Sungho Shin

Sungho Shin, Joosoon Lee, Junseok Lee et al.

Deep learning has achieved outstanding performance for face recognition benchmarks, but performance reduces significantly for low resolution (LR) images. We propose an attention similarity knowledge distillation approach, which transfers attention maps obtained from a high resolution (HR) network as a teacher into an LR network as a student to boost LR recognition performance. Inspired by humans being able to approximate an object's region from an LR image based on prior knowledge obtained from HR images, we designed the knowledge distillation loss using the cosine similarity to make the student network's attention resemble the teacher network's attention. Experiments on various LR face related benchmarks confirmed the proposed method generally improved recognition performances on LR settings, outperforming state-of-the-art results by simply transferring well-constructed attention maps. The code and pretrained models are publicly available in the https://github.com/gist-ailab/teaching-where-to-look.

Sungho Shin, Yeonguk Yu, Kyoobin Lee

In this study, we introduce a feature knowledge distillation framework to improve low-resolution (LR) face recognition performance using knowledge obtained from high-resolution (HR) images. The proposed framework transfers informative features from an HR-trained network to an LR-trained network by reducing the distance between them. A cosine similarity measure was employed as a distance metric to effectively align the HR and LR features. This approach differs from conventional knowledge distillation frameworks, which use the L_p distance metrics and offer the advantage of converging well when reducing the distance between features of different resolutions. Our framework achieved a 3% improvement over the previous state-of-the-art method on the AgeDB-30 benchmark without bells and whistles, while maintaining a strong performance on HR images. The effectiveness of cosine similarity as a distance metric was validated through statistical analysis, making our approach a promising solution for real-world applications in which LR images are frequently encountered. The code and pretrained models are publicly available on https://github.com/gist-ailab/feature-similarity-KD.

Sungho Shin, Yiheng Lin, Guannan Qu et al.

This paper studies the trade-off between the degree of decentralization and the performance of a distributed controller in a linear-quadratic control setting. We study a system of interconnected agents over a graph and a distributed controller, called $κ$-distributed control, which lets the agents make control decisions based on the state information within distance $κ$ on the underlying graph. This controller can tune its degree of decentralization using the parameter $κ$ and thus allows a characterization of the relationship between decentralization and performance. We show that under mild assumptions, including stabilizability, detectability, and a subexponentially growing graph condition, the performance difference between $κ$-distributed control and centralized optimal control becomes exponentially small in $κ$. This result reveals that distributed control can achieve near-optimal performance with a moderate degree of decentralization, and thus it is an effective controller architecture for large-scale networked systems.

Sungho Shin, François Pacaud, Mihai Anitescu

This paper introduces a framework for solving alternating current optimal power flow (ACOPF) problems using graphics processing units (GPUs). While GPUs have demonstrated remarkable performance in various computing domains, their application in ACOPF has been limited due to challenges associated with porting sparse automatic differentiation (AD) and sparse linear solver routines to GPUs. We address these issues with two key strategies. First, we utilize a single-instruction, multiple-data abstraction of nonlinear programs. This approach enables the specification of model equations while preserving their parallelizable structure and, in turn, facilitates the parallel AD implementation. Second, we employ a condensed-space interior-point method (IPM) with an inequality relaxation. This technique involves condensing the Karush--Kuhn--Tucker (KKT) system into a positive definite system. This strategy offers the key advantage of being able to factorize the KKT matrix without numerical pivoting, which has hampered the parallelization of the IPM algorithm. By combining these strategies, we can perform the majority of operations on GPUs while keeping the data residing in the device memory only. Comprehensive numerical benchmark results showcase the advantage of our approach. Remarkably, our implementations -- MadNLP.jl and ExaModels.jl -- running on NVIDIA GPUs achieve an order of magnitude speedup compared with state-of-the-art tools running on contemporary CPUs.

Yeonguk Yu, Sungho Shin, Seongju Lee et al.

