IJCNN 2011 - Plenary Talks

IJCNN 2011 will feature invited plenary talks by the following speakers:

Brains, Machines and Buildings
Michael Arbib	University of Southern California

Challenges for Computational Vision: From Random Dots to the Wagon Wheel Illusion
Leon Glass	McGill University

Cognitive Computing: Neuroscience, Supercomputing, Nanotechnology
Dharmendra Modha	IBM Almaden Research Center

Deep Learning and Unsupervised Feature Learning
Andrew Ng	Stanford University

Learning Motor Skills in Humans and Humanoids
Stefan Schaal	University of Southern California

Neural Network ReNNaissance
Juergen Schmidhuber	IDSIA, Switzerland

David Rumelhart Memorial Session

On Wednesday, August 3, IJCNN 2011 will feature a special plenary session honoring the life and work of the Late David Rumelhart, one of the true pioneers in the field of neural networks and a major contributor to the field's modern renaissance.

Speaker: Michael I. Jordan (University of California, Berkeley)

Featured Plenary Session: The Emergence of Mind

On Thursday, August 4, IJCNN 2011 will feature a special plenary session focusing on how higher mental functions such as perception, cognition and consciousness emerge from the neural substrate of the brain.

Speakers: Stephen Grossberg (Boston University)
Walter J. Freeman (University of California, Berkeley)
Bernard J. Baars (The Neuroscience Institute)

Abstracts of Plenary Talks

Learning Motor Skills in Humans and Humanoids
8:00 AM -9:00 AM Monday, August 1
	Stefan Schaal University of Southern California Los Angeles, California, USA


Neural Network ReNNaissance
1:50 PM - 2:50 PM Monday, August 1
	Juergen Schmidhuber Swiss Institute for Artificial Intelligence (IDSIA) Lugano, Switzerland
Our fast deep / recurrent neural nets recently achieved numerous 1st ranks in many pattern recognition competitions and benchmarks, without any unsupervised pre-training, sometimes (but not always) profiting from weight sharing & convolution, contrast enhancement, max-pooling, and sparse network connectivity. GPUs speed up learning by a factor of up to 50, thus contributing to the ongoing second Neural Network ReNNaissance. The future, however, will belong to active systems that learn to sequentially shift attention towards informative inputs, not only solving externally posed tasks, but also their own self-generated tasks designed to improve their understanding of the world according to our Formal Theory of Fun and Creativity, which requires two interacting modules: (1) an adaptive (possibly neural) predictor or compressor or model of the growing data history as the agent is interacting with its environment, and (2) a (possibly neural) reinforcement learner. The learning progress of (1) is the FUN or intrinsic reward of (2). That is, (2) is motivated to invent skills leading to interesting or surprising novel patterns that (1) does not yet know but can easily learn (until they become boring). We discuss how this principle explains science & art & music & humor.

Brains, Machines and Buildings
8:00 AM -9:00 AM Tuesday, August 2
	Michael Arbib University of Southern California Los Angeles, California, USA
Michael Arbib is a University Professor, Professor of Neuroscience and the Fletcher Jones Professor of Computer Science at the University of Southern California. The scope of his career was defined in the title of his first book Brains, Machines, and Mathematics. This talk will turn from mathematics to buildings, seeking to understand the implications for architecture of his work on computational modeling of brain mechanisms and mirror neurons, relating vision to action, emotion and language.

The talk will introduce Neuromorphic Architecture, exploring ways to incorporate "brains" into buildings, developing the view that future buildings are to be constructed as perceiving, acting and adapting entities. The discussion is grounded in an exposition of Ada - the intelligent space, a pavilion visited by over 550,000 guests at the Swiss National Exhibition of 2002. She had a "brain" based (in part) on neural networks, had "emotions" and "wanted" to play with her visitors. Dramatic new developments will emerge as we explore the lessons from neuroscience on how the brain supports an animal's interactions with its physical and social world to develop brain operating principles that lead to new algorithms for a neuromorphic architecture which supports the "social interaction" of rooms with people and other rooms to constantly adapt buildings to the needs of their inhabitants and enhance interactions between the people who use them and their environment.

Cognitive Computing: Neuroscience, Supercomputing, Nanotechnology
1:50 PM - 2:50 PM Tuesday, August 2
	Dharmendra Modha Cognitive Computing Group IBM Almaden Research Center
The ultimate goal of the DARPA SyNAPSE project is to build brain-like cognitive computing chips that scale to human cortex by moving beyond the von Neumann architecture and become the brains behind IBM's Smarter Planet vision. The project leverages neuroscience, supercomputing, and nanotechnology and is currently a collaboration of four universities (Cornell, Columbia, Wisconsin-Madison, UC Merced) and five IBM sites (Almaden, Austin, East Fishkill, India, Yorktown).

Challenges for Computational Vision: From Random Dots to the Wagon Wheel Illusion
8:00 AM -9:00 AM Wednesday, August 3
	Leon Glass McGill University Montreal, CANADA
Even understanding the way we perceive very simple images presents a major challenge for both neurophysiologists and computer scientists. In this talk I will discuss two visual effects. In one random dots are superimposed on themselves following a linear transformation (1,2). In the second, a rotating disk with radial spokes is viewed under stroboscopic illumination, where the frequency and duration of the stroboscopic flash are varied (3,4). Though these phenomena are very different, in both correlation plays a major role in defining the structure of the image. In this talk, I will give demonstrations of these phenomena and discuss related experimental and theoretical work by ourselves and others. In particular, I focus on recent analysis that uses the theory of forced nonlinear oscillations to predict the percept of rotating disks during stroboscopic illumination over a wide range of disk rotation speeds and strobe frequencies (4). Finally, I suggest that the anatomical structure of the human visual system plays a major role in enabling the amazingly rapid and accurate computation of spatial and time dependent correlation functions carried out by the visual system. 1. L. Glass. Moire effect from random dots. Nature 223, 578580 (1969). 2. L. Glass, R. Perez. Perception of random dot interference patterns. Nature 246, 360-362 (1973). 3. R.M. Shymko, L. Glass. Negative images in stroboscopy. Optical Engineering 14, 506-507 (1975). 4. P. Martineau, M. Aguilar, L. Glass. Predicting perception of the wagon wheel illusion. Physical Review Letters, 103:2 (2009).


Deep Learning and Unsupervised Feature Learning
1:50 PM - 2:50 PM Wednesday, August 3
	Andrew Ng Stanford University Palo Alto, California, USA
Machine learning often works very well, but can be a lot of work to apply because it requires spending a long time engineering the input representation (or "features") for each specific problem. This is true for machine learning applications in vision, audio, text/NLP and other problems. To address this, researchers have recently developed "unsupervised feature learning" and "deep learning" algorithms that can automatically learn feature representations from unlabeled data, thus bypassing much of this time-consuming engineering. Many of these algorithms are developed using simple simulations of cortical (brain) computations, and build on such ideas as sparse coding, self-taught learning, and deep belief networks. By doing so, they exploit large amounts of unlabeled data (which is cheap and easy to obtain) to learn a good feature representation. These methods have also surpassed the previous state-of-the-art on a number of problems in vision, audio, and text. In this talk, I describe some of the key ideas behind unsupervised feature learning and deep learning, and present a few algorithms.

David Rumelhart Memorial Session Wednesday, August 3, 2011 6:15 PM - 7:30 PM

Learning Natural Language Semantics
	Michael I Jordan University of California Berkeley Berkeley, California, USA
What is the total population of the ten largest capitals in the US? Answering free-form questions such as this requires modeling the deep semantics of language. But is it possible for a learner to acquire these deep semantics from only surface-level supervision, e.g., question/answer pairs? We answer affirmatively, by developing a new tree-based semantic representation with favorable linguistic and computational properties, along with an algorithm that induces this hidden representation. Using our approach, we obtain significantly higher accuracy on the task of question answering compared to existing state-of-the-art methods, despite using less supervision (Joint work with Percy Liang and Dan Klein).


Featured Session: The Emergence of Mind Thursday, August 4, