解开视觉关系推理的几何
Jiaqi Shang1, Gabriel Kreiman2,3,4, Haim Sompolinsky4,5
1Program in Neuroscience, Harvard Medical School, Boston, Massachusetts & 02115, United States.
ArXiv
|March 10, 2025
概括
神经网络与抽象关系的泛化作斗争,不像人类. 散射组合学习者 (SCL) 架构在新的基准上显示了最类似人类的性能,为AI推理提供了几何洞察力.
科学领域:
- 人工智能的人工智能
- 认知科学 认知科学
- 计算机视觉 计算机视觉
背景情况:
- 人类擅长抽象的关系推理,这种能力仍然是当前人工神经网络面临的挑战.
- 现有的基准没有充分捕捉到人类类AI所需的抽象关系概括的细微差别.
研究的目的:
- 引入一个新的基准,简化RPM,用于系统地评估神经网络如何将抽象关系概括.
- 将不同神经网络架构的性能与人类关系推理能力进行比较.
- 为预测概括性能的神经表示提供几何洞察力.
主要方法:
- 开发简化RPM基准和并行的人体实验,以建立关系难度基线.
- 评估了四种不同的神经网络架构:ResNet-50,视觉转换器,野生关系网络和散射组合学习器 (SCL).
- 对表示几何学和层级策略的分析,以了解概括机制,包括提出一个新的目标函数,SNRloss.
主要成果:
- 分散组合学习器 (SCL) 在测试的架构中展示了与人类行为和优越的概括能力的最佳对齐.
- 确定了与模型相关的表示几何学,并预测了跨模型的概括性能.
- 层次分析揭示了各种推理策略,以及在训练对齐子空间内压缩未见的规则表示.
结论:
- 分散组合学习器 (SCL) 为开发具有更类似人类抽象关系推理的AI系统提供了一个有希望的方向.
- 对神经表示的几何洞察力对于理解和改进AI中的概括至关重要.
- 拟议的SNRloss目标函数有助于平衡表示几何,以增强AI推理能力.
相关概念视频
01:16Deductive Reasoning
54.8K
Deductive reasoning, or deduction, is the type of logic used in hypothesis-based science. In deductive reasoning, the pattern of thinking moves in the opposite direction as compared to inductive reasoning, which means that it uses a general principle or law to predict specific results. From those general principles, a scientist can deduce and predict the specific results that would be valid as long as the general principles are valid.
For example, a researcher can deduce specific predictions...
For example, a researcher can deduce specific predictions...
54.8K
01:12Visual Agnosia
173
Visual agnosia is a condition characterized by the inability to recognize visually presented objects despite having normal vision. For instance, a person with visual agnosia can describe the shape and color of an object but cannot identify or name it. This impairment does not affect their visual field, acuity, color vision, brightness discrimination, language, or memory. An example of this condition in a social setting is someone at a dinner party asking for "that silver thing with a round...
173
01:21Gestalt Principles of Perception
268
Gestalt principles provide a framework for understanding how humans perceive objects as unified wholes within their context. These principles are essential in explaining the cognitive processes that make sense of complex visual stimuli by organizing them into coherent groups. One fundamental principle is proximity, which posits that objects located close to each other are perceived as a collective group. For instance, when dots are positioned near one another, the visual system interprets them...
268
02:57Coordination Number and Geometry
15.4K
For transition metal complexes, the coordination number determines the geometry around the central metal ion. Table 1 compares coordination numbers to molecular geometry. The most common structures of the complexes in coordination compounds are octahedral, tetrahedral, and square planar.
15.4K
00:59Inductive Reasoning
59.8K
Inductive reasoning is a form of logical thinking that uses related observations to arrive at a general conclusion. It is uncertain and operates in degrees to which the conclusions are credible. As such, inductive arguments can be weak or strong, rather than valid or invalid, and conclusions can be used to formulate testable, falsifiable hypotheses.
Inductive reasoning is common in descriptive science. A life scientist makes observations and records them. This data can be qualitative or...
Inductive reasoning is common in descriptive science. A life scientist makes observations and records them. This data can be qualitative or...
59.8K
01:15Depth Perception and Spatial Vision
503
Depth perception is the ability to perceive objects three-dimensionally. It relies on two types of cues: binocular and monocular. Binocular cues depend on the combination of images from both eyes and how the eyes work together. Since the eyes are in slightly different positions, each eye captures a slightly different image. This disparity between images, known as binocular disparity, helps the brain interpret depth. When the brain compares these images, it determines the distance to an object.
503