Detecting out-of-distribution (OOD) inputs during the inference stage is crucial for deploying neural networks in the real world. Previous methods commonly relied on the output of a network derived from the highly activated feature map. In this study, we first revealed that a norm of the feature map obtained from the other block than the last block can be a better indicator of OOD detection. Motivated by this, we propose a simple framework consisting of FeatureNorm: a norm of the feature map and NormRatio: a ratio of FeatureNorm for ID and OOD to measure the OOD detection performance of each block. In particular, to select the block that provides the largest difference between FeatureNorm of ID and FeatureNorm of OOD, we create Jigsaw puzzle images as pseudo OOD from ID training samples and calculate NormRatio, and the block with the largest value is selected. After the suitable block is selected, OOD detection with the FeatureNorm outperforms other OOD detection methods by reducing FPR95 by up to 52.77% on CIFAR10 benchmark and by up to 48.53% on ImageNet benchmark. We demonstrate that our framework can generalize to various architectures and the importance of block selection, which can improve previous OOD detection methods as well.

Dirk Lauinger, Deepjyoti Deka, Sungho Shin

Electricity distribution companies deploy battery storage to defer grid upgrades by reducing peak demand. In deregulated jurisdictions, such storage often sits idle because regulatory constraints bar participation in electricity markets. Here, we develop an optimization framework that, to our knowledge, provides the first formal model of market participation constraints within storage investment and operation planning. Applying the framework to a Massachusetts case study, we find that market participation delivers similar savings as peak demand reduction. Under current conditions, market participation does not increase storage investment, but at very low storage costs, could incentivize deployment beyond local distribution needs. This might run contrary to the separation of distribution from generation in deregulated markets. Our framework can mitigate this concern by identifying investment levels appropriate for local distribution needs.

Seunghyeok Back, Sangbeom Lee, Kangmin Kim et al.

Accurate and efficient segmentation of unknown objects in unstructured environments is essential for robotic manipulation. Unknown Object Instance Segmentation (UOIS), which aims to identify all objects in unknown categories and backgrounds, has become a key capability for various robotic tasks. However, existing methods struggle with over-segmentation and under-segmentation, leading to failures in manipulation tasks such as grasping. To address these challenges, we propose QuBER (Quadruple Boundary Error Refinement), a novel error-informed refinement approach for high-quality UOIS. QuBER first estimates quadruple boundary errors-true positive, true negative, false positive, and false negative pixels-at the instance boundaries of the initial segmentation. It then refines the segmentation using an error-guided fusion mechanism, effectively correcting both fine-grained and instance-level segmentation errors. Extensive evaluations on three public benchmarks demonstrate that QuBER outperforms state-of-the-art methods and consistently improves various UOIS methods while maintaining a fast inference time of less than 0.1 seconds. Furthermore, we show that QuBER improves the success rate of grasping target objects in cluttered environments. Code and supplementary materials are available at https://sites.google.com/view/uois-quber.

Yeonguk Yu, Sungho Shin, Seunghyeok Back et al.

Test-time adaptation (TTA) aims to adapt a pre-trained model to a new test domain without access to source data after deployment. Existing approaches typically rely on self-training with pseudo-labels since ground-truth cannot be obtained from test data. Although the quality of pseudo labels is important for stable and accurate long-term adaptation, it has not been previously addressed. In this work, we propose DPLOT, a simple yet effective TTA framework that consists of two components: (1) domain-specific block selection and (2) pseudo-label generation using paired-view images. Specifically, we select blocks that involve domain-specific feature extraction and train these blocks by entropy minimization. After blocks are adjusted for current test domain, we generate pseudo-labels by averaging given test images and corresponding flipped counterparts. By simply using flip augmentation, we prevent a decrease in the quality of the pseudo-labels, which can be caused by the domain gap resulting from strong augmentation. Our experimental results demonstrate that DPLOT outperforms previous TTA methods in CIFAR10-C, CIFAR100-C, and ImageNet-C benchmarks, reducing error by up to 5.4%, 9.1%, and 2.9%, respectively. Also, we provide an extensive analysis to demonstrate effectiveness of our framework. Code is available at https://github.com/gist-ailab/domain-specific-block-selection-and-paired-view-pseudo-labeling-for-online-TTA.

