Development of a Bidirectional Associative Memory for Supervised, Unsupervised and Reinforcement Learning

双向联想记忆在监督学习、非监督学习和强化学习中的发展。尽管记忆是我们大多数人每天都不费吹灰之力就能完成的事情，但它实际上是一项极其复杂的任务。例如，存储和检索对象的简单行为是我们还没有设法通过编程让机器达到接近人类效率和健壮性的事情。我的研究旨在更好地理解人类认知系统如何完成从模式中创建（和增强）表示以及识别，识别，分类和分类的复杂任务。为了理解人类认知系统是如何工作的，我们需要建立正式的模型。如今，没有正式的模型可以在不损害简单性和自一致性的情况下考虑到这种行为的多样性。在本研究中，我们将使用并行的人工神经网络，其中信息分布在单元之间。由于大脑对不同的任务使用相似的计算单元和机制，因此该方法使用双向联想记忆（BAM）模型作为通用认知架构的构建块。更准确地说，这个研究项目的重点是开发一个通用的BAM，它可以包含监督学习、无监督学习和强化学习。重点是给出了有趣的非线性动力学系统的观点，其中时间和变化是关键变量。在这个透视图中，记忆被表示为不变或非周期状态。此外，我们希望通过扩展BAM来处理峰值表示，并将模型实现到虚拟和物理机器人中，从而为神经科学建立更好的基础。因此，希望通过开发这样的BAM，我们将更好地了解大脑和思维是如何工作的，以及如何将这种理解转化为人工智能。

Development of a Bidirectional Associative Memory for Supervised, Unsupervised and Reinforcement Learning. Although it is something most of us do every day without effort, memorizing is in fact an incredibly complex task. For instance, the simple act of storing and retrieving an object is something we have not yet managed to program a machine to do with anything approaching human efficiency and robustness. My research aims to better understand how human cognitive system accomplishes the complex task of create (and enhance) representation from patterns as well as recognize, identify, categorize and classify them. To understand how the human cognitive system works, we need to develop formal models. Nowadays, no formal models can take into account this variety of behavior without scarifying simplicity and self-consistency. In this research, we will use artificial neural networks that are parallel and where information is distributed among the units. Since the brain uses similar computing units and mechanisms for different tasks, this approach uses a bidirectional associative memory (BAM) model as building blocks for a general cognitive architecture. More precisely, this research program focuses on developing a general BAM that can encompass supervised, unsupervised and reinforcement learning. Focus is given on interesting nonlinear dynamics system perspective where time and change are the key variables. Within that perspective memory are represented as invariant or aperiodic states. Moreover, we want to build a better grounding into neuroscience by extending the BAM to handle spiking representation and implement the model into virtual and physical robots. Therefore, it is hoped that by developing such BAM we will have a better understanding on how the brain and mind work, and, how this understanding can be translated into artificial general intelligence.