On the Fly Sampling

The area of on-the-fly sampling, compression of EO/IR Video, and RF-ITV Streams with targets, necessitates efficient and flexible methodologies that adapt to the nature of the data and the underlying dynamics of the task. Within this context, our research has developed two innovative methods that have relevance to this problem domain, even though they have not been directly applied to EO/IR video and RF-ITV streams. Both contributions represent significant advancements in optimization and data handling that align with the requirements of on-the-fly sampling and compression ¹. The integration of these methods into the specific domain of EO/IR video and RF-ITV streams warrants further investigation, potentially leading to substantial improvements in efficiency, flexibility, and robustness of the associated processes.

A Particle-based Sparse Gaussian Process Optimizer

The first method introduced in this domain is a Particle-based Sparse Gaussian Process Optimizer (Bajaj et al., 2022). Traditional swarm-based optimization methods, while successful in many applications, can suffer inefficiency or local-minima entrapment. In order to overcome these challenges, we developed a new framework that utilizes Gaussian Process Regression to gain a deeper understanding of the underlying dynamical process of descent.

This method offers greater exploration around the current state before determining the direction of descent. Such an explorative approach has the potential to escape from local minima, allowing for more robust optimization. We empirically showed the superiority of our method in non-convex optimization scenarios and tested its applicability in high-dimensional spaces, such as image classification. Further, it has a potential application to the sampling and compression of EO/IR Video and RF-ITV Streams.

Learning Generative Embeddings using an Optimal Subsampling Policy for Tensor Sketching

The second contribution to on-the-fly sampling is our research "Learning Generative Embeddings using an Optimal Subsampling Policy for Tensor Sketching" (Bajaj et al., 2022). With data tensors of order 3 and above becoming increasingly prevalent, particularly in fields like images, videos, and geographic data, accessing such large data collections has become increasingly prohibitive.

Our proposed method learns approximate full-rank and compact tensor sketches, providing efficient space, time, and spectral embeddings. By constructing tensor sketches from a sample-efficient sub-sampling of tensor slices, and employing an adaptable stochastic Thompson sampling with Dirichlet distributions and conjugate priors, we manage to produce optimal rank-r Tucker decompositions.

The implications of this method in the context of on-the-fly sampling and compression of EO/IR Video and RF-ITV Streams can be seen in the efficient handling and querying of large data collections. The use of generative sketches allows for streamlined processing without losing critical information, a crucial aspect in the real-time handling of such complex data.

References

• Bajaj, C., Heo, T., & Avlur, R. (2022). Learning Generative Embeddings using an Optimal Subsampling Policy for Tensor Sketching. ArXiv Preprint ArXiv: 2209.00372.

• Bajaj, C., Vaidya, O. B., & Wang, Y. (2022). A Particle-based Sparse Gaussian Process Optimizer. ArXiv Preprint ArXiv:2211.14517.