object files
**Object Files** are a **cognitive science concept applied to artificial intelligence — discrete internal representations that bind together the distinct attributes (color, shape, position, velocity, identity) of a single entity into a unified, persistent data structure** — enabling neural networks to maintain separate, non-interfering representations for each object in a scene, preventing the catastrophic attribute mixing that occurs when all object information is compressed into a single global feature vector.
**What Are Object Files?**
- **Definition**: Borrowed from cognitive psychology (Kahneman, Treisman & Gibbs, 1992), an object file is a temporary episodic representation that binds together all the properties of a perceived object — its color, shape, location, trajectory, and identity — into a single coherent "file" that persists across time and viewpoint changes. In AI, this concept is implemented as a dedicated vector (slot) per object that is maintained and updated independently.
- **Binding Problem**: The binding problem is the fundamental challenge of associating the correct attributes with the correct objects. When a scene contains a "red circle" and a "blue square," a global feature vector risks confusing attributes — producing hallucinated "red squares" or "blue circles." Object files solve this by maintaining separate representations where each file exclusively owns its object's attributes.
- **Persistence**: Object files persist across time — when a red ball moves behind an occluder and re-emerges, the same object file continues tracking it, providing object permanence. This temporal persistence is critical for video understanding, physical prediction, and interactive planning.
**Why Object Files Matter**
- **Attribute Binding Accuracy**: Global representations (average pooling, CLS tokens) compress all scene information into a single vector, making it impossible to accurately answer "What color is the object left of the cube?" when multiple objects are present. Object files maintain separate attribute bindings, enabling precise per-object queries.
- **Relational Reasoning**: Reasoning about relationships ("Is the red ball above the blue cube?") requires comparing attributes of distinct entities. Object files provide the discrete representations needed for pairwise comparison, unlike global features where entity boundaries are lost.
- **Physical Prediction**: Predicting future states of multi-object scenes (balls bouncing, objects falling) requires tracking each object's position and velocity independently. Object files provide the per-object state vectors that physics prediction networks (Interaction Networks, graph neural networks) operate on.
- **Cognitive Alignment**: Object files align AI representations with human cognitive architecture, enabling more natural human-AI interaction. Humans naturally think in terms of discrete objects with bound properties — AI systems that share this representation can better communicate reasoning processes.
**AI Implementations of Object Files**
| Architecture | Mechanism | Key Property |
|-------------|-----------|--------------|
| **Slot Attention** | Competitive attention assigns pixels to slots | Unsupervised object discovery |
| **RIMs (Recurrent Independent Mechanisms)** | Independent recurrent modules with sparse communication | Modular temporal processing |
| **MONET (Multi-Object Networks)** | VAE with attention-based decomposition | Generative object-centric model |
| **SAVi (Slot Attention for Video)** | Temporal slot attention with optical flow conditioning | Video object tracking |
| **STEVE** | Slot-based transformer encoder for video entities | Scalable video decomposition |
**Object Files** are **digital tracking cards** — maintaining a separate, persistent data folder for every object in the scene, binding attributes to their correct entities and preventing the information mixing that makes global representations unreliable for compositional visual reasoning.