Junseok Lee, Sungho Shin, Seongju Lee et al.

Single-domain generalization is essential for object detection, particularly when training models on a single source domain and evaluating them on unseen target domains. Domain shifts, such as changes in weather, lighting, or scene conditions, pose significant challenges to the generalization ability of existing models. To address this, we propose Cross-Domain Feature Knowledge Distillation (CD-FKD), which enhances the generalization capability of the student network by leveraging both global and instance-wise feature distillation. The proposed method uses diversified data through downscaling and corruption to train the student network, whereas the teacher network receives the original source domain data. The student network mimics the features of the teacher through both global and instance-wise distillation, enabling it to extract object-centric features effectively, even for objects that are difficult to detect owing to corruption. Extensive experiments on challenging scenes demonstrate that CD-FKD outperforms state-of-the-art methods in both target domain generalization and source domain performance, validating its effectiveness in improving object detection robustness to domain shifts. This approach is valuable in real-world applications, like autonomous driving and surveillance, where robust object detection in diverse environments is crucial.

Shaohui Liu, Sungho Shin, Deepjyoti Deka

Enabling continued data-center growth under increasing grid stress motivates closer coordination between flexible computing demand and co-located battery energy storage systems (BESS) to improve site operations and provide grid services. This paper develops a robust co-optimization framework for day-ahead operation of data centers with co-located BESS under utility-imposed interconnection limits on peak load and ramping. The model jointly considers deadline-constrained computing workloads, managed through workload scheduling and dynamic voltage and frequency scaling (DVFS), together with degradation-aware BESS dispatch to enable cost optimization and participation in ancillary-service markets. Case studies based on real-world market and workload data show that the proposed framework yields feasible day-ahead schedules across a range of operating conditions, with substantially larger benefits when interconnection constraints become binding. Under baseline conditions, BESS value is derived from both ancillary-service participation and improved workload and energy management. Under stressed peak-load and ramping limits, however, the daily value of BESS increases by a factor of two or more, driven primarily \revise{by BESS actions to reduce the potential incompletion in the schedulable workload while complying with interconnection constraints}. Under tight peak-load caps, workload composition also matters where a higher share of non-schedulable jobs can increase operating cost by more than 25\% relative to more flexible workload mixes. \revise{Additionally, DVFS studies further show that processor-level control is a material flexibility lever under tight load limits.} These results demonstrate that coordinated compute-storage flexibility can materially expand the operational headroom and grid value of data centers, especially under increasingly scarce grid capacity.

Sanjay Johnson, Dirk Lauinger, Sungho Shin et al.

As GPU-accelerated mathematical programming techniques mature, there is growing interest in utilizing them to address the computational challenges of power system optimization. This paper introduces ExaModelsPower.jl, an open-source modeling library for creating GPU-compatible nonlinear AC optimal power flow models. Built on ExaModels.jl, ExaModelsPower.jl provides a high-level interface that automatically generates all necessary callback functions for GPU solvers. The library is designed for large-scale problem instances, which may include multiple time periods and security constraints. Using ExaModelsPower.jl, we benchmark GPU and CPU solvers on open-source test cases. Our results show that GPU solvers can deliver up to two orders of magnitude speedups compared to alternative tools on CPU for problems with more than 20,000 variables and a solution precision of up to $10^{-4}$, while performance for smaller instances or tighter tolerances may vary.

Flemming Holtorf, Sungho Shin

We present an approximate dynamic programming framework for designing degradation-aware market participation policies for battery energy storage systems. The approach employs a tailored value function approximation that reduces the state space to state of charge and battery health, while performing dynamic programming along a pseudo-time axis encoded by state of health. This formulation enables an offline/online computation split that separates long-term degradation dynamics (months to years) from short-term market dynamics (seconds to minutes) -- a timescale mismatch that renders conventional predictive control and dynamic programming approaches computationally intractable. The main computational effort occurs offline, where the value function is approximated via coarse-grained backward induction along the health dimension. Online decisions then reduce to a real-time tractable one-step predictive control problem guided by the precomputed value function. This decoupling allows the integration of high-fidelity physics-informed degradation models without sacrificing real-time feasibility. Backtests on historical market data show that the resulting policy outperforms several benchmark strategies with optimized hyperparameters.

