Kaist

KAIST Reveals Mathematical Principle behind AI’s ‘..

(from left: Professor Jong Chul Ye, PhD candidates Yoseob Han and Eunju Cha) A KAIST research team identified the geometrical structure of artificial intelligence (AI) and discovered the mathematical principles of highly performing artificial neural networks, which can be applicable in fields such as medical imaging. Deep neural networks are an exemplary method of implementing deep learning, which is at the core of the AI technology, and have shown explosive growth in recent years. This technique has been used in various fields, such as image and speech recognition as well as image processing. Despite its excellent performance and usefulness, the exact working principles of deep neural networks has not been well understood, and they often suffer from unexpected results or errors. Hence, there is an increasing social and technical demand for interpretable deep neural network models. To address these issues, Professor Jong Chul Ye from the Department of Bio & Brain Engineering and his team attempted to find the geometric structure in a higher dimensional space where the structure of the deep neural network can be easily understood. They proposed a general deep learning framework, called deep convolutional framelets, to understand the mathematical principle of a deep neural network in terms of the mathematical tools in Harmonic analysis. As a result, it was found that deep neural networks’ structure appears during the process of decomposition of high dimensionally lifted signal via Hankel matrix, which is a high-dimensional structure formerly studied intensively in the field of signal processing. In the process of decomposing the lifted signal, two bases categorized as local and non-local basis emerge. The researchers found that non-local and local basis functions play a role in pooling and filtering operation in convolutional neural network, respectively. Previously, when implementing AI, deep neural networks were usually constructed through empirical trial and errors. The significance of the research lies in the fact that it provides a mathematical understanding on the neural network structure in high dimensional space, which guides users to design an optimized neural network. They demonstrated improved performance of the deep convolutional framelets’ neural networks in the applications of image denoising, image pixel in painting, and medical image restoration. Professor Ye said, “Unlike conventional neural networks designed through trial-and-error, our theory shows that neural network structure can be optimized to each desired application and are easily predictable in their effects by exploiting the high dimensional geometry. This technology can be applied to a variety of fields requiring interpretation of the architecture, such as medical imaging.” This research, led by PhD candidates Yoseob Han and Eunju Cha, was published in the April 26th issue of the SIAM Journal on Imaging Sciences. Figure 1. The design of deep neural network using mathematical principles Figure 2. The results of image noise cancelling Figure 3. The artificial neural network restoration results in the case where 80％ of the pixels are lost

Deep learning predicts drug-drug and drug-food int..

