skip to main content


The NSF Public Access Repository (NSF-PAR) system and access will be unavailable from 11:00 PM ET on Friday, April 12 until 2:00 AM ET on Saturday, April 13 due to maintenance. We apologize for the inconvenience.

Title: Rep-Net: Efficient On-Device Learning via Feature Reprogramming
Transfer learning, where the goal is to transfer the well-trained deep learning models from a primary source task to a new task, is a crucial learning scheme for on-device machine learning, due to the fact that IoT/edge devices collect and then process massive data in our daily life. However, due to the tiny memory constraint in IoT/edge devices, such on-device learning requires ultra-small training memory footprint, bringing new challenges for memory-efficient learning. Many existing works solve this problem by reducing the number of trainable parameters. However, this doesn't directly translate to memory-saving since the major bottleneck is the activations, not parameters. To develop memory-efficient on-device transfer learning, in this work, we are the first to approach the concept of transfer learning from a new perspective of intermediate feature reprogramming of a pre-trained model (i.e., backbone). To perform this lightweight and memory-efficient reprogramming, we propose to train a tiny Reprogramming Network (Rep-Net) directly from the new task input data, while freezing the backbone model. The proposed Rep-Net model interchanges the features with the backbone model using an activation connector at regular intervals to mutually benefit both the backbone model and Rep-Net model features. Through extensive experiments, we validate each design specs of the proposed Rep-Net model in achieving highly memory-efficient on-device reprogramming. Our experiments establish the superior performance (i.e., low training memory and high accuracy) of Rep-Net compared to SOTA on-device transfer learning schemes across multiple benchmarks.  more » « less
Award ID(s):
1931871 2144751
Author(s) / Creator(s):
; ;
Date Published:
Journal Name:
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
Page Range / eLocation ID:
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Nowadays, one practical limitation of deep neural network (DNN) is its high degree of specialization to a single task or domain (e.g., one visual domain). It motivates researchers to develop algorithms that can adapt DNN model to multiple domains sequentially, while still performing well on the past domains, which is known as multi-domain learning. Almost all conventional methods only focus on improving accuracy with minimal parameter update, while ignoring high computing and memory cost during training, which makes it difficult to deploy multi-domain learning into more and more widely used resource-limited edge devices, like mobile phone, IoT, embedded system, etc. During our study in multi-domain training process, we observe that large memory used for activation storage is the bottleneck that largely limits the training time and cost on edge devices. To reduce training memory usage, while keeping the domain adaption accuracy performance, we propose Dynamic Additive Attention Adaption (DA3), a novel memory-efficient on-device multi-domain learning method. DA3 learns a novel additive attention adaptor module, while freezing the weights of the pre-trained backbone model for each domain. Differentiating from prior works, such module not only mitigates activation memory buffering for reducing memory usage during training, but also serves as a dynamic gating mechanism to reduce the computation cost for fast inference. We validate DA3 on multiple datasets against state-of-the-art methods, which shows great improvement in both accuracy and training time. Moreover, we deployed DA3 into the popular NIVDIA Jetson Nano edge GPU, where the measured experimental results show our proposed \mldam reduces the on-device training memory consumption by 19x-37x, and training time by 2x, in comparison to the baseline methods (e.g., standard fine-tuning, Parallel and Series Res. adaptor, and Piggyback). 
    more » « less
  2. Transfer learning on edge is challenging due to on-device limited resources. Existing work addresses this issue by training a subset of parameters or adding model patches. Developed with inference in mind, Inverted Residual Blocks (IRBs) split a convolutional layer into depthwise and pointwise convolutions, leading to more stacking layers, e.g., convolution, normalization, and activation layers. Though they are efficient for inference, IRBs require that additional activation maps are stored in memory for training weights for convolution layers and scales for normalization layers. As a result, their high memory cost prohibits training IRBs on resource-limited edge devices, and making them unsuitable in the context of transfer learning. To address this issue, we present MobileTL, a memory and computationally efficient on-device transfer learning method for models built with IRBs. MobileTL trains the shifts for internal normalization layers to avoid storing activation maps for the backward pass. Also, MobileTL approximates the backward computation of the activation layer (e.g., Hard-Swish and ReLU6) as a signed function which enables storing a binary mask instead of activation maps for the backward pass. MobileTL fine-tunes a few top blocks (close to output) rather than propagating the gradient through the whole network to reduce the computation cost. Our method reduces memory usage by 46% and 53% for MobileNetV2 and V3 IRBs, respectively. For MobileNetV3, we observe a 36% reduction in floating-point operations (FLOPs) when fine-tuning 5 blocks, while only incurring a 0.6% accuracy reduction on CIFAR10. Extensive experiments on multiple datasets demonstrate that our method is Pareto-optimal (best accuracy under given hardware constraints) compared to prior work in transfer learning for edge devices. 
