Introduction
Modern Artificial Intelligence, particularly Deep Learning, is currently dependent on "Big Data." This reliance on massive, static data collections creates several systemic issues. Centralization of power occurs because only a few massive corporations and governments can afford the infrastructure to collect, store, and process gargantuan datasets. Privacy risks are inherent when personal sensor data (cameras, microphones) is stored for training. Furthermore, training Large Language Models (LLMs) is energy-intensive due to the iterative processing of static collections. Biologically, this diverges from how animals learn—not by "downloading" a shuffled database, but through continuous, temporal interaction with their environment. This paper proposes "Collectionless AI," a paradigm shift that decouples intelligence from data accumulation by enforcing a strict "no-storage" rule.
The Collectionless Principle
The core of this framework is the Collectionless Principle, defined by three fundamental "Rules of the Game":
- Data Collections are Banned: Agents access data only as sequential, one-way streams. They cannot store these streams for offline training or re-shuffling.
- No Sensor Storage: Agents must process information at the moment of acquisition; they cannot record temporal information from sensors for subsequent usage.
- No Internal Sequence Storage: The restriction extends to internal states; an agent cannot store a history of its own internal representations to simulate a memory of raw data.
Technically, this imposes an O(1) space complexity for buffering sensorial data. The memory used to hold incoming data must remain constant, regardless of the learning duration. Agents are expected to develop "self-organized memorization skills" at an abstract level, such as updating neural weights, rather than relying on a database.
Evaluation and Benchmarking
As a position paper, it moves away from standard static benchmarks like ImageNet or COCO. Instead, it proposes a new approach to evaluation:
- "En Plein Air" Evaluation: Inspired by Impressionist painters, the authors suggest evaluating agents in the wild, observing their behavior and progress in real-time rather than on a fixed test set.
- Signal Generation by Target Emulation: A proposed task where an agent must learn to mimic a complex signal (like music or a wave) from a live stream without storing the history of that signal.
- Extreme Collectionless Signal Generation (ECSG): A benchmark where the buffer size is exactly one sample ($n=1$), forcing the model to update its internal state instantaneously.
Conceptual Pillars
Collectionless AI redefines the roles of environment, time, and learning to align more closely with natural intelligence.
- Time as the Protagonist: In standard ML, time is often discarded through shuffling. In Collectionless AI, the temporal order of data is the primary source of information, mirroring how humans perceive the world.
- The "Trump Card": While the renunciation of memory seems like a handicap, it forces the development of models that are naturally better at personal interaction, customizability, and causal reasoning.
- Unification of AI: This approach bridges the gap between Symbolic AI (traditionally collectionless and logic-based) and Sub-symbolic AI (neural networks), placing them under the same umbrella.
- Geopolitical Impact: By moving computation to "thin clients" (edge devices) that don't need to communicate with a central cloud, the concentration of power in large data silos is naturally reduced.
Conclusion and Future Outlook
The paper formalizes the collectionless protocol, demonstrating that tasks like autonomous signal generation and unsupervised optical flow can be mathematically formulated within these constraints. By design, privacy, control, and customizability are integrated into the architecture rather than being added as afterthoughts.
However, the authors acknowledge the "Slowdown Challenge": learning in a truly online, lifelong manner is significantly slower than the high-speed batch processing used today. Furthermore, the computational complexity of organizing information dynamically without a buffer remains a significant research hurdle. Transitioning the AI community from static datasets to "en plein air" evaluation requires a total overhaul of measurement standards. Ultimately, Collectionless AI calls for a return to the roots of intelligence—interaction with a changing world in real-time.
Summary updated