(Distinguished Professor Sang Yup Lee) A Korean research team from KAIST developed a computational framework, DeepDDI, that accurately predicts and generates 86 types of drug-drug and drug-food interactions as outputs of human-readable sentences, which allows in-depth understanding of the drug-drug and drug-food interactions. Drug interactions, including drug-drug interactions (DDIs) and drug-food constituent interactions (DFIs), can trigger unexpected pharmacological effects, including adverse drug events (ADEs), with causal mechanisms often unknown. However, current prediction methods do not provide sufficient details beyond the chance of DDI occurrence, or require detailed drug information often unavailable for DDI prediction. To tackle this problem, Dr. Jae Yong Ryu, Assistant Professor Hyun Uk Kim and Distinguished Professor Sang Yup Lee, all from the Department of Chemical and Biomolecular Engineering at Korea Advanced Institute of Science and Technology (KAIST), developed a computational framework, named DeepDDI, that accurately predicts 86 DDI types for a given drug pair. The research results were published online in Proceedings of the National Academy of Sciences of the United States of America (PNAS) on April 16, 2018, which is entitled “Deep learning improves prediction of drug-drug and drug-food interactions.” DeepDDI takes structural information and names of two drugs in pair as inputs, and predicts relevant DDI types for the input drug pair. DeepDDI uses deep neural network to predict 86 DDI types with a mean accuracy of 92.4％ using the DrugBank gold standard DDI dataset covering 192,284 DDIs contributed by 191,878 drug pairs. Very importantly, DDI types predicted by DeepDDI are generated in the form of human-readable sentences as outputs, which describe changes in pharmacological effects and/or the risk of ADEs as a result of the interaction between two drugs in pair. For example, DeepDDI output sentences describing potential interactions between oxycodone (opioid pain medication) and atazanavir (antiretroviral medication) were generated as follows: “The metabolism of Oxycodone can be decreased when combined with Atazanavir”; and “The risk or severity of adverse effects can be increased when Oxycodone is combined with Atazanavir”. By doing this, DeepDDI can provide more specific information on drug interactions beyond the occurrence chance of DDIs or ADEs typically reported to date. DeepDDI was first used to predict DDI types of 2,329,561 drug pairs from all possible combinations of 2,159 approved drugs, from which DDI types of 487,632 drug pairs were newly predicted. Also, DeepDDI can be used to suggest which drug or food to avoid during medication in order to minimize the chance of adverse drug events or optimize the drug efficacy. To this end, DeepDDI was used to suggest potential causal mechanisms for the reported ADEs of 9,284 drug pairs, and also predict alternative drug candidates for 62,707 drug pairs having negative health effects to keep only the beneficial effects. Furthermore, DeepDDI was applied to 3,288,157 drug-food constituent pairs (2,159 approved drugs and 1,523 well-characterized food constituents) to predict DFIs. The effects of 256 food constituents on pharmacological effects of interacting drugs and bioactivities of 149 food constituents were also finally predicted. All these prediction results can be useful if an individual is taking medications for a specific (chronic) disease such as hypertension or diabetes mellitus type 2. Distinguished Professor Sang Yup Lee said, “We have developed a platform technology DeepDDI that will allow precision medicine in the era of Fourth Industrial Revolution. DeepDDI can serve to provide important information on drug prescription and dietary suggestions while taking certain drugs to maximize health benefits and ultimately help maintain a healthy life in this aging society.” Figure 1. Overall scheme of Deep DDDI and prediction of food constituents that reduce the in vivo concentration of approved drugs

KAIST Develops Sodium Ion Batteries using Copper S..

(Dr. Sungjoo Kim, PhD candidates Hyeon Kook Seo, Jae Yeol Park and Joon Ha Chang, Emeritus Professor Jeong Yong Lee and Professor Jong Min Yuk) A KAIST research team recently developed sodium ion batteries using copper sulfide anode. This finding will contribute to advancing the commercialization of sodium ion batteries (SIBs) and reducing the production cost of any electronic products with batteries. Professor Jong Min Yuk and Emeritus Professor Jeong Yong Lee from Department of Materials Science and Engineering succeeded in developing a new anode material suitable for use in a SIB. Compared to the existing anode materials, the copper sulfide anode was measured to exhibit 1.5 times better cyclability with projected 40％ reduction in cost. Batteries used in various applications including mobile phones are lithium ion batteries, mostly referred as Li-ion batteries or LIBs. Though they are popularly used until now, large-scale energy storage systems require much inexpensive and abundant materials. Hence, a SIB has attracted enormous attention for their advantage over a lithium counterpart. However, one main obstacle to commercialization of SIB is the lack of suitable anodes that exhibit high capacity and the cycling stability of the battery. Hence, the research team recognized this need for a good anode material that could offer high electrical conductivity and theoretical capacity. The material was found to be copper sulfide, preferably in nanoplates, which “prefers to make an alloy with sodium and is thus promising for high capacity and long-term cyclability.” Further analysis presented in the study reveals that copper sulfide undergoes crystallographic tuning to make a room for sodium insertion. Results indicate that the sodium ion-insertion capacity of copper sulfide is as much as 1.5 times that of lithium ions for graphite. Furthermore, a battery with this new anode material retains 90％ of its original capacity for 250 charge-discharge cycles. With the natural abundance of sodium in seawater, this development may contribute to reduction in battery costs, which can be translated into up to 30％ cut in the price of various consumer electronics. Professor Lee expressed his hope for “the production of next-generation, high-performance sodium ion batteries”. Professor Yuk said, “These days, people are showing a great deal of interest in products related to renewable energy due to recent micro-dust issues ongoing in Korea. This study may help Korea get a head-start on renewable energy products”. This research, led by PhD candidate Jae Yeol Park and Dr. Sung Joo Kim, was published online in Nature Communications on March 2. Figure 1. The sodiation process of copper sulfide

KAIST Succeeds in Producing 50x More Stable Adsorb..