    more » « less
  3. The high energy cost of processing deep convolutional neural networks impedes their ubiquitous deployment in energy-constrained platforms such as embedded systems and IoT devices. This article introduces convolutional layers with pre-defined sparse 2D kernels that have support sets that repeat periodically within and across filters. Due to the efficient storage of our periodic sparse kernels, the parameter savings can translate into considerable improvements in energy efficiency due to reduced DRAM accesses, thus promising significant improvements in the trade-off between energy consumption and accuracy for both training and inference. To evaluate this approach, we performed experiments with two widely accepted datasets, CIFAR-10 and Tiny ImageNet in sparse variants of the ResNet18 and VGG16 architectures. Compared to baseline models, our proposed sparse variants require up to ∼82% fewer model parameters with 5.6× fewer FLOPs with negligible loss in accuracy for ResNet18 on CIFAR-10. For VGG16 trained on Tiny ImageNet, our approach requires 5.8× fewer FLOPs and up to ∼83.3% fewer model parameters with a drop in top-5 (top-1) accuracy of only 1.2% ( ∼2.1% ). We also compared the performance of our proposed architectures with that of ShuffleNet and MobileNetV2. Using similar hyperparameters and FLOPs, our ResNet18 variants yield an average accuracy improvement of ∼2.8% . 
    more » « less
  4. null (Ed.)
    User authentication is a critical process in both corporate and home environments due to the ever-growing security and privacy concerns. With the advancement of smart cities and home environments, the concept of user authentication is evolved with a broader implication by not only preventing unauthorized users from accessing confidential information but also providing the opportunities for customized services corresponding to a specific user. Traditional approaches of user authentication either require specialized device installation or inconvenient wearable sensor attachment. This article supports the extended concept of user authentication with a device-free approach by leveraging the prevalent WiFi signals made available by IoT devices, such as smart refrigerator, smart TV, and smart thermostat, and so on. The proposed system utilizes the WiFi signals to capture unique human physiological and behavioral characteristics inherited from their daily activities, including both walking and stationary ones. Particularly, we extract representative features from channel state information (CSI) measurements of WiFi signals, and develop a deep-learning-based user authentication scheme to accurately identify each individual user. To mitigate the signal distortion caused by surrounding people’s movements, our deep learning model exploits a CNN-based architecture that constructively combines features from multiple receiving antennas and derives more reliable feature abstractions. Furthermore, a transfer-learning-based mechanism is developed to reduce the training cost for new users and environments. Extensive experiments in various indoor environments are conducted to demonstrate the effectiveness of the proposed authentication system. In particular, our system can achieve over 94% authentication accuracy with 11 subjects through different activities. 
    more » « less
  5. Opportunistic Physics-mining Transfer Mapping Architecture (OPTMA) is a hybrid architecture that combines fast simplified physics models with neural networks in order to provide significantly improved generalizability and explainability compared to pure data-driven machine learning (ML) models. However, training OPTMA remains computationally inefficient due to its dependence on gradient-free solvers or back-propagation with supervised learning over expensively pre-generated labels. This paper presents two extensions of OPTMA that are not only more efficient to train through standard back-propagation but are readily deployable through the state-of-the-art library, PyTorch. The first extension, OPTMA-Net, presents novel manual reprogramming of the simplified physics model, expressing it in Torch tensor compatible form, thus naturally enabling PyTorch's in-built Auto-Differentiation to be used for training. Since manual reprogramming can be tedious for some physics models, a second extension called OPTMA-Dual is presented, where a highly accurate internal neural net is trained apriori on the fast simplified physics model (which can be generously sampled), and integrated with the transfer model. Both new architectures are tested on analytical test problems and the problem of predicting the acoustic field of an unmanned aerial vehicle. The interference of the acoustic pressure waves produced by multiple monopoles form the basis of the simplified physics for this problem statement. An indoor noise monitoring setup in motion capture environment provided the ground truth for target data. Compared to sequential hybrid and pure ML models, OPTMA-Net/Dual demonstrate several fold improvement in performing extrapolation, while providing orders of magnitude faster training times compared to the original OPTMA. 
    more » « less