultaneously. But different answers result in conflict which may be the foundation of intelligence but which is inconsistent with conventional computer architecture. The brain is organized into a conscious and subconscious mind, with the subconscious mind utilizing the greatest amount of parallelism. Insight to how the brain achieves intelligence may lie in an examination of conflict in the subconscious – the production, accommodation and resolution of multiple, competing answers. Paradox is a logical result which produces a simultaneous true and false result – something computer circuitry cannot accommodate. While computer programs force such conflict resolution, the brain likely holds multiple results without resolving them, instead holding and developing multiple, conflicting models which continue to develop, over time, and are selected on an as needed basis. Instead of resolving the conflict, the brain picks and chooses, depending upon circumstances. Fundamentally, paradox produces multiple, conflicting answers to the same question. While computer programs operate on vast amounts of logic, they cannot accommodate paradox. Instead of holding the conflict, they force an answer. By focusing on paradox, the structural properties of the brain in the context computer architecture can be examined. Thus, the significance of the structural organization of the brain will be better understood, future computer architecture can utilize parallelism towards the order of the brain, and true machine intelligence can be studied in a new light. By focusing on a striking property of the brain, that of how the brain accommodates paradox, the brain likely functions as an MISD (Multiple Instruction Single Datastream) computer. Machine Learning results in algorithmic solutions, but does not provide insight into how the structure of the brain achieves true intelligence, including creativity. At the same time, conventional computer architecture has been unable to accommodate the vast amount of parallelism and complexity found in the human brain. MISD models of computing have been considered non-sensical because, like paradox, they result in multiple conflicting results. However, because they produce different results and do not need to converge, MISD computing can be the foundation of nearly perfect parallelism, thus may hold the secret to how the brain accommodates vast degrees of parallelism as well as accommodate conflict. The brain must also operate in real-time, and thus likely holds different solutions to the same set of inputs with some accommodating a quick answer, but not the best overall, but one which is necessary to survival. This work will focus on algorithm/processor pairs such as sort and speech recognition for which different algorithms process the same set of inputs, producing different results. The best result is then chosen as a tuple of time and quality, where quality is sacrificed in the interest of time. For example, the spoken word might be processed imprecisely at first, but in a better than nothing mode. However, more thinking might result in a different answer. These two modes of processing would process the same set of inputs, but be held in different parts of the brain that operate in nearly perfect independence. In this way, novel brain-inspired computer architecture can result in which the subconscious mind contains vast amounts of MISD parallel algorithms that operate in nearly perfect independence, and the conscious mind is viewed as a selector of the many possible answers. Since creativity also requires conflict, this research may also reveal how the brain’s structure results in creativity, which is required for true machine intelligence. This project will conduct experiments to demonstrate problems that have multiple possible algorithmic solutions with varying degree of accuracy dependent upon time and resources. This will form the basis of a high-level foundation for a new architecture that accommodates high degrees of parallelism, accommodates time and quality trade-offs, resolves conflict resolution and leads to an initial investigation of how the brain creates. Experiments will be conducted to illustrate our approach and generalization of the architecture and approach will be developed.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.