(From left: Professor Minkee Choi and PhD candidate Woosung Choi) A KAIST research team developed a technology to increase the stability of amine-containing adsorbents by fifty times, moving one step further toward commercializing stable adsorbents that last longer. Professor Minkee Choi from the Department of Chemical and Biomolecular Engineering and his team succeeded in developing amine-containing adsorbents that show high oxidative stability. The capture of the greenhouse gas carbon dioxide is an active ongoing research field, and some of the latest advancements point to amine-containing adsorbents as an efficient and environment-friendly way to capture carbon dioxide. However, existing amine-containing adsorbents are known to be unstable under oxidation, which chemically breaks down the adsorbent, thereby making it difficult to rely on amine-containing adsorbents for repeated and continued use. The researchers have discovered that the miniscule amount of iron and copper present in the amine accelerate the oxidative breakdown of the amine-containing adsorbent. Upon this discovery, they proposed the use of a chelator substance, which essentially suppresses the activation of the impurities. The team demonstrates that the proposed method renders the adsorbent up to 50 times slower in its deactivation rate due to oxidation, compared to conventional polyethyleneimine (PEI) / silica adsorbents. Figure 1 illustrates the superior performance of this oxidation-stable amine-containing adsorbent (shown in black squares), whose carbon dioxide-capturing capacity deteriorates by only a small amount (~8％). Meanwhile, the carbon dioxide-capturing capacity of the PEI/silica adsorbent (shown in red diamonds) degrades dramatically after being exposed to oxidative aging for 30 days. This stability under oxidation is expected to have brought amine-containing adsorbents one step closer to commercialization. As such, first author Woosung Choi describes the significance of this study as “having brought solid carbon dioxide adsorbents to commercializable standards”. In fact, Professor Choi explains that commercialization steps for his team’s carbon dioxide adsorbents are already underway. He further set forth his aim to “develop the world’s best carbon dioxide capture adsorbent”. This research, led by the PhD candidate Woosung Choi, was published online in Nature Communications on February 20. Figure 1. Carbon dioxide working capacity against oxidative aging time. Performance of the proposed method (black) degrades much more slowly (~50x) than that of existing methods. The novel adsorbent is thus shown to be more robust to oxidation.

KAIST Discloses the Formation of Burning Ice in Oc..

(from left: Professor Tae-Hyuk Kwon and PhD candidate Taehyung Park) ?A KAIST research team has identified the formation of natural gas hydrates, so-called flammable ice, formed in oceans. Professor Tae-Hyuk Kwon from the Department of Civil & Environmental Engineering and his team found that clay minerals in oceanic clay-rich sedimentary deposits promote formation of gas hydrates and proposed the principle of gas hydrate formation in the clayey sedimentary layers. Gas hydrates are ice-like crystalline structures composed of hydrogen-bonded water molecules encapsulating gas molecules. They are also known as burning ice. Their deposits are so huge that they gain attention for alternative energy. Conventionally, it was believed that formation of gas hydrates is limited in clay sedimentary deposits; however, unexpected abundance of natural gas hydrates in oceanic clay-rich sedimentary deposits raised the issue of how they formed. The surfaces of natural clay minerals are negatively charged and, thus, unavoidably generate physicochemical interactions between clay and water. Such clay-water interactions have a critical role in the occurrence of natural gas hydrates in clay-rich sedimentary formations. However, there has been experimental difficulty in analyzing hydrate formation because of the cations contained in clay particles, which balance the clay surface charges. Therefore, clay particles inevitably release the cations when mixed with water, which complicates the interpretation of experimental results. To overcome this limitation, the team polarized water molecules with an electric field and monitored the induction times of water molecules forming gas hydrates. They found that the 10 kV/m of electric field promoted gas hydrate nucleation under certain conditions rather than slowing it down, due to the partial breakage of the hydrogen bonded water clusters and the lowered thermal energy of water molecules. Professor Kwon said, “Through this research, we gained better insight into the origin of gas hydrates occurrence in clay-rich sedimentary deposits. In the near future, we will soon be able to commercially produce methane gas from natural gas hydrate deposits.” This research, led by PhD candidate Taehyung Park, was published online in Environmental Science and Technology on February 3. (doi: 10.1021/acs.est.7b05477) Figure 1. Formation of gas hydrates with water molecules Figure 2. Enhancement and inhibition of gas hydrates

