arXiv subjects

Taehoon Kim

Publications and source records attributed to Taehoon Kim.

43 records · Page 3Linked to original sources

arXiv2019

Quantifying Generalization in Reinforcement Learning

In this paper, we investigate the problem of overfitting in deep reinforcement learning. Among the most common benchmarks in RL, it is customary to use the same environments for both training and testing. This practice offers relatively little insight into an agent's ability to generalize. We address this issue by using procedurally generated environments to construct distinct training and test sets. Most notably, we introduce a new environment called CoinRun, designed as a benchmark for generalization in RL. Using CoinRun, we find that agents overfit to surprisingly large training sets. We then show that deeper convolutional architectures improve generalization, as do methods traditionally found in supervised learning, including L2 regularization, dropout, data augmentation and batch normalization.

cs.LG↗

arXiv2019

Transfer Learning via Unsupervised Task Discovery for Visual Question Answering

We study how to leverage off-the-shelf visual and linguistic data to cope with out-of-vocabulary answers in visual question answering task. Existing large-scale visual datasets with annotations such as image class labels, bounding boxes and region descriptions are good sources for learning rich and diverse visual concepts. However, it is not straightforward how the visual concepts can be captured and transferred to visual question answering models due to missing link between question dependent answering models and visual data without question. We tackle this problem in two steps: 1) learning a task conditional visual classifier, which is capable of solving diverse question-specific visual recognition tasks, based on unsupervised task discovery and 2) transferring the task conditional visual classifier to visual question answering models. Specifically, we employ linguistic knowledge sources such as structured lexical database (e.g. WordNet) and visual descriptions for unsupervised task discovery, and transfer a learned task conditional visual classifier as an answering unit in a visual question answering model. We empirically show that the proposed algorithm generalizes to out-of-vocabulary answers successfully using the knowledge transferred from the visual dataset.

cs.LG↗

arXiv2019

An Enhanced Random Access with Preamble-Assisted Short-Packet Transmissions for Cellular IoT Communications

We propose an enhanced random access (RA) with preamble-assisted short-packet transmissions to support cellular Internet-of-things (IoT) communications. A key feature of the proposed scheme is that the base station (e.g., eNodeB in LTE networks) utilizes the RA preambles as uplink reference signals to estimate uplink channel state information of IoT devices, which subsequently provides additional opportunities to transmit short-packets during the RA procedure without extra signaling (e.g., connection and scheduling requests). Through simulations, we evaluate the performance of the proposed scheme in terms of a channel estimation mean squared error, a bit error rate, a preamble collision probability, and a success probability of short-packet transmissions. The results show that the proposed scheme can support reliable and low-latency featured short-packet transmissions during the RA procedure by efficiently incorporating both multiple antenna and detection techniques.

cs.IT↗

arXiv2018

Music Source Separation Using Stacked Hourglass Networks

In this paper, we propose a simple yet effective method for multiple music source separation using convolutional neural networks. Stacked hourglass network, which was originally designed for human pose estimation in natural images, is applied to a music source separation task. The network learns features from a spectrogram image across multiple scales and generates masks for each music source. The estimated mask is refined as it passes over stacked hourglass modules. The proposed framework is able to separate multiple music sources using a single network. Experimental results on MIR-1K and DSD100 datasets validate that the proposed method achieves competitive results comparable to the state-of-the-art methods in multiple music source separation and singing voice separation tasks.

cs.SD↗

arXiv2017

An Enhanced Access Reservation Protocol with a Partial Preamble Transmission Mechanism in NB-IoT Systems

In this letter, we propose an enhanced Access Reservation Protocol (ARP) with a partial preamble transmission (PPT) mechanism for the narrow band Internet of Things (NB-IoT) systems. The proposed ARP can enhance the ARP performance by mitigating the occurrence of preamble collisions, while being compatible with the conventional NB-IoT ARP. We provide an analytical model that captures the performance of the proposed ARP in terms of false alarm, mis-detection and collision probabilities. Moreover, we investigate a trade-off between the mis-detection and the collision probabilities, and optimize the proposed ARP according to the system loads. The results show that the proposed ARP outperforms the conventional NB-IoT ARP, in particular at heavier system loads.

cs.IT↗

arXiv2011

Ultra-thin SWNTs Films with Tunable, Anisotropic Transport Properties

Directional transport properties at the nanoscale remain a challenge primarily due to issues associated with control over the underlying anisotropy and scalability to macroscopic scales. In this letter, we develop a facile approach based on template-guided fluidic assembly of high mobility building blocks - single walled carbon nanotubes (SWNTs) - to fabricate ultra-thin and anisotropic SWNT films. A major advancement is the complete control over the anisotropy in the assembled nanostructure, realized by three-dimensional engineering of dip-coated SWNT thin films into alternating hydrophilic and hydrophobic micro-line patterns with prescribed intra/inter-line widths and line thicknesses. Variations in the contact line profile results in an evaporation-controlled assembly mechanism that leads to the formation of an alternating, and more importantly, contiguous SWNT network. Evidently, the nanoscopic thickness modulations are direct reflections of the substrate geometry and chemistry. The nanostructured film exhibits significant anisotropy in their electrical and thermal transport properties as well as optical transparency, as revealed by characterization studies. The direct interplay between the anisotropy and the 3D micro-line patterns of the substrate combined with the wafer-level scalability of the fluidic assembly allows us to tune the transport properties for a host of nanoelectronic applications.

cond-mat.mes-hall↗

arXiv2010

Topological transitions in carbon nanotube networks via nanoscale confinement

Efforts aimed at large-scale integration of nanoelectronic devices that exploit the superior electronic and mechanical properties of single-walled carbon nanotubes (SWCNTs) remain limited by the difficulties associated with manipulation and packaging of individual SWNTs. Alternative approaches based on ultra-thin carbon nanotube networks (CNNs) have enjoyed success of late with the realization of several scalable device applications. However, precise control over the network electronic transport is challenging due to i) an often uncontrollable interplay between network coverage and its topology and ii) the inherent electrical heterogeneity of the constituent SWNTs. In this letter, we use template-assisted fluidic assembly of SWCNT networks to explore the effect of geometric confinement on the network topology. Heterogeneous SWCNT networks dip-coated onto sub-micron wide ultra-thin polymer channels exhibit a topology that becomes increasingly aligned with decreasing channel width and thickness. Experimental scale coarse-grained computations of interacting SWCNTs show that the effect is a reflection of an aligned topology that is no longer dependent on the network density, which in turn emerges as a robust knob that can induce semiconductor-to-metallic transitions in the network response. Our study demonstrates the effectiveness of directed assembly on channels with varying degrees of confinement as a simple tool to tailor the conductance of the otherwise heterogeneous network, opening up the possibility of robust large-scale CNN-based devices.

cond-mat.mtrl-sci↗