Symposium
“Towards integration of neuroscience and machine intelligence”
シンポジウム
“脳科学と機械知能の融合に向けて” |
|---|
| 7月25日（木）14:40～14:50 第6会場（朱鷺メッセ 2F 201A）
1S06a-1
機械学習がヒトイメージング研究に与えた影響
Okito Yamashita（山下 宙人）1,2
1理研革新知能統合研究センター
2国際電気通信基礎技術研究所 脳情報解析研究所
The artificial intelligence technologies are rapidly developing in recent years. The key techniques such as convolutional network, long short-term memory, and reservoir computing are based on the integration of neuroscience knowledge and machine learning methods. They have achieved performance exceeding human in applications such as image recognition, speech recognition, and games. However, their success relies on more than millions of labeled data and many applications do not have such big data. It is an open problem how to achieve good performance with a small number of data like a human can do. Development of advanced machine learning technology that is learned from human learning mechanism is studied all over the world.
On the other hand, in the field of neuroscience, the importance of machine learning methods is increasing as acquired data increases and becomes complicated. By letting computers learn patterns hidden in data using example data and mathematical models, the machine learning methods provide a means to process and interpret the vast amount of data that cannot be handled manually. Various applications such as automation of electron microscope connectomics, development of the brain-machine interface, and development of diagnostic biomarker are progressing and the utilization of machine learning technologies will be further increased in the future.
Our research group has been studying machine learning methods to understand the brain for more than 15 years. In particular, we have proposed a method that utilizes sparse Bayesian learning called the automatic relevance determination method (proposed by D.Mackay,1992) as one of the effective solutions for a small sample problem. We showed its effectiveness with the brain-machine interface, brain information decoding, biomarker development, and current source estimation. Although the Sparse Bayesian learning became an old method, there are still a lot of applications this method works effectively. In this presentation, we outline the impact that Sparse Bayes learning has on human imaging research and explain the importance of developing the new machine learning method. | | 7月25日（木）14:50～15:12 第6会場（朱鷺メッセ 2F 201A）
1S06a-2
弱教師付き分類
Takashi Ishida（石田 隆）
東京大院新領域創成科学複雑理工
Classification problems have been studied for decades by machine learning researchers, but in order for the technology to be successful, massive data with high-quality labels is usually needed. However, preparing such datasets can be unrealistic in many practical domains. In this talk, we will introduce two classification techniques from weak supervision: binary classification from positive-confidence data and multi-class classification from complementary labels.
In binary classification, the usual requirement is to have both positive and negative training data to learn a classifier. However in real world machine learning, we are often faced with problems where we only have positive data and no negative data is available, e.g., in purchase prediction, we can easily collect customer data from our own company (positive data), but not from rival companies (negative data). Previous classification methods could not cope with such situations, but we have achieved to learn with only positive data, as long as we have a confidence score for our positive data, constructed from information such as buying intention. With our new method, we can learn a classifier only from positive data equipped with confidence.
In multi-class classification, we consider learning from another type of weak but natural supervision called complementary-label learning, where the label only specifies one of the classes that the pattern does not belong to. In contrast to the ordinary case where the true class is given to each data (which often needs to be chosen out of many candidate classes precisely), collecting these complementary labels is obviously much easier and less costly. With our new method, we show we can learn a classifier from complementary labels. | | 7月25日（木）15:12～15:34 第6会場（朱鷺メッセ 2F 201A）
1S06a-3
脳の仕組みに基づいたDeep Neural Networkの表現学習法
Takashi Shinozaki（篠崎 隆志）1,2
1情報通信研
2大阪大
Convolutional neural networks (CNNs) and back propagation (BP) learning are the most powerful combination of recent machine learning methods. However, it requires a huge amount of labeled data for the training, and its requirement sometimes become a barrier for real world applications. On the other hand, biological neural networks work with very limited labeled data by utilizing unlabeled data more effectively. To realize the biological level effectiveness, exploration of more biologically plausible learning method than BP is required. One of the promised candidates is competitive learning which is an unsupervised leaning method used in neocognitron (Fukushima, 1980) and self organizing map (SOM) (Kohonen, 1980). In our previous study, we combined unsupervised competitive learning with supervised BP learning, and enabled semi-supervised learning for conventional CNNs (Shinozaki, 2017). However, its competitive leaning is only effective in the first convolutional layer, because the sparse dynamics of ReLU, an activation function of conventional CNNs, is strongly depended on bias modulation by BP learning. In this study, we introduce winners share all (WSA) mechanism into CNNs as an activation function. Since the dynamics of WSA is simple thresholding which does not require the bias modulation as BP learning, it enables competitive learning among multi-layers. We verified the method by a image discrimination task with CIFAR-10 dataset. As the result, the pre-training by competitive learning can obtain high-level features from unlabeled data, and drastically accelerated the learning speed especially in the early stage of the fine-tuning by BP learning. Since the representation learning by competitive learning is much more powerful than that by BP learning, the proposed method could be useful for applying various types of limited labeled data: motion data, time coursed of sensor data, time series, medical data, etc. Despite this study used just the simplest competitive learning, SOM-like competitive learning could be also possible to apply. It interpolates the filter space in CNNs, and may enable more continual and detailed representation than conventional CNNs. Further studies are required. | | 7月25日（木）15:34～15:56 第6会場（朱鷺メッセ 2F 201A）
1S06a-4
Reservoir Computingにおけるカオスの役割
Kohei Nakajima（中嶋 浩平）
東京大院情報理工知能機械情報
Reservoir computing (RC), which originated from the pioneering study on echo state network and liquid state machine in the early 2000s, has been proposed as a framework to train recurrent neural networks. In this framework, the low-dimensional input is projected to high-dimensional dynamical systems, which are typically referred to as a reservoir. If the dynamics of the reservoir involve adequate nonlinearity and memory, emulating nonlinear dynamical systems only requires adding a linear, static readout from the high-dimensional state space of the reservoir. Because of its generic nature, RC is not limited to neural networks, but any high-dimensional dynamical system can be used as a reservoir if it has the appropriate properties. Accordingly, the framework is now attracting attention to broad range of research field, not only in the neuroscience and machine learning community but also from mathematics, physics, chemistry, nanotechnology, spintronics, and soft robotics. It is important to explore and extend the framework to cope with this increasing diversity of the community. In this presentation, we discuss the role of chaos in the reservoir computing framework. Chaos is usually considered as a drawback for computational systems because it avoids reproducible responses to the same input in general. We show, however, through a number of experiments that chaos can be harnessed and positively exploited to increase the expressive power of computational system if we deal with it appropriately. | | 7月25日（木）15:56～16:18 第6会場（朱鷺メッセ 2F 201A）
1S06a-5
Topological complexity in the brain: Fragility, volatility, and a hierarchy of timescales
Leonardo Gollo（Gollo Leonardo）
Systems Neuroscience Group, QIMR Berghofer Medical Research Institute, Brisbane, Australia
The next era of connectomics will be driven by studies that consider both spatial and temporal components of brain organization simultaneously--elucidating their interactions and dependencies. Utilizing this guiding principle, we explored the origin of a hierarchy of timescales in the brain. This is a fundamental link between brain structure and dynamics that exhibits patterns of slow fluctuations within the topological core of hub regions, and fast fluctuations at the periphery. We find that a constellation of densely interconnected regions plays a central role in promoting a stable dynamical core. This core is crucial for integration of activity from specialized and segregated areas, and it mostly operates in slow timescales. These slow timescales are well matched to the regulation of internal visceral states, corresponding to the somatic correlates of mood and anxiety. In contrast, feeder regions surrounding this stable core show unstable and rapidly fluctuating dynamics, which is crucial for fast perceptual processes. Another implication of this relationship between structure and dynamics is revealed by studying the changes in brain dynamics following local stimulation. We find that the same stimulation protocol may cause opposite effects in functional connectivity if the target region is at the core or at the periphery of the network. These contrary effects are part of a continuous tuning curve that represents how different brain regions respond to stimulation depending on their position in the hierarchy. We also study the fragility and volatility of hub regions to random variants of the human connectome that introduce subtle perturbations to network topology while preserving the geometrical embedding and the wiring cost of the brain. If the variation in structure is permitted to accumulate, strong peripheral connections progressively connect to central nodes, and hubs shift toward the geometric center of the brain. Moreover, the fragility of hubs to disconnections shows a significant association with regions that exhibit accelerated thinning of gray-matter volume in schizophrenia. Finally, we discuss potential applications of brain hierarchy and its integrating core to artificial intelligence and beyond. | | 7月25日（木）16:18～16:40 第6会場（朱鷺メッセ 2F 201A）
1S06a-6
作用素論的方法に基づくデータ駆動による非線形力学系の解析
Yoshinobu Kawahara（河原 吉伸）1,2
1九州大学 マス・フォア・インダストリ研究所
2理化学研究所 革新知能統合研究センター
Data-driven modeling of complex systems have received much attention over the recent years, largely due to the availability of large datasets. In this context, operator-theoretic analysis of nonlinear dynamical systems has been actively discussed in applied mathematics and various scientific fields including neuroscience. This is because it can provide physical interpretations of the dynamics based on deep theoretical backgrounds and is endowed with prominent estimation methods such as dynamic mode decomposition (DMD). DMD is a numerical method for estimating spectra of Koopman operators and has been attracting attention as a way of obtaining global modal descriptions of nonlinear dynamical systems from data without requiring explicit prior knowledge.
In this talk, I first overview the recent advances on this research topic, especially focusing on spectral analysis of dynamical systems with Koopman operators and DMD. Then, I describe several recently-proposed DMD algorithms using machine learning techniques. For example, I describe a variant of DMD using reproducing kernels, which could avoid the issue in DMD when selecting nonlinear observables to estimate accurately spectra of Koopman operators from finite data and also can provide a more flexible scheme applicable to structured data sequences such as graph sequences. In the talk, I occasionally show some applications of these method to several real-world data, such as ones from collective motions. Finally, I also discuss several open problems and perspectives related to the context. | |
|
|