Printed Thermo-Plasmonic Heat Patterns for Neurolo..

(Professor Nam and Dr. Kang, right) A KAIST team presented a highly customizable neural stimulation method. The research team developed a technology that can print the heat pattern on a micron scale to enable the control of biological activities remotely. The researchers integrated a precision inkjet printing technology with bio-functional thermo-plasmonic nanoparticles to achieve a ‘selective nano-photothermal neural stimulation method.’ The research team of Professor Yoonkey Nam at the Department of Bio and Brain Engineering expects this will serve as an enabling technology for personalized precision neuromodulation therapy for patients with neurological disorders. The nano-photothermal neural stimulation method uses the thermo-plasmonic effect of metal nanoparticles to modulate the activities of neuronal networks. With the thermo-plasmonic effect, metal nanoparticles can absorb specific wavelength of illuminated light to efficiently generate localized heat. The research team discovered the inhibitory behavior of spontaneous activities of neurons upon photothermal stimulation four years ago. Since then, they have developed this technology to control hyperactive behaviors of neurons and neural circuits, which is often found in neurological disorders such as epilepsy. In order to overcome the limitation on the spatial selectivity and resolution of the previously developed nano-photothermal method, the team adopted an inkjet printing technology to micro pattern the plasmonic nanoparticles (a few tens of microns), and successfully demonstrated that the nano-photothermal stimulation can be selectively applied according to the printed patterns. The researchers applied a polyelectrolyte layer-by-layer coating method to printing substrates in a way to improve the pattern fidelity and achieve the uniform assembly of nanoparticles. The electrostatic attraction between the printed nanoparticles and the coated printing substrate also helped the stability of the attached nanoparticles. Because the polyelectrolyte coating is biocompatible, biological experiments including cell culture are possible with the technology developed in this work. Using printed gold nanorod particles in a few tens of microns resolution over a several centimeters area, the researchers showed that highly complex heat patterns can be precisely formed upon light illumination according to the printing image. Lastly, the team confirmed that the printed heat patterns can selectively and instantaneously inhibit the activities of cultured hippocampal neurons upon near-infrared light illumination. Because the printing process is applicable to thin and flexible substrates, the technology can be easily applied to implantable neurological disorder treatment devices and wearable devices. By selectively applying the heat patterns to only the desired cellular areas, customized and personalized photothermal neuromodulation therapy can be applied to patients. “The fact that any desired heat patterns can be simply ‘printed’ anywhere broadens the applicability of this technology in many engineering fields. In bioengineering, it can be applied to neural interfaces using light and heat to modulate physiological functions. As another engineering application, for example, printed heat patterns can be used as a new concept of anti-counterfeit applications,” said the principal investigator, Yoonkey Nam at KAIST. This work, led mainly by Dr. Hongki Kang, was published in ACS Nano on February 5th 2018.

Recognizing Seven Different Face Emotions on a Mob..

(Professor Hoi-Jun Yoo) A KAIST research team succeeded in achieving face emotion recognition on a mobile platform by developing an AI semiconductor IC that processes two neural networks on a single chip. Professor Hoi-Jun Yoo and his team (Primary researcher: Jinmook Lee Ph. D. student) from the School of Electrical Engineering developed a unified deep neural network processing unit (UNPU). Deep learning is a technology for machine learning based on artificial neural networks, which allows a computer to learn by itself, just like a human. The developed chip adjusts the weight precision (from 1 bit to 16 bit) of a neural network inside of the semiconductor in order to optimize energy efficiency and accuracy. With a single chip, it can process a convolutional neural network (CNN) and recurrent neural network (RNN) simultaneously. CNN is used for categorizing and recognizing images while RNN is for action recognition and speech recognition, such as time-series information. Moreover, it enables an adjustment in energy efficiency and accuracy dynamically while recognizing objects. To realize mobile AI technology, it needs to process high-speed operations with low energy, otherwise the battery can run out quickly due to processing massive amounts of information at once. According to the team, this chip has better operation performance compared to world-class level mobile AI chips such as Google TPU. The energy efficiency of the new chip is 4 times higher than the TPU. In order to demonstrate its high performance, the team installed UNPU in a smartphone to facilitate automatic face emotion recognition on the smartphone. This system displays a user’s emotions in real time. The research results for this system were presented at the 2018 International Solid-State Circuits Conference (ISSCC) in San Francisco on February 13. Professor Yoo said, “We have developed a semiconductor that accelerates with low power requirements in order to realize AI on mobile platforms. We are hoping that this technology will be applied in various areas, such as object recognition, emotion recognition, action recognition, and automatic translation. Within one year, we will commercialize this technology.”

Aqueous Storage Device Needs Only 20 Seconds to Go

(from left: PhD candidate Il Woo Ock and Professor Jeung Ku Kang) A KAIST research team developed a new hybrid energy storage device that can be charged in less than half a minute. It employs aqueous electrolytes instead of flammable organic solvents, so it is both environmentally friendly and safe. It also facilitates a boosting charge with high energy density, which makes it suitable for portable electronic devices. Professor Jeung Ku Kang and his team from the Graduate School of Energy, Environment, Water, and Sustainability developed this hybrid energy storage with high energy and power densities along over a long cycle life by assembling fibre-like polymer chain anodes and sub-nanoscale metal oxide cathodes on graphene. Conventional aqueous electrolyte-based energy storage devices have a limitation for boosting charges and high energy density due to low driving voltage and a shortage of anode materials. Energy storage device capacity is determined by the two electrodes, and the balance between cathode and anode leads to high stability. In general, two electrodes show differences in electrical properties and differ in ion storage mechanism processes, resulting in poor storage and stability from the imbalance. The research team came up with new structures and materials to facilitate rapid speed in energy exchange on the surfaces of the electrodes and minimize the energy loss between the two electrodes. The team made anodes with graphene-based polymer chain materials. The web-like structure of graphene leads to a high surface area, thereby allowing higher capacitance. For cathode materials, the team used metal oxide in sub-nanoscale structures to elevate atom-by-ion redox reactions. This method realized higher energy density and faster energy exchange while minimizing energy loss. The developed device can be charged within 20 to 30 seconds using a low-power charging system, such as a USB switching charger or a flexible photovoltaic cell. The developed aqueous hybrid energy device shows more than 100-fold higher power density compared to conventional aqueous batteries and can be rapidly recharged. Further, the device showed high stability with its capacity maintained at 100％ at a high charge/discharge current. Professor Kang said, “This eco-friendly technology can be easily manufactured and is highly applicable. In particular, its high capacity and high stability, compared to existing technologies, could contribute to the commercialization of aqueous capacitors. The device can be rapidly charged using a low-power charging system, and thus can be applied to portable electronic device.” This research, led by a PhD candidate Il Woo Ock, was published in Advanced Energy Materials on January 15. Figure 1. Switching wearable LED kit with two AHCs in series charged by a flexible photovoltaic cell ?Figure 2. Schematic diagram for aqueous hybrid capacitors Figure 3. TEM images of anode and cathode