With the growing adoption of privacy-preserving machine learning algorithms, such as Differentially Private Stochastic Gradient Descent (DP-SGD), training or fine-tuning models on private datasets has become increasingly prevalent. This shift has led to the need for models offering varying privacy guarantees and utility levels to satisfy diverse user requirements. Managing numerous versions of large models introduces significant operational challenges, including increased inference latency, higher resource consumption, and elevated costs. Model deduplication is a technique widely used by many model serving and database systems to support high-performance and low-cost inference queries and model diagnosis queries. However, none of the existing model deduplication works has considered privacy, leading to unbounded aggregation of privacy costs for certain deduplicated models and inefficiencies when applied to deduplicate DP-trained models. We formalize the problem of deduplicating DP-trained models for the first time and propose a novel privacy- and accuracy-aware deduplication mechanism to address the problem. We developed a greedy strategy to select and assign base models to target models to minimize storage and privacy costs. When deduplicating a target model, we dynamically schedule accuracy validations and apply the Sparse Vector Technique to reduce the privacy costs associated with private validation data. Compared to baselines, our approach improved the compression ratio by up to 35× for individual models (including large language models and vision transformers). We also observed up to 43× inference speedup due to the reduction of I/O operations.

In recent years, numerous neural network models have been put forth, with an emphasis on the applications of raster imagery and spatiotemporal non-imagery datasets. Implementing these models using existing deep learning frame-works, such as PyTorch and TensorFlow, requires nontrivial coding efforts from the developers although these deep learning frameworks support the implementation of various state-of-the-art machine learning models, such as neural networks, hidden Markov models, and support vector machines. This is due to the fact that the models emphasized on spatiotemporal applications differ extensively from state-of-the-art models supported by existing deep learning frameworks. Moreover, existing deep learning frameworks lack the support for scalable data preprocessing, a mandatory step for converting spatiotemporal datasets into trainable tensors. Considering the limitations of existing deep learning frameworks, we present GeoTorchAI, a framework for deep learning and scalable data processing on raster imagery and spatiotemporal non-imagery datasets. GeoTorchAI enables machine learning practitioners to implement spatiotemporal deep learning models with minimum coding efforts on top of PyTorch. It provides state-of-the-art neural network models, ready-to-use benchmark datasets, and transformation operations for raster imagery and spatiotemporal non-imagery datasets. Besides deep learning, GeoTorchAI contains a data preprocessing module and a DFtoTorch Converter module that enable the formation of trainable spatiotemporal vector datasets and the mapping of preprocessed DataFrames into PyTorch tensors, respectively.

Deep learning frameworks, such as PyTorch and TensorFlow, support the implementation of various state-of-the-art machine learning models such as neural networks, hidden Markov models, and support vector machines. In recent years, many extensions of neural network models have been proposed in the literature targeting the applications of raster and spatiotemporal datasets. Implementing these models using existing deep learning frameworks requires nontrivial coding efforts from the developers because these extensions either are hybrid combinations of various categories of neural network models or differ extensively from state-of-the-art models supported by existing deep learning frameworks. Moreover, existing deep learning frameworks lack the support for scalable data preprocessing required to form trainable tensors from raw spatiotemporal datasets. To enable easy implementation of these neural network extensions, we present GeoTorch, a framework for deep learning and scalable data processing on raster and spatiotemporal datasets. Along with the state-of-the-art spatiotemporal models and ready-to-use benchmark datasets, we propose a data preprocessing module that allows the processing and transformation of spatiotemporal datasets in a cluster computing setting.

Spatial datasets are used extensively to train machine learning (ML) models for applications such as spatial regression, classification, clustering, and deep learning. Most of the real-world spatial datasets are often too large, and many spatial ML algorithms represent the geographical region as a grid consisting of several spatial cells. If the granularity of the grid is too fine, that results in a large number of grid cells leading to long training time and high memory consumption issues during the model training. To alleviate this problem, we propose a machine learning-aware spatial data re-partitioning framework that substantially reduces the granularity of the spatial grid. Our spatial data re-partitioning approach combines fine-grained, adjacent spatial cells from a grid into coarser cells prior to training an ML model. During this re-partitioning phase, we keep the information loss within a user-defined threshold without significantly degrading the accuracy of the ML model. According to the empirical evaluation performed on several real-world datasets, the best results achieved by our spatial re-partitioning framework show that we can reduce the data volume and training time by up to 81%, while keeping the difference in prediction or classification error below 5% as compared to a model that is trained on the original input dataset, for most of the ML applications. Our re-partitioned framework also outperforms the state-of-the-art data reduction baselines by 2% to 20% w.r.t. prediction and classification errors.

Prediction of the next SQL query from the user, given her sequence of queries until the current timestep, during an ongoing interaction session of the user with the database, can help in speculative query processing and increased interactivity. While existing machine learning-- (ML) based approaches use recommender systems to suggest relevant queries to a user, there has been no exhaustive study on applying temporal predictors to predict the next user issued query. In this work, we experimentally compare ML algorithms in predicting the immediate next future query in an interaction workload, given the current user query or the sequence of queries in a user session thus far. As a part of this, we propose the adaptation of two powerful temporal predictors: (a) Recurrent Neural Networks (RNNs) and (b) a Reinforcement Learning approach called Q-Learning that uses Markov Decision Processes. We represent each query as a comprehensive set of fragment embeddings that not only captures the SQL operators, attributes, and relations but also the arithmetic comparison operators and constants that occur in the query. Our experiments on two real-world datasets show the effectiveness of temporal predictors against the baseline recommender systems in predicting the structural fragments in a query w.r.t. both quality and time. Besides showing that RNNs can be used to synthesize novel queries, we find that exact Q-Learning outperforms RNNs despite predicting the next query entirely from the historical query logs.

In this paper, we demonstrate Tabula, a middleware that sits between the data system and the geospatial visualization dashboard to increase user interactivity. The proposed system adopts a sampling cube approach that stores prematerialized spatial samples and allows data scientists to define their own accuracy loss function such that the produced samples can be used for various user-defined visualization tasks. The system ensures that the difference between the sample fed into the dashboard and the raw query answer never exceeds the user-specified loss threshold. For demonstration purposes, we connect Apache Zeppelin, a visualization dashboard, to the system and show how Tabula accelerates interactive visualizations on NYC Taxi Trip data, Yelp review data and San Diego Smart Streetlights data.