研究深度生成模型的内在自上而下的动态
Lorenzo Tausani1,2, Alberto Testolin3,4, Marco Zorzi5,6
1Department of General Psychology and Padova Neuroscience Center, University of Padova, Padova, Italy.
Scientific reports
|January 22, 2025
概括
像代深信网络 (iDBN) 这样的层次生成模型可以生成多样化的数据原型. 从"奇默状态"开始生成增强了这种多样性,支持持续学习,并提供了对大脑动态的洞察力.
科学领域:
- 计算神经科学是一种计算神经科学.
- 深度学习是一种深度学习.
- 人工智能的人工智能是人工智能.
背景情况:
- 层次生成模型学习数据分布,并可以解释自发的大脑活动.
- 深度信念网络 (DBN) 是无监督的,基于能量的模型学习等级表示.
- 当前的理论将静止状态大脑活动与自上而下的生成模型动力学联系起来.
研究的目的:
- 调查代深信网络 (iDBN) 的生成动态.
- 探索在层次模型中增强数据生成多样性的方法.
- 将生成模型动力学与神经认知发展理论联系起来.
主要方法:
- 在手写数字和面部上训练代的深信网络 (iDBN) 模型.
- 分析了自上而下的抽样动态和国家访问.
- 实验了从"奇美拉状态" (组合的高层特征) 的初始化生成.
- 将iDBN动态与一个浅的受限制的博尔兹曼机器进行比较.
主要成果:
- 从"奇米拉状态"进行iDBN采样的初始化增加了生成数据原型的多样性.
- 与浅层模型相比,iDBN表现出更丰富的自上而下的动态.
- 生成的样本支持通过生成重复进行持续学习.
- 模型动态受到能量函数形状,架构深度和数据结构的影响.
结论:
- 代深信网络 (iDBN) 通过偏向的初始状态显示出增强的生成能力.
- 该研究提供了一个计算框架,将生成模型与大脑功能和发育联系起来.
- 研究结果表明,iDBN在持续学习和理解神经动态方面有潜力.
相关概念视频
01:19Graded Potential
3.7K
Graded potentials are localized fluctuations in the cell membrane's electrical charge, commonly found in the dendrites of neurons. The magnitude of these potential changes depends on the strength of the initiating stimulus. In a membrane at its resting potential, a graded potential signifies a voltage shift either above -70 mV or below -70 mV.
Graded potentials fall into two categories: depolarizing and hyperpolarizing. Depolarizing graded potentials typically occur when sodium (Na+) or...
Graded potentials fall into two categories: depolarizing and hyperpolarizing. Depolarizing graded potentials typically occur when sodium (Na+) or...
3.7K
01:14Steps in the Modeling Process
190
Albert Bandura's theory of observational learning identifies four critical processes: attention, retention, motor reproduction, and reinforcement or motivation.
Attention is the first necessary component for observational learning. It involves focusing on what the model is doing and saying. For example, if you decide to take a drawing class to enhance your skills, you need to pay close attention to the instructor's words and hand movements. The characteristics of the model significantly...
Attention is the first necessary component for observational learning. It involves focusing on what the model is doing and saying. For example, if you decide to take a drawing class to enhance your skills, you need to pay close attention to the instructor's words and hand movements. The characteristics of the model significantly...
190