Architectures and algorithms for user customization of CNNs

Harris, Barend; Moghaddam, Mansureh S.; Kang, Duseok; Bae, Inpyo; Kim, Euiseok; Min, Hyemi; Cho, Hansu; Kim, Sukjin; Egger, Bernhard; Ha, Soonhoi; Choi, Ki‐Young

doi:10.1109/aspdac.2018.8297379

Cited by 6 publications

(19 citation statements)

References 19 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In fact, the customization of both our work and [7] can be viewed as containing elements from transfer learning. The general idea of transfer learning is to improve the learning of target task on target domain by utilizing the knowledge learned from source domain regarding to source task [16].…”

Section: Related Workmentioning

confidence: 99%

“…The Extended MNIST (EMNIST) dataset [3] is used as the generic dataset during training. The images of handwritten digits and letters released by [7] are used as the customized dataset for customized training and testing. Generic Dataset: The EMNIST dataset is generated by applying Gaussian blurring, centering, padding, and down sampling to all the images in NIST Special dataset 19 [6].…”

Section: Generic and Customized Datasetsmentioning

confidence: 99%

“…Table 3 indicates that the overhead of network size and energy brought by LE + GN are minimal, at 2.52% and 2.45% respectively, relative to GE. [7] and fine-tuning the last few layers of pre-trained DNN. Besides the proposed MOE model, there exists other alternative approaches for user customization of pre-trained large DNNs.…”

Section: Classification Accuracies Of Various Components In Moementioning

confidence: 99%

“…We use EM-NIST [3] as the generic dataset and representative of many users. We then use a small customized dataset provided by [7] which is representative handwritings of 10 users. In our experiments, we show the overhead of LE and GN is about 2.5% of GE in terms of energy and the network size (represented by the number of distinct parameters).…”

Section: Introductionmentioning

confidence: 99%

“…Compared to a recent work [7] which also focuses on customization, the MOE architecture provides more flexibility and results in achieving higher classification accuracy over the generic dataset [7]. Overall, the summary of our contributions are as follows:…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

A Mixture of Experts Approach for Low-Cost DNN Customization

Zhang

Davoodi

2021

IEEE Des. Test

View full text Add to dashboard Cite

The ability to customize a trained Deep Neural Network (DNN) locally using user-specific data may greatly enhance user experiences, reduce development costs, and protect user's privacy. In this work, we propose to incorporate a novel Mixture of Experts (MOE) approach to accomplish this goal. This architecture comprises of a Global Expert (GE), a Local Expert (LE) and a Gating Network (GN). The GE is a trained DNN developed on a large training dataset representative of many potential users. After deployment on an embedded edge device, GE will be subject to customized, user-specific data (e.g., accent in speech) and its performance may suffer. This problem may be alleviated by training a local DNN (the local expert, LE) on a small size customized training data to correct the errors made by GE. A gating network then will be trained to determine whether an incoming data should be handled by GE or LE. Since the customized dataset is in general very small, the cost of training LE and GN would be much lower than that of re-training of GE. The training of LE and GN thus can be performed at local device, properly protecting the privacy of customized training data. In this work, we developed a prototype MOE architecture for handwritten alphanumeric character recognition task. We use EMNIST as the generic dataset, LeNet5 as GE, and handwritings of 10 users as the customized dataset. We show that with the LE and GN, the classification accuracy is significantly enhanced over the customized dataset with almost no degradation of accuracy over the generic dataset. In terms of energy and network size, the overhead of LE and GN is around 2.5% compared to those of GE.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Generic and Customized Datasetsmentioning

confidence: 99%

Section: Classification Accuracies Of Various Components In Moementioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations