Title: Weak Detection of Signal in the Spiked Wigner Model
Authors: Hye Won Chung, Ji Oon Lee
We consider the problem of detecting the presence of the signal in a rank-one signal-plus-noise data matrix. In case the signal-to-noise ratio is under the threshold below which a reliable detection is impossible, we propose a hypothesis test based on the linear spectral statistics of the data matrix. When the noise is Gaussian, the error of the proposed test is optimal as it matches the error of the likelihood ratio test that minimizes the sum of the Type-I and Type-II errors. The test is data-driven and does not depend on the distribution of the signal or the noise. If the density of the noise is known, it can be further improved by an entrywise transformation to lower the error of the test.
One of the fundamental questions in machine learning is to detect signals from given data. If the data is of ‘signal-plus-noise’ type, the model is often referred to as a ‘spiked model.’ If the strength of the signal is considerably stronger than that of the noise, we can reliably detect the signal and also recover the signal from the noisy data. On the other hand, if the noise dominates the signal, it is impossible to detect the presence of the signal from the data, which is indistinguishable from pure noise.
In this paper, we consider the case that the strengths of the signal and of the noise are comparable. It was known that there is a certain threshold for the signal-to-noise ratio (SNR) above which the reliable detection, or the strong detection is available, whereas the strong detection is impossible if SNR is below the threshold. In the latter case, we try the weak detection to determine whether the signal is present in the given data. More precisely, we propose a hypothesis test with low computational complexity whose probability of error is minimal. The test is based on state-of-the-art techniques from random matrix theory.
If the noise is non-Gaussian, the test can be further improved by suitably processing the given data. Such a procedure, which we call an entrywise transformation in our work, effectively increases SNR. In case the noise has exponential decay, the entrywise transformation corresponds to applying a function similar to the hyperbolic tangent function (tanh) to each data entry. Our test is expected to be used in various problems with noisy high dimensional data such as community detection and angular synchronization.
Figure1: The limiting density of the proposed test statistic under H0 (when the signal is not present) and under H1 (when the signal is present)
Figure2: The limiting errors of the proposed algorithms (Alg 1: without entrywise transformation, Alg2: with entrywise transformation.)
Title: Dimension-Wise Importance Sampling Weight Clipping for Sample-Efficient Reinforcement Learning
Authors: Seungyul Han and Youngchul Sung
In importance sampling (IS)-based reinforcement learning algorithms such as Proximal Policy Optimization (PPO), IS weights are typically clipped to avoid large variance in learning. However, policy update from clipped statistics induces large bias in tasks with high action dimensions, and bias from clipping makes it difficult to reuse old samples with large IS weights. In this work, we propose the Dimension-wise Importance Sampling Weight Clipping (DISC) algorithm based on PPO, a representative on-policy algorithm, by applying dimension-wise IS weight clipping which separately clips the IS weight of each action dimension to avoid large bias and adaptively controls the IS weight to bound policy update from the current policy. This new technique enables efficient learning in high action-dimensional tasks and reusing old samples like in off-policy learning to significantly increase the sample efficiency. Numerical results show that the proposed DISC algorithm outperforms other state-of-the-art RL algorithms in various Open AI Gym tasks.
Title: TapNet: Neural Network Augmented with Task-Adaptive Projection for Few-Shot Learning
Authors: Sung Whan Yoon, Jun Seo, Jaekyun Moon
Few-shot learning promises to allow machines to carry out tasks that are previously unencountered, using only a small number of relevant examples. As such, few-shot learning
finds wide applications, where labeled data are scarce or expensive, which is far more often the case than not. Unfortunately, despite immense interest and active research in recent years, few-shot learning remains an elusive challenge to machine learning community. For example, while
deep networks now routinely offer near-perfect classification scores on standard image test datasets given ample training, reported results on few-shot learning still fall well below the levels that would be considered reliable in crucial real world settings.
We propose TapNets, neural networks augmented with task-adaptive projection for improved few-shot learning. See Figure. Here, employing a meta-learning strategy with episode-based training, a network and a set of per-class reference vectors are learned across widely varying tasks. At the same time, for every episode, features in the embedding space are linearly projected into a new space as a form of quick task-specific conditioning. Excellent generalization results with this combination. When tested on standard datasets, we obtain state of the art classification accuracies under various few-shot scenarios. As seen in Table, our method gives the best accuracy when compared with existing world-renowned few-shot learners.
Title: Deep Cooperative Sensing: Cooperative Spectrum Sensing Based on Convolutional Neural Networks
Authors: Woongsup Lee, Minhoe Kim, Dong-Ho Cho
In this paper, we investigate cooperative spectrum sensing (CSS) in a cognitive radio network (CRN) where multiple secondary users (SUs) cooperate in order to detect a primary user, which possibly occupies multiple bands simultaneously. Deep cooperative sensing (DCS), which constitutes the first CSS framework based on a convolutional neural network (CNN), is proposed. In DCS, instead of the explicit mathematical modeling of CSS, the strategy for combining the individual sensing results of the SUs is learned autonomously with a CNN using training sensing samples regardless of whether the individual sensing results are quantized or not. Moreover, both spectral and spatial correlation of individual sensing outcomes are taken into account such that an environment-specific CSS is enabled in DCS. Through simulations, we show that the performance of CSS can be greatly improved by the proposed DCS.
Figure 1. CSS with correlated individual spectrum sensing.
Title: Semidynamic cell-clustering algorithm based on reinforcement learning in cooperative transmission system
Authors: Byung Chang Chung, Dong-Ho Cho
In this paper, we propose a novel method of managing a semidynamic cluster through the use of a reinforcement learning. We derive some concepts from reinforcement learning that could be suitable for cooperative networks. We also verify the performance of proposed algorithm by means of a simulation, in which we examined spectral efficiency and convergence properties. The proposed algorithm represents a considerable improvement for edge users in particular. In addition, we analyze the complexity of the clustering schemes. Our proposed algorithm is effective in the environment where there is a limited computational resource.
2018년 8월 22일 열린 제1회 ‘AI 월드컵 2018’결선에서 KAIST 전기및전자공학부의 통신 디비젼 소속 AFC-WlSRL팀과 신호 디비젼 소속 SIIT팀이 1, 2위를 차지했다. AI 월드컵은 인공지능 플레이어 5개가 한팀으로 구성되어 기계학습과 인공지능으로 훈련을 받고 외부 지도 없이 축구를 하는 게임이다. AI 축구경기는 AI 전술과 실시간 행동 제어 알고리즘을 테스트해 비교하는 신흥 AI 플랫폼이다. 제 1회 인공지능 월드컵 2018에는 12개국 29개 팀이 참가해 우승을 겨뤘다. 1위와 2위 모두 KAIST 전기및전자공학부 출신인 것으로 밝혀져 KAIST 전기및전자공학부의 활발한 AI 연구 성과를 입증하고 위상을 드높였다.
Authors: Minhoe Kim, Nam-I Kim, Woongsup Lee, Dong-Ho Cho
Sparse code multiple access (SCMA) is a promising code-based non-orthogonal multiple-access technique that can provide improved spectral efficiency and massive connectivity meeting the requirements of 5G wireless communication systems. We propose a deep learning-aided SCMA (D-SCMA) in which the codebook that minimizes the bit error rate (BER) is adaptively constructed, and a decoding strategy is learned using a deep neural network-based encoder and decoder. One benefit of D-SCMA is that the construction of an efficient codebook can be achieved in an automated manner, which is generally difficult due to the non-orthogonality and multi-dimensional traits of SCMA. We use simulations to show that our proposed scheme provides a lower BER with a smaller computation time than conventional schemes.
Title: A novel PAPR reduction scheme for OFDM system based on deep learning
Authors: Minhoe Kim, Woongsup Lee, Dong-Ho Cho
High peak-to-average power ratio (PAPR) has been one of the major drawbacks of orthogonal frequency division multiplexing (OFDM) systems. In this letter, we propose a novel PAPR reduction scheme, known as PAPR reducing network (PRNet), based on the autoencoder architecture of deep learning. In the PRNet, the constellation mapping and demapping of symbols on each subcarrier is determined adaptively through a deep learning technique, such that both the bit error rate (BER) and the PAPR of the OFDM system are jointly minimized. We used simulations to show that the proposed scheme outperforms conventional schemes in terms of BER and PAPR.
Figure 1. Structure of proposed PRNet with DNN encoder and DNN decoder.