Yeonguk Yu, Minhwan Ko, Sungho Shin et al.

Deep neural networks have demonstrated remarkable performance in various vision tasks, but their success heavily depends on the quality of the training data. Noisy labels are a critical issue in medical datasets and can significantly degrade model performance. Previous clean sample selection methods have not utilized the well pre-trained features of vision foundation models (VFMs) and assumed that training begins from scratch. In this paper, we propose CUFIT, a curriculum fine-tuning paradigm of VFMs for medical image classification under label noise. Our method is motivated by the fact that linear probing of VFMs is relatively unaffected by noisy samples, as it does not update the feature extractor of the VFM, thus robustly classifying the training samples. Subsequently, curriculum fine-tuning of two adapters is conducted, starting with clean sample selection from the linear probing phase. Our experimental results demonstrate that CUFIT outperforms previous methods across various medical image benchmarks. Specifically, our method surpasses previous baselines by 5.0%, 2.1%, 4.6%, and 5.8% at a 40% noise rate on the HAM10000, APTOS-2019, BloodMnist, and OrgancMnist datasets, respectively. Furthermore, we provide extensive analyses to demonstrate the impact of our method on noisy label detection. For instance, our method shows higher label precision and recall compared to previous approaches. Our work highlights the potential of leveraging VFMs in medical image classification under challenging conditions of noisy labels.

Joosoon Lee, Seongju Lee, Seunghyeok Back et al.

Understanding assembly instruction has the potential to enhance the robot s task planning ability and enables advanced robotic applications. To recognize the key components from the 2D assembly instruction image, We mainly focus on segmenting the speech bubble area, which contains lots of information about instructions. For this, We applied Cascade Mask R-CNN and developed a context-aware data augmentation scheme for speech bubble segmentation, which randomly combines images cuts by considering the context of assembly instructions. We showed that the proposed augmentation scheme achieves a better segmentation performance compared to the existing augmentation algorithm by increasing the diversity of trainable data while considering the distribution of components locations. Also, we showed that deep learning can be useful to understand assembly instruction by detecting the essential objects in the assembly instruction, such as tools and parts.

Yoonho Boo, Sungho Shin, Jungwook Choi et al.

The quantization of deep neural networks (QDNNs) has been actively studied for deployment in edge devices. Recent studies employ the knowledge distillation (KD) method to improve the performance of quantized networks. In this study, we propose stochastic precision ensemble training for QDNNs (SPEQ). SPEQ is a knowledge distillation training scheme; however, the teacher is formed by sharing the model parameters of the student network. We obtain the soft labels of the teacher by changing the bit precision of the activation stochastically at each layer of the forward-pass computation. The student model is trained with these soft labels to reduce the activation quantization noise. The cosine similarity loss is employed, instead of the KL-divergence, for KD training. As the teacher model changes continuously by random bit-precision assignment, it exploits the effect of stochastic ensemble KD. SPEQ outperforms the existing quantization training methods in various tasks, such as image classification, question-answering, and transfer learning without the need for cumbersome teacher networks.

Wonyong Sung, Iksoo Choi, Jinhwan Park et al.

The stochastic gradient descent (SGD) method is most widely used for deep neural network (DNN) training. However, the method does not always converge to a flat minimum of the loss surface that can demonstrate high generalization capability. Weight noise injection has been extensively studied for finding flat minima using the SGD method. We devise a new weight-noise injection-based SGD method that adds symmetrical noises to the DNN weights. The training with symmetrical noise evaluates the loss surface at two adjacent points, by which convergence to sharp minima can be avoided. Fixed-magnitude symmetric noises are added to minimize training instability. The proposed method is compared with the conventional SGD method and previous weight-noise injection algorithms using convolutional neural networks for image classification. Particularly, performance improvements in large batch training are demonstrated. This method shows superior performance compared with conventional SGD and weight-noise injection methods regardless of the batch-size and learning rate scheduling algorithms.

Jongwon Kim, Sungho Shin, Yeonguk Yu et al.

Privacy issues were raised in the process of training deep learning in medical, mobility, and other fields. To solve this problem, we present privacy-preserving distributed deep learning method that allow clients to learn a variety of data without direct exposure. We divided a single deep learning architecture into a common extractor, a cloud model and a local classifier for the distributed learning. First, the common extractor, which is used by local clients, extracts secure features from the input data. The secure features also take the role that the cloud model can employ various task and diverse types of data. The feature contain the most important information that helps to proceed various task. Second, the cloud model including most parts of the whole training model gets the embedded features from the massive local clients, and performs most of deep learning operations which takes severe computing cost. After the operations in cloud model finished, outputs of the cloud model send back to local clients. Finally, the local classifier determined classification results and delivers the results to local clients. When clients train models, our model does not directly expose sensitive information to exterior network. During the test, the average performance improvement was 2.63% over the existing local training model. However, in a distributed environment, there is a possibility of inversion attack due to exposed features. For this reason, we experimented with the common extractor to prevent data restoration. The quality of restoration of the original image was tested by adjusting the depth of the common extractor. As a result, we found that the deeper the common extractor, the restoration score decreased to 89.74.

Yoonho Boo, Sungho Shin, Wonyong Sung

Quantized deep neural networks (QDNNs) are necessary for low-power, high throughput, and embedded applications. Previous studies mostly focused on developing optimization methods for the quantization of given models. However, quantization sensitivity depends on the model architecture. Therefore, the model selection needs to be a part of the QDNN design process. Also, the characteristics of weight and activation quantization are quite different. This study proposes a holistic approach for the optimization of QDNNs, which contains QDNN training methods as well as quantization-friendly architecture design. Synthesized data is used to visualize the effects of weight and activation quantization. The results indicate that deeper models are more prone to activation quantization, while wider models improve the resiliency to both weight and activation quantization. This study can provide insight into better optimization of QDNNs.

Sen Na, Sungho Shin, Mihai Anitescu et al.

We study the convergence properties of an overlapping Schwarz decomposition algorithm for solving nonlinear optimal control problems (OCPs). The algorithm decomposes the time domain into a set of overlapping subdomains, and solves all subproblems defined over subdomains in parallel. The convergence is attained by updating primal-dual information at the boundaries of overlapping subdomains. We show that the algorithm exhibits local linear convergence, and that the convergence rate improves exponentially with the overlap size. We also establish global convergence results for a general quadratic programming, which enables the application of the Schwarz scheme inside second-order optimization algorithms (e.g., sequential quadratic programming). The theoretical foundation of our convergence analysis is a sensitivity result of nonlinear OCPs, which we call "exponential decay of sensitivity" (EDS). Intuitively, EDS states that the impact of perturbations at domain boundaries (i.e. initial and terminal time) on the solution decays exponentially as one moves into the domain. Here, we expand a previous analysis available in the literature by showing that EDS holds for both primal and dual solutions of nonlinear OCPs, under uniform second-order sufficient condition, controllability condition, and boundedness condition. We conduct experiments with a quadrotor motion planning problem and a PDE control problem to validate our theory; and show that the approach is significantly more efficient than ADMM and as efficient as the centralized solver Ipopt.

Sungho Shin, Qiugang Lu, Victor M. Zavala

This paper presents unifying results for subspace identification (SID) and dynamic mode decomposition (DMD) for autonomous dynamical systems. We observe that SID seeks to solve an optimization problem to estimate an extended observability matrix and a state sequence that minimizes the prediction error for the state-space model. Moreover, we observe that DMD seeks to solve a rank-constrained matrix regression problem that minimizes the prediction error of an extended autoregressive model. We prove that existence conditions for perfect (error-free) state-space and low-rank extended autoregressive models are equivalent and that the SID and DMD optimization problems are equivalent. We exploit these results to propose a SID-DMD algorithm that delivers a provably optimal model and that is easy to implement. We demonstrate our developments using a case study that aims to build dynamical models directly from video data.

