Research map / Architectures and efficiency / RNNs and normalization

SuBiTO: Synopsis-based Training Optimization for Continuous Real-Time Neural Learning over Big Streaming Data

AAAI 2025 · Architectures and efficiency › RNNs and normalization · approach: Efficient network compression and training · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Presents SuBiTO, a demo framework that uses Bayesian optimization over a Gaussian Process to tune NN architecture, stream synopsis size, and epochs for continuous real-time training on streaming data, with concept drift detection.

Problem

Continuously training neural networks on big, high-velocity data streams requires balancing training time against accuracy, and existing sampling-based speedups provide no a priori known accuracy vs training-time trade-off for continuously trained NNs.

Approach

Automatically learn accuracy vs training-time trade-offs as data streams in by running Bayesian optimization trials over NN layer configuration, stream synopsis parameters (load shedding), and epoch count, triggered on KL-divergence concept drift detection, and let a human operator pick a top-ranked parameter set to deploy on-the-fly.

Key results

• No quantitative results are reported; the paper is a demonstration description of the SuBiTO dashboard and architecture.
• Optimizer scores parameter sets with Score(c) = λ·accuracy(c) − (1−λ)·tanh(training_time(c)/TargetLatency − 1) and the operator deploys one of the top-3 alternatives.
• Demonstration uses NSFW Detect and UCF50 real-world labeled streams for the Training Pipeline.

Limitations

• No quantitative evaluation or comparison against baselines is presented.
• Concept drift detection currently relies only on KL-divergence.
• Prediction Pipeline streams are simulated via a GAN or dataset split rather than live data.

Experimental setup

Datasets: NSFW Detect (HuggingFace) UCF50

Metrics: accuracy training time epoch duration training loss

Baselines: sampling-based NN training speedups (Zhong et al. 2023; Wang et al. 2018; Huang

Builds on: Bayesian optimization Kafka

Closest papers by meaning

• Improved Algorithms for Neural Active Learning (NeurIPS 2022 · similarity 0.84)
• Algorithm Development in Neural Networks: Insights from the Streaming Parity Task (ICML 2025 · similarity 0.83)
• Don’t Think It Twice: Exploit Shift Invariance for Efficient Online Streaming Inference of CNNs (AAAI 2025 · similarity 0.83)
• Learning from One Continuous Video Stream (CVPR 2024 · similarity 0.82)
• MIND: Multi-Task Incremental Network Distillation (AAAI 2024 · similarity 0.82)
• TinySubNets: An Efficient and Low Capacity Continual Learning Strategy (AAAI 2025 · similarity 0.82)