mlsanityData to 7 Oct 2026

POET: Training Neural Networks on Tiny Devices with Integrated Rematerialization and Paging

arXiv 2207.07697 ↗First posted 15 Jul 2022cs.LG cs.CV cs.DCcited by 26

Not on the map. Why?

Abstract (arXiv)

Fine-tuning models on edge devices like mobile phones would enable privacy-preserving personalization over sensitive data. However, edge training has historically been limited to relatively small models with simple architectures because training is both memory and energy intensive. We present POET, an algorithm to enable training large neural networks on memory-scarce battery-operated edge devices. POET jointly optimizes the integrated search search spaces of rematerialization and paging, two algorithms to reduce the memory consumption of backpropagation. Given a memory budget and a run-time constraint, we formulate a mixed-integer linear program (MILP) for energy-optimal training. Our approach enables training significantly larger models on embedded devices while reducing energy consumption while not modifying mathematical correctness of backpropagation. We demonstrate that it is possible to fine-tune both ResNet-18 and BERT within the memory constraints of a Cortex-M class embedded device while outperforming current edge training methods in energy efficiency. POET is an open-source project available at https://github.com/ShishirPatil/poet

Key facts. Cited by 26 papers in the map (data through 7 Oct 2026). Builds on 20 of its 40 references that are in the map. First posted on arXiv 15 Jul 2022.

Builds on

20 of 40 references.

References · 20

  1. 2015Deep Residual Learning for Image Recognitioncited by 51,809
  2. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,204
  3. 2014Very Deep Convolutional Networks for Large-Scale Image Recognitioncited by 15,267
  4. 2018The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networkscited by 2,250
  5. 2021EfficientNetV2: Smaller Models and Faster Trainingcited by 1,127
  6. 2016SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model sizecited by 1,054
  7. 2016Training Deep Nets with Sublinear Memory Costcited by 825
  8. 2018ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardwarecited by 666
  9. 2020What is the State of Neural Network Pruning?cited by 578
  10. 2018Federated Learning for Mobile Keyboard Predictioncited by 565
  11. 2019HAWQ: Hessian AWare Quantization of Neural Networks with Mixed-Precisioncited by 387
  12. 2021ZeRO-Offload: Democratizing Billion-Scale Model Trainingcited by 339
8 more
  1. 2022The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrinkcited by 183
  2. 2019Checkmate: Breaking the Memory Wall with Optimal Tensor Rematerializationcited by 90
  3. 2021Federated Evaluation and Tuning for On-Device Personalization: System Design & Applicationscited by 84
  4. 2021ActNN: Reducing Training Memory Footprint via 2-Bit Activation Compressed Trainingcited by 61
  5. 2020Dynamic Tensor Rematerializationcited by 45
  6. 2019E2-Train: Training State-of-the-art CNNs with Over 80% Energy Savingscited by 36
  7. 2019On-Device Neural Net Inference with Mobile GPUscited by 32
  8. 2020Memory Optimization for Deep Networkscited by 15

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2022: 3; 2023: 7; 2024: 7; 2025: 5; 2026: 4.

26 papers.

Sort:

By area:

  1. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,767
  2. 2024Tiny Machine Learning: Progress and Futurescited by 40
  3. 2022On-device Training: A First Overview on Existing Systemscited by 18
  4. 2024Asteroid: Resource-Efficient Hybrid Pipeline Parallelism for Collaborative DNN Training on Heterogeneous Edge Devicescited by 13
  5. 2024Training Machine Learning models at the Edge: A Surveycited by 11
  6. 2022OLLA: Optimizing the Lifetime and Location of Arrays to Reduce the Memory Usage of Neural Networkscited by 6
  7. 2023Colossal-Auto: Unified Automation of Parallelization and Activation Checkpoint for Large-scale Modelscited by 5
  8. 2024On-Device Training of Fully Quantized Deep Neural Networks on Cortex-M Microcontrollerscited by 5
  9. 2025SlimPipe: Memory-Thrifty and Efficient Pipeline Parallelism for Long-Context LLM Trainingcited by 5
  10. 2023Rockmate: an Efficient, Fast, Automatic and Generic Tool for Re-materialization in PyTorchcited by 4
  11. 2023Peer-to-Peer Deep Learning for Beyond-5G IoTcited by 4
  12. 2025FlexInfer: Breaking Memory Constraint via Flexible and Efficient Offloading for On-Device LLM Inferencecited by 4
14 more
  1. 2025FedEL: Federated Elastic Learning for Heterogeneous Devicescited by 3
  2. 2022XEngine: Optimal Tensor Rematerialization for Neural Networks in Heterogeneous Environmentscited by 2
  3. 2023Towards Federated Learning Under Resource Constraints via Layer-wise Training and Depth Dropoutcited by 2
  4. 2023Breaking On-device Training Memory Wall: A Systematic Surveycited by 1
  5. 2024Resource-Efficient Personal Large Language Models Fine-Tuning with Collaborative Edge Computingcited by 1
  6. 2026TrainDeeploy: Hardware-Accelerated Parameter-Efficient Fine-Tuning of Small Transformer Models at the Extreme Edgecited by 1
  7. 2023Moccasin: Efficient Tensor Rematerialization for Neural Networkscited by 0
  8. 2024Privacy-Enhanced Training-as-a-Service for On-Device Intelligence: Concept, Architectural Scheme, and Open Problemscited by 0
  9. 2024Online Resource Allocation for Edge Intelligence with Colocated Model Retraining and Inferencecited by 0
  10. 2025MUDAS: Mote-scale Unsupervised Domain Adaptation in Multi-label Sound Classificationcited by 0
  11. 2025FTTE: Enabling Federated and Resource-Constrained Deep Edge Intelligencecited by 0
  12. 2026What changes after deployment? A survey on On-device Learning in TinyMLcited by 0
  13. 2026QScheduler: Adaptive Gradient Sampling for Zeroth-Order On-Device Training on INT8 NPUscited by 0
  14. 2026GaLe: memory-efficient Global Approximate and Local Exact featurescited by 0