Sungho Shin, Alex D. Smith, S. Joe Qin et al.

Dynamic inner principal component analysis (DiPCA) is a powerful method for the analysis of time-dependent multivariate data. DiPCA extracts dynamic latent variables that capture the most dominant temporal trends by solving a large-scale, dense, and nonconvex nonlinear program (NLP). A scalable decomposition algorithm has been recently proposed in the literature to solve these challenging NLPs. The decomposition algorithm performs well in practice but its convergence properties are not well understood. In this work, we show that this algorithm is a specialized variant of a coordinate maximization algorithm. This observation allows us to explain why the decomposition algorithm might work (or not) in practice and can guide improvements. We compare the performance of the decomposition strategies with that of the off-the-shelf solver Ipopt. The results show that decomposition is more scalable and, surprisingly, delivers higher quality solutions.

Sungho Shin, Yoonho Boo, Wonyong Sung

Designing a deep neural network (DNN) with good generalization capability is a complex process especially when the weights are severely quantized. Model averaging is a promising approach for achieving the good generalization capability of DNNs, especially when the loss surface for training contains many sharp minima. We present a new quantized neural network optimization approach, stochastic quantized weight averaging (SQWA), to design low-precision DNNs with good generalization capability using model averaging. The proposed approach includes (1) floating-point model training, (2) direct quantization of weights, (3) capturing multiple low-precision models during retraining with cyclical learning rates, (4) averaging the captured models, and (5) re-quantizing the averaged model and fine-tuning it with low-learning rates. Additionally, we present a loss-visualization technique on the quantized weight domain to clearly elucidate the behavior of the proposed method. Visualization results indicate that a quantized DNN (QDNN) optimized with the proposed approach is located near the center of the flat minimum in the loss surface. With SQWA training, we achieved state-of-the-art results for 2-bit QDNNs on CIFAR-100 and ImageNet datasets. Although we only employed a uniform quantization scheme for the sake of implementation in VLSI or low-precision neural processing units, the performance achieved exceeded those of previous studies employing non-uniform quantization.

Sungho Shin, Yoonho Boo, Wonyong Sung

Knowledge distillation (KD) is a very popular method for model size reduction. Recently, the technique is exploited for quantized deep neural networks (QDNNs) training as a way to restore the performance sacrificed by word-length reduction. KD, however, employs additional hyper-parameters, such as temperature, coefficient, and the size of teacher network for QDNN training. We analyze the effect of these hyper-parameters for QDNN optimization with KD. We find that these hyper-parameters are inter-related, and also introduce a simple and effective technique that reduces \textit{coefficient} during training. With KD employing the proposed hyper-parameters, we achieve the test accuracy of 92.7% and 67.0% on Resnet20 with 2-bit ternary weights for CIFAR-10 and CIFAR-100 data sets, respectively.

Sungho Shin, Yoonho Boo, Wonyong Sung

Fixed-point optimization of deep neural networks plays an important role in hardware based design and low-power implementations. Many deep neural networks show fairly good performance even with 2- or 3-bit precision when quantized weights are fine-tuned by retraining. We propose an improved fixedpoint optimization algorithm that estimates the quantization step size dynamically during the retraining. In addition, a gradual quantization scheme is also tested, which sequentially applies fixed-point optimizations from high- to low-precision. The experiments are conducted for feed-forward deep neural networks (FFDNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs).

