KAIST Bio-IT Healthcare Initiative International Workshop Series : Neural Computation
2:00pm-5:30pm, NOV 2 Wednesday, 2016
Dream Hall (1st floor), CHUNG Moon Soul building (E16), KAIST
The KAIST Bio-IT Healthcare Initiative International Workshop, Neural Computation, is a half-day event to discuss various approaches in theoretical neuroscience and machine learning to understand brain functions.
Topics we will discuss include:
- 3D structures of neurons
- functional organization of neural networks
- cognitive control processes
- dimensionality of neural activities
- distributed information encoding of neurons
*Hosts: KAIST Bio-IT Healthcare Initiative, Department of Bio and Brain Engineering, KAIST
*Organizer: Sang Wan Lee (inquiries: Hana Park, email@example.com)
LIST OF SPEAKERS
Se-Bum Paik (KAIST, South Korea)
He is an assistant professor of the department of bio and brain engineering at KAIST. He received his PhD in physics at University of California, Berkeley. He was a postdoctoral associate at UCLA. He was awarded POSCO TJ Park Science Fellowship. His research interest is computational neuroscience, focusing on the dynamics of the functional organization of the neural networks in visual system.
Sang Wan Lee (KAIST, South Korea)
He is an assistant professor of the department of bio and brain engineering at KAIST. He received his PhD in electrical engineering and computer science at KAIST. He was a postdoctoral associate at MIT, and a Della Martin postdoctoral scholar at Caltech. His research interests include brain-inspired artificial intelligence and computational neuroscience.
Jinseop Kim (Korea Brain Research Institute, South Korea)
He is a senior research members at the Korea Brain Research Institute. He received his PhD in physics at Seoul National University. He was a postdoctoral researcher at MIT and Princeton. His research interests include computational neuroscience, neuroanatomy, and connectomics. He was a leading scientist of the brain-mapping game EyeWire to precisely map the cells of a mouse retina and shed a light on how those cells detect motion (Nature 2014).
Mattia Rigotti (IBM T.J. Watson Research Center, USA)
He is a research staff member in the Theory and Computational Physics Group at the IBM T. J. Watson Research Center. He received his PhD in neuroscience and M. Phil. in neurobiology and behavior at Columbia University. He was an associate research fellow at the Center for Theoretical Neuroscience and The Italian Academy for Advanced Studies at Columbia University. His research interests include neuromorphic engineering, computational neuroscience, neural networks and machine learning. He applies machine learning theory on electrophysiology data to understand neural coding schemes underlying rule-based behavior (Nature 2013, Nature 2015, and Neuron 2015).
14:00-14:10 [10min] Opening remarks (Sang Wan Lee)
14:10-14:40 [30min] Se-Bum Paik (KAIST, South Korea)
Functional structure of cortical neural networks for visual information processing
In the primary visual cortex of higher mammals, neurons are often spatially organized by their stimulus preference, forming various types of functional maps. Explaining the origin and role of these functional structures is crucial to understanding how various components of information are being processed in the brain. Our research over the past few years has focused on developing a new theoretical model of the initial development of functional structures in the primary visual cortex (V1). We built a unified model of developmental mechanism of various cortical functional maps, which not only successfully explains the observed geometrical correlations between different functional maps, but also accounts for the intriguing developmental mechanism of functional circuits in the brain for effective information processing. This theoretical model may significantly change our current view of the functional structure of the brain, and helps us better understand how we take theoretical and computational approaches to study the neural circuits for various brain functions.
14:40-15:10 [30min] Sang Wan Lee (KAIST, South Korea)
Markov decision process unfolds within the human prefrontal cortex
Recent advances in artificial intelligence (AI) research have paved the way for developing human-level intelligent systems. This begs the question of how the human brain handles a wide variety of decision making tasks, whereas conventional AI systems need to function in a task-specific way. This talk puts together ideas from AI and neuroscience to appreciate the human brain’s ability to achieve diversity of learning and inference. I will present evidence suggesting that the brain consists of multiple subsystems for distinctive types of learning and inference: model-based and model-free reinforcement learning, and incremental and one-shot causal inference. Based upon a combination of functional magnetic resonance imaging (fMRI) and computational modeling, I will then show evidence supporting the view that the human lateral prefrontal cortex (LPFC) influences the brain's various learning systems, placing it as the brain's “meta-controller”. These findings may help to explain how the LPFC might allocate control over behavior to brain’s subsystems in a way that mounts efficient operation under constraints on performance, time, and energy.
15:10-16:10 [60min] Jinseop Kim (Korea Brain Research Institute, South Korea)
Neural codes of visual perception: single cells and neural circuits in the retina
The neural computation of visual perception already begins in the retina. Retinal ganglion cells (RGCs) are the outputs of the retina, each type of which is a channel that distinct visual information is processed through. While classifying the types is crucial to understanding of their roles in vision, the catalog of types remains incomplete. In this talk I will present our recent effort to anatomically classify unbiased sample of 400 RGCs reconstructed from a stack of electron microscopic images of a small patch of mouse retina. In all, the classification resulted in more than 40 distinct types. The visual response properties of the types could be understood from their structural features and connectivity with other cells. This research showcases that functions of neural systems are encoded in their structures.
16:10-16:30 [30min] Coffee break
16:30-17:30 [60min] Mattia Rigotti (IBM T.J. Watson Research Center, USA)
High and low-dimensional neural responses for learning and implementing context-dependent behavior
Single-neuron activity in frontal cortex is very complex: in animals engaged in cognitive behavior, responses are reliably but idiosyncratically tuned to mixtures of multiple task-related aspects. We will see that such "mixed selectivity" at the level of individual cells is a signature of high-dimensionality at the level of the population activity. We will consider theoretical models and empirical evidence suggesting that such high-dimensionality confers impressive computational advantages to a neural circuit in terms of the richness of the set of implementable downstream responses. We will then examine recurrent network models that shape the mixed selectivity inherent in high-dimensional patterns to generate targeted low-dimensional representations that support abstraction, context-dependent behavior and generalization. The predictions from these models will then be compared with experimental results obtained in classical conditioning and navigation tasks.