一般口演
学習理論 / 神経回路モデル化と人工知能 1
Learning Theory / Neural Network Modeling and Artificial Intelligence 1
座長：下野 昌宣（京都大学） |
|---|
| 2022年7月1日 10:00～10:15 沖縄コンベンションセンター 会議場B5～7 第4会場
2O04m2-01
ゲート付海馬-前頭葉-視床回路モデルによる空間情報の階層的学習
Hierarchical spatial information learning in a gated prefrontal-thalamo-hippocampal circuit model
*高久 宗矩(1)、深井 朋樹(1)
1. 沖縄科学技術大学院
*Munenori Takaku(1), Tomoki Fukai(1)
1. Okinawa Institute of Science and Technology
Keyword: Hippocampus, Prefrontal cortex, Thalamus, Recurrent neural network
We encounter many different environments in our daily lives. It is one of the excellent abilities of human beings to cope with the unknown difficulties in those environments based on prior experience. In recent years, machine learning researchers have been trying to apply such human problem-solving abilities to machines. However, the proposed models might miss the contextual information of episodes, which is essential for retrieving prior experiences appropriately. In the neuroscience area, it is suggested that the prefrontal cortex, hippocampus, thalamus, and subregions of these areas relate to the memory function that requires contextual information. If the role of these interactions is clarified, we can construct a good memory model for machine learning.In this project, we developed a machine learning-based neuronal model for contextual memory and tested its abilities. The brain regions we will focus on are the hippocampus, prefrontal cortex, and thalamus, modeled by a network of Long Short-Term Memory (LSTM) units. They are considered as the modules in the proposed model and are connected to each other.At the core of the designed LSTM network is a gating function regulated by a thalamic input and a cortical unit that contains multi states of neuronal activities. With this model, We tried to reproduce the result of Delayed-nonmatch to Sample experiments on rodents, which requires contextual information, and check the activities of the model. As a result, the proposed model shows better results than the comparison model with fewer training sessions. Analysis of the firing rate of each neuron predicts the aspects of the modules; the thalamus module isswitching the state, the prefrontal cortex module constitutes a contextual attractor, and the hippocampus module can encode location information. This implies that transfer learning should be easy, and in fact, we have shown good results with a small number of training sessions on tasks with different trajectories.By comparing the proposed model and simple model, which doesn’t include the function of the thalamus module, the presence of the thalamus is expected to assist in constructing associative memory.Further study of this model is expected to investigate the prefrontal-thalamo-hippocampal circuit's role and construct a flexible memory system for artificial intelligence. | | 2022年7月1日 10:15～10:30 沖縄コンベンションセンター 会議場B5～7 第4会場
2O04m2-02
Development of a Data-driven Prediction Model for the Evolution of White Matter Hyperintensities using Deep Learning: Progress and Challenges
*Muhammad Febrian Rachmadi(1), Maria del C Valdés-Hernández(2), Joanna Wardlaw(2), Stephen Makin(3), Henrik Skibbe(1)
1. Brain Image Analysis Unit, RIKEN Center for Brain Science, Wako, Japan, 2. Centre for Clinical Brain Sciences, University of Edinburgh, Edinburgh, UK, 3. University of Aberdeen, Aberdeen, UK
Keyword: WHITE MATTER HYPERINTENSITIES, PREDICTION OF DISEASE PROGRESSION, DEEP LEARNING, SMALL VESSEL DISEASE
White matter hyperintensities (WMHs) are neuroradiological features often seen in T2-FLAIR brain MRI, characteristic of small vessel disease (SVD), and associated with stroke and dementia progression. Clinical studies indicate that the volume of WMHs on a patient may shrink, remain unchanged, or grow over time. We call this the “evolution of WMH”. Predicting the evolution of WMHs is challenging because the rate and direction of WMHs’ evolution vary considerably across individual brains, individuals, and clinical studies.
In this study, we introduce ours works that have been done on developing deep learning models for predicting these dynamic changes of WMHs, and the challenges ahead. To benchmark our models, we predicted the evolution of WMHs based on one single brain MRI scan sequence from each patient. We used brain MRI data from stroke patients enrolled in a study of stroke mechanisms, imaged by a GE 1.5T scanner at three time points (baseline scan, about 3 months, and a year after). We tested different input modalities ranging from a T2-FLAIR MRI, a probability map (i.e., output from a deep learning model for WMHs segmentation), and an irregularity map (i.e., output from an unsupervised model for WMHs segmentation called LOTS-IM).
Our studies have shown that (1) incorporating risk factors of WMHs evolution and (2) modeling uncertainty are important to improve the prediction of the evolution of WMHs.
(1) Several factors have been indicated by clinical studies to be associated with the evolution of WMHs, such as baseline WMHs volume and presence of stroke lesions. Thus, we incorporated such factors as additional model inputs.
(2) Furthermore, our study exposed that predicting the evolution of WMHs involves some levels of uncertainty, especially when predicting the areas of shrinking and growing WMHs. The uncertainty is due to the difficulty of distinguishing textures/intensities of shrinking and growing WMHs from the T2-FLAIR brain MRI sequence.
Our recent model, equipped with a conditional variational autoencoder for modelling the uncertainty and additional inputs for incorporating clinical risk factors, has achieved the best performance compared to the other deep learning models, including the vanilla deep learning model of U-Net. | | 2022年7月1日 10:30～10:45 沖縄コンベンションセンター 会議場B5～7 第4会場
2O04m2-03
階層事前分布を導入した構造化ベイズモデルによるMEG信号源推定のノイズ耐性と情報拡散
Noise tolerance and information spreading in MEG source estimation using a structured Bayesian model with hierarchical prior
*宮﨑 海(1)、韮澤 駿(1)、赤松 和昌(1)、山下 宙人(2,3)、宮脇 陽一(1)
1. 電気通信大学、2. 国際電気通信基礎技術研究所、3. 革新知能統合研究センター
*Kai Miyazaki(1), Shun Nirasawa(1), Kazuaki Akamatsu(1), Okito Yamashita(2,3), Yoichi Miyawaki(1)
1. The University of Electro-Communications, 2. Advanced Telecommunications Research Institutes International, 3. RIKEN Center for Advanced Intelligence Project
Keyword: MEG, SOURCE LOCALIZATION, MACHINE LEARNING, DECODING
Magnetoencephalography (MEG) acquires human brain activity at a high temporal resolution, but its spatial resolution is insufficient to reveal neural mechanisms at a fine scale. A method to resolve this issue is source estimation of MEG signals (Hämäläinen et al., 1994, Iwaki et al., 1998, Sato et al., 2004). However, our previous study (Sato et al., 2018) showed that the combination of the source estimation and multivariate decoding analysis produces “information spreading,” a phenomenon showing significant decoding accuracy outside of true source locations, resulting in false-positive interpretations about informative brain areas. To suppress this phenomenon, we proposed a structured Bayesian source estimation model using grouped automatic relevance determination (gARD) (Yu et al., 2015), demonstrating better performance in suppression of information spreading than conventional source estimation models (Ishibashi et al., 2020). However, we also found that our model is susceptible to observation noise. To resolve this issue, we here introduced a hyperprior that represented possible source locations in the Bayesian statistics framework and aimed at achieving accurate source localization and noise tolerance. Results showed that the model with the hyperprior achieved better source localization and tolerance even under noisy conditions than did the other conventional models. Searchlight decoding (Kriegeskorte et al., 2006) showed significant prediction over the true source location, but its extent spread larger than the original model without the hyperprior. This counterintuitive result might partially originate from the introduction of hyperprior because it generally makes a model prefer non-sparse solutions. These results indicate that our hierarchical model may be useful for identifying source locations even under a low signal-to-noise environment but still insufficient for suppressing information spreading, suggesting that the model needs appropriate balance to control how much hyperprior information is recruited to achieve better source localization with narrow information spreading. | | 2022年7月1日 10:45～11:00 沖縄コンベンションセンター 会議場B5～7 第4会場
2O04m2-04
確率的神経システムにおける状態遷移の制御コスト
Control Cost for State Transitions in Stochastic Neural Systems
*神谷 俊輔(1)、川北 源二(1)、笹井 俊太朗(2)、北園 淳(1)、大泉 匡史(1)
1. 東京大学、2. 株式会社アラヤ
*Shunsuke Kamiya(1), Genji Kawakita(1), Shuntaro Sasai(2), Jun Kitazono(1), Masafumi Oizumi(1)
1. The University of Tokyo, 2. Araya Inc.
Keyword: BRAIN STATE TRANSITION, STATE TRANSITION COST, CONTROL THEORY, SCHRODINGER'S BRIDGE
Our brain is considered a dynamical system that flexibly transitions to various states. Among these states, some are difficult to reach while others are easy. Evaluating difficulty on transitions can deepen our understanding of brain dynamics, thus quantifying brain state transition costs is important. For this goal, a metric called control energy has been proposed by applying network control theory to neuroscience. However, the control energy, based on a deterministic framework, can neither take into account stochasticity in the brain nor can assess costs necessary to change co-variation among the brain regions, i.e., functional connectivity (FC). Here we propose a novel method for evaluating transition costs based on a stochastic model. We applied a framework by our previous study (Kawakita et al. Network Neuroscience. 2022), where the cost is quantified as the minimal Kullback-Leibler (KL) divergence between the uncontrolled and controlled processes, to a linear stochastic dynamics. Utilizing the equivalence of the KL divergence and the expected total control input, we analytically derived the optimal total input and defined a stochastic version of the control cost. Furthermore, we found that this cost can be decomposed into two parts, the cost for controlling the mean brain activity (mean control) and the covariance among brain areas (covariance control). We proved that the former is identical to the control energy that has conventionally used in the deterministic setting, making the stochastic cost a natural extension of its deterministic counterpart. The covariance control corresponds to the cost to control FC, which has not been quantified in the past studies. We also demonstrate that this decomposition of mean and covariance can be performed at regional level, i.e., we can quantify control inputs in a brain region that contribute to alter mean brain activities and FC. As a proof of our novel concept, we applied this metric for functional magnetic resonance imaging (fMRI) data in Human Connectome Project. We showed that the stochastic cost can correspond to the difficulty of cognitive tasks. We also found that the areas that contribute to alter mean brain activity are close to the areas that get active in the same task. Plus, we revealed the posterior cingulate cortex (PCC) is related to regulating covariance in many tasks, implying that PCC plays important roles in FC reconfiguration in brain state transitions. | |
|
|