Sungho Shin, Kyuyeon Hwang, Wonyong Sung

The complexity of deep neural network algorithms for hardware implementation can be lowered either by scaling the number of units or reducing the word-length of weights. Both approaches, however, can accompany the performance degradation although many types of research are conducted to relieve this problem. Thus, it is an important question which one, between the network size scaling and the weight quantization, is more effective for hardware optimization. For this study, the performances of fully-connected deep neural networks (FCDNNs) and convolutional neural networks (CNNs) are evaluated while changing the network complexity and the word-length of weights. Based on these experiments, we present the effective compression ratio (ECR) to guide the trade-off between the network size and the precision of weights when the hardware resource is limited.

Minjae Lee, Kyuyeon Hwang, Jinhwan Park et al.

In this paper, a neural network based real-time speech recognition (SR) system is developed using an FPGA for very low-power operation. The implemented system employs two recurrent neural networks (RNNs); one is a speech-to-character RNN for acoustic modeling (AM) and the other is for character-level language modeling (LM). The system also employs a statistical word-level LM to improve the recognition accuracy. The results of the AM, the character-level LM, and the word-level LM are combined using a fairly simple N-best search algorithm instead of the hidden Markov model (HMM) based network. The RNNs are implemented using massively parallel processing elements (PEs) for low latency and high throughput. The weights are quantized to 6 bits to store all of them in the on-chip memory of an FPGA. The proposed algorithm is implemented on a Xilinx XC7Z045, and the system can operate much faster than real-time.

Sungho Shin, Wonyong Sung

Gesture recognition is a very essential technology for many wearable devices. While previous algorithms are mostly based on statistical methods including the hidden Markov model, we develop two dynamic hand gesture recognition techniques using low complexity recurrent neural network (RNN) algorithms. One is based on video signal and employs a combined structure of a convolutional neural network (CNN) and an RNN. The other uses accelerometer data and only requires an RNN. Fixed-point optimization that quantizes most of the weights into two bits is conducted to optimize the amount of memory size for weight storage and reduce the power consumption in hardware and software based implementations.

Sungho Shin, Kyuyeon Hwang, Wonyong Sung

In this paper, we propose a generative knowledge transfer technique that trains an RNN based language model (student network) using text and output probabilities generated from a previously trained RNN (teacher network). The text generation can be conducted by either the teacher or the student network. We can also improve the performance by taking the ensemble of soft labels obtained from multiple teacher networks. This method can be used for privacy conscious language model adaptation because no user data is directly used for training. Especially, when the soft labels of multiple devices are aggregated via a trusted third party, we can expect very strong privacy protection.

Sungho Shin, Kyuyeon Hwang, Wonyong Sung

Recurrent neural networks have shown excellent performance in many applications, however they require increased complexity in hardware or software based implementations. The hardware complexity can be much lowered by minimizing the word-length of weights and signals. This work analyzes the fixed-point performance of recurrent neural networks using a retrain based quantization method. The quantization sensitivity of each layer in RNNs is studied, and the overall fixed-point optimization results minimizing the capacity of weights while not sacrificing the performance are presented. A language model and a phoneme recognition examples are used.

Wonyong Sung, Sungho Shin, Kyuyeon Hwang

The complexity of deep neural network algorithms for hardware implementation can be much lowered by optimizing the word-length of weights and signals. Direct quantization of floating-point weights, however, does not show good performance when the number of bits assigned is small. Retraining of quantized networks has been developed to relieve this problem. In this work, the effects of retraining are analyzed for a feedforward deep neural network (FFDNN) and a convolutional neural network (CNN). The network complexity is controlled to know their effects on the resiliency of quantized networks by retraining. The complexity of the FFDNN is controlled by varying the unit size in each hidden layer and the number of layers, while that of the CNN is done by modifying the feature map configuration. We find that the performance gap between the floating-point and the retrain-based ternary (+1, 0, -1) weight neural networks exists with a fair amount in 'complexity limited' networks, but the discrepancy almost vanishes in fully complex networks whose capability is limited by the training data, rather than by the number of connections. This research shows that highly complex DNNs have the capability of absorbing the effects of severe weight quantization through retraining, but connection limited networks are less resilient. This paper also presents the effective compression ratio to guide the trade-off between the network size and the precision when the hardware resource is limited.