An official website of the United States government
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
Selecting and Composing Learning Rate Policies for Deep Neural Networks
The choice of learning rate (LR) functions and policies has evolved from a simple fixed LR to the decaying LR and the cyclic LR, aiming to improve the accuracy and reduce the training time of Deep Neural Networks (DNNs). This article presents a systematic approach to selecting and composing an LR policy for effective DNN training to meet desired target accuracy and reduce training time within the pre-defined training iterations. It makes three original contributions. First, we develop an LR tuning mechanism for auto-verification of a given LR policy with respect to the desired accuracy goal under the pre-defined training time constraint. Second, we develop an LR policy recommendation system (LRBench) to select and compose good LR policies from the same and/or different LR functions through dynamic tuning, and avoid bad choices, for a given learning task, DNN model, and dataset. Third, we extend LRBench by supporting different DNN optimizers and show the significant mutual impact of different LR policies and different optimizers. Evaluated using popular benchmark datasets and different DNN models (LeNet, CNN3, ResNet), we show that our approach can effectively deliver high DNN test accuracy, outperform the existing recommended default LR policies, and reduce the DNN training time by 1.6-6.7× to meet a targeted model accuracy.
more »
« less
- Award ID(s):
- 2038029
- PAR ID:
- 10475333
- Publisher / Repository:
- ACM
- Date Published:
- Journal Name:
- ACM Transactions on Intelligent Systems and Technology
- Volume:
- 14
- Issue:
- 2
- ISSN:
- 2157-6904
- Page Range / eLocation ID:
- 1 to 25
- Subject(s) / Keyword(s):
- Ensemble learning
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Deep neural networks have become increasingly popular in radar micro-Doppler classification; yet, a key challenge, which has limited potential gains, is the lack of large amounts of measured data that can facilitate the design of deeper networks with greater robustness and performance. Several approaches have been proposed in the literature to address this problem, such as unsupervised pre-training and transfer learning from optical imagery or synthetic RF data. This work investigates an alternative approach to training which involves exploitation of “datasets of opportunity” – micro-Doppler datasets collected using other RF sensors, which may be of a different frequency, bandwidth or waveform - for the purposes of training. Specifically, this work compares in detail the cross-frequency training degradation incurred for several different training approaches and deep neural network (DNN) architectures. Results show a 70% drop in classification accuracy when the RF sensors for pre-training, fine-tuning, and testing are different, and a 15% degradation when only the pre-training data is different, but the fine-tuning and test data are from the same sensor. By using generative adversarial networks (GANs), a large amount of synthetic data is generated for pre-training. Results show that cross-frequency performance degradation is reduced by 50% when kinematically-sifted GAN-synthesized signatures are used in pre-training.more » « less
-
Variability-induced accuracy degradation of RRAMbased DNNs is of great concern due to their significant potential for use in future energy-efficient machine learning architectures. To address this, we propose a two-step process. First, an enhanced testing procedure is used to predict DNN accuracy from a set of compact test stimuli (images). This test response (signature) is simply the concatenated vectors of output neurons of intermediate and final DNN layers over the compact test images applied. DNNs with a predicted accuracy below a threshold are then tuned based on this signature vector. Using a clustering based approach, the signature is mapped to the optimal tuning parameter values of the DNN (determined using off-line training of the DNN via backpropagation) in a single step, eliminating any post-manufacture training of the DNN weights (expensive). The tuning parameters themselves consist of the gains and offsets of the ReLU activation of neurons of the DNN on a per-layer basis and can be tuned digitally. Tuning is achieved in less than a second of tuning time, with yield improvements of over 45% with a modest accuracy reduction of 4% compared to digital DNNs.more » « less
-
Variability-induced accuracy degradation of RRAM based DNNs is of great concern due to their significant potential for use in future energy-efficient machine learning architectures. To address this, we propose a two-step process. First, an enhanced testing procedure is used to predict DNN accuracy from a set of compact test stimuli (images). This test response (signature) is simply the concatenated vectors of output neurons of intermediate final DNN layers over the compact test images applied. DNNs with a predicted accuracy below a threshold are then tuned based on this signature vector. Using a clustering based approach, the signature is mapped to the optimal tuning parameter values of the DNN (determined using off-line training of the DNN via backpropagation) in a single step, eliminating any post-manufacture training of the DNN weights (expensive). The tuning parameters themselves consist of the gains and offsets of the ReLU activation of neurons of the DNN on a per-layer basis and can be tuned digitally. Tuning is achieved in less than a second of tuning time, with yield improvements of over 45% with a modest accuracy reduction of 4% compared to digital DNNs.more » « less
-
Object insertion under tight tolerances (less than 1mm) is an important but challenging assembly task as even small errors can result in undesirable contacts. Recent efforts focused on Reinforcement Learning (RL), which often depends on careful definition of dense reward functions. This work proposes an effective strategy for such tasks that integrates traditional model-based control with RL to achieve improved insertion accuracy. The policy is trained exclusively in simulation and is zero-shot transferred to the real system. It employs a potential field-based controller to acquire a model-based policy for inserting a plug into a socket given full observability in simulation. This policy is then integrated with residual RL, which is trained in simulation given only a sparse, goal-reaching reward. A curriculum scheme over observation noise and action magnitude is used for training the residual RL policy. Both policy components use as input the SE(3) poses of both the plug and the socket and return the plug's SE(3) pose transform, which is executed by a robotic arm using a controller. The integrated policy is deployed on the real system without further training or fine-tuning, given a visual SE(3) object tracker. The proposed solution and alternatives are evaluated across a variety of objects and conditions in simulation and reality. The proposed approach outperforms recent RL-based methods in this domain and prior efforts with hybrid policies. Ablations highlight the impact of each component of the approach.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.1145/3570508
