There is a widening gap between teams that wait for data to accumulate before analysis and those that capture distributional shifts the moment they occur. Forcing legacy algorithms optimized for static datasets into real-time streaming environments inevitably leads to memory bloat and unacceptable latency. In the realm of Optimal Transport (OT), this distinction is no longer just a performance metric—it is a prerequisite for system stability.
The Reliability of Sliced Optimal Transport
Traditional Wasserstein distance is a robust mathematical tool for comparing probability distributions, but its computational cost in high-dimensional spaces is notoriously high. To bypass this, researchers introduced the Sliced Wasserstein (SW) distance. By projecting high-dimensional data onto random 1D lines, SW transforms a complex optimization problem into a simple sorting task.
Developers embraced SW because 1D optimal transport is solved simply by sorting samples, which drastically reduces complexity. For offline training or static generative modeling, SW was the gold standard—offering mathematical rigor without the overhead of the full Monge-Kantorovich problem. At the time, the assumption that the entire dataset would reside in memory for sorting was rarely questioned.
The Hidden Cost of Batching in Production
However, the paradigm shifts when data flows continuously. In environments like IoT sensor monitoring, real-time log analysis, or high-frequency trading, the concept of a "complete dataset" is nonexistent. To apply traditional SW in these scenarios, one must store every incoming sample, leading to a memory complexity of $O(N)$.
As $N$ grows into the millions, the system eventually hits a wall. Re-sorting the entire history every time a new sample arrives creates a latency spike that compounds over time. In my own experience building real-time anomaly detection, I often found myself trapped in a trade-off: small windows led to poor statistical power, while large windows caused the service to crash. This dilemma is an inherent flaw of batch-centric algorithms applied to streaming problems.
Streaming Sliced Wasserstein: Efficiency in Motion
Streaming Sliced Wasserstein (Stream-SW) was proposed to break this cycle. The core innovation lies in updating the SW estimate incrementally without storing the entire sample history. Instead of relying on a full sort of all past data, it utilizes techniques to maintain a running estimate of the distribution's projections.
From a resource perspective, the advantage is clear: memory usage remains constant or near-constant regardless of how much data has passed through the system. By replacing $O(N)$ storage with a streamlined update mechanism, Stream-SW ensures that the system's footprint doesn't expand indefinitely. This enables sophisticated distribution comparison even on edge devices with limited RAM, moving the computation to where the data lives.
Practical Hurdles and Migration Strategy
Transitioning from batch-based SW to Stream-SW requires a strategic approach. The first "gotcha" is the approximation error. Since you are no longer looking at the global dataset, the estimate will have some variance. It is crucial to monitor how quickly the algorithm adapts to "concept drift"—sudden changes in the underlying data distribution.
Secondly, manage your projection directions carefully. The accuracy of any sliced method depends on the number of projections used. In a streaming context, you must decide whether to keep these projections fixed or refresh them periodically. In my view, the most stable approach is to start with a fixed set of projections to establish a baseline and then introduce dynamic updates only if the data's dimensionality or complexity demands it.
Final Thoughts: From Accumulation to Flow
To be blunt, Stream-SW is not a silver bullet for every use case. If your data is static and your training is offline, the simplicity of batch SW is still preferable. But if you are managing a live pipeline where infrastructure costs and real-time responsiveness are paramount, the shift is non-negotiable.
Machine learning is evolving from a discipline of "collect and process" to one of "observe and adapt." Efficiency is no longer just about raw FLOPS; it's about how gracefully an algorithm handles the relentless flow of information. Check your memory profiles today. If your RAM usage is creeping upward as your data grows, you've found your primary candidate for a streaming upgrade.
Reference: arXiv CS.LG (Machine Learning)