Soyoung Yang

Bio: Soyoung Yang is an academic researcher from KAIST. The author has contributed to research in topics: Machine translation & Meta learning (computer science). The author has an hindex of 2, co-authored 8 publications receiving 390 citations.

Restoring and Mining the Records of the Joseon Dynasty via Neural Language Modeling and Machine Translation.

Abstract: Understanding voluminous historical records provides clues on the past in various aspects, such as social and political issues and even natural science facts. However, it is generally difficult to fully utilize the historical records, since most of the documents are not written in a modern language and part of the contents are damaged over time. As a result, restoring the damaged or unrecognizable parts as well as translating the records into modern languages are crucial tasks. In response, we present a multi-task learning approach to restore and translate historical documents based on a self-attention mechanism, specifically utilizing two Korean historical records, ones of the most voluminous historical records in the world. Experimental results show that our approach significantly improves the accuracy of the translation task than baselines without multi-task learning. In addition, we present an in-depth exploratory analysis on our translated results via topic modeling, uncovering several significant historical events.

Unsupervised Neural Machine Translation for Low-Resource Domains via Meta-Learning

Abstract: Unsupervised machine translation, which utilizes unpaired monolingual corpora as training data, has achieved comparable performance against supervised machine translation. However, it still suffers from data-scarce domains. To address this issue, this paper presents a novel meta-learning algorithm for unsupervised neural machine translation (UNMT) that trains the model to adapt to another domain by utilizing only a small amount of training data. We assume that domain-general knowledge is a significant factor in handling data-scarce domains. Hence, we extend the meta-learning algorithm, which utilizes knowledge learned from high-resource domains, to boost the performance of low-resource UNMT. Our model surpasses a transfer learning-based approach by up to 2-3 BLEU scores. Extensive experimental results show that our proposed algorithm is pertinent for fast adaptation and consistently outperforms other baselines.

Crystallization and preliminary X-ray crystallographic studies of the PDZ domain of Shank1 from Rattus norvegicus.

Abstract: Shank proteins are a new family of scaffold proteins interacting with various membrane and cytoplasmic proteins. Shank contains multiple protein–protein interaction sites, including ankyrin repeats, an SH3 domain, a PDZ domain, a long proline-rich region and an SAM domain. The PDZ domain of Shank binds to the C-terminus of guanylate kinase-associated protein (GKAP). The PDZ domain of Shank1 from Rattus norvegicus and its complex with the C-terminal octapeptide of GKAP were crystallized at 294 K using polyethylene glycol 20 000 and 6000 as precipitants. Diffraction data sets from a peptide-free crystal and a complex crystal were collected to 1.8 and 3.2 A resolution, respectively, using synchrotron radiation. The peptide-free crystal belongs to space group P21, with unit-cell parameters a = 42.0, b = 50.3, c = 51.8 A, β = 106.3°. The complex crystal belongs to space group P212121, with unit-cell parameters a = 89.4, b = 97.5, c = 108.3 A.

Meta-Learning for Low-Resource Unsupervised Neural MachineTranslation.

Abstract: Unsupervised machine translation, which utilizes unpaired monolingual corpora as training data, has achieved comparable performance against supervised machine translation. However, it still suffers from data-scarce domains. To address this issue, this paper presents a meta-learning algorithm for unsupervised neural machine translation (UNMT) that trains the model to adapt to another domain by utilizing only a small amount of training data. We assume that domain-general knowledge is a significant factor in handling data-scarce domains. Hence, we extend the meta-learning algorithm, which utilizes knowledge learned from high-resource domains to boost the performance of low-resource UNMT. Our model surpasses a transfer learning-based approach by up to 2-4 BLEU scores. Extensive experimental results show that our proposed algorithm is pertinent for fast adaptation and consistently outperforms other baseline models.

Structural basis of long-term potentiation in single dendritic spines

Abstract: Dendritic spines of pyramidal neurons in the cerebral cortex undergo activity-dependent structural remodelling that has been proposed to be a cellular basis of learning and memory. How structural remodelling supports synaptic plasticity, such as long-term potentiation, and whether such plasticity is input-specific at the level of the individual spine has remained unknown. We investigated the structural basis of long-term potentiation using two-photon photolysis of caged glutamate at single spines of hippocampal CA1 pyramidal neurons. Here we show that repetitive quantum-like photorelease (uncaging) of glutamate induces a rapid and selective enlargement of stimulated spines that is transient in large mushroom spines but persistent in small spines. Spine enlargement is associated with an increase in AMPA-receptor-mediated currents at the stimulated synapse and is dependent on NMDA receptors, calmodulin and actin polymerization. Long-lasting spine enlargement also requires Ca2+/calmodulin-dependent protein kinase II. Our results thus indicate that spines individually follow Hebb's postulate for learning. They further suggest that small spines are preferential sites for long-term potentiation induction, whereas large spines might represent physical traces of long-term memory.

MAPK cascade signalling and synaptic plasticity

Abstract: The mitogen-activated protein kinase (MAPK) cascade that leads to the activation of extracellular signal-regulated kinases-1 and -2 (ERK1 and ERK2) has a key role in the differentiation of some cell types and the proliferation of others. However, several recent reports implicate this cascade in the control of synaptic plasticity in the adult brain. ERK signalling seems to be essential for characterized neuronal transcriptional events, and might also regulate synaptic targets to control plasticity. Another recently emerging story is the involvement of a 'parallel' but distinct kinase cascade leading to the activation of p38 MAPK, which might control distinct forms of synaptic plasticity.

Dendritic spines: structure, dynamics and regulation.

Abstract: Dendritic spines are tiny protrusions that receive excitatory synaptic input and compartmentalize postsynaptic responses. Heterogeneous in size and shape, and modifiable by activity and experience, dendritic spines have long been thought to provide a morphological basis for synaptic plasticity. Although advanced imaging techniques have highlighted the rapid and regulated motility of spines in living neurons, the functional significance of spine plasticity remains elusive. Recent insights into the molecular mechanisms that regulate spine morphogenesis offer potential ways to manipulate dendritic spines in vivo and to explore their physiological roles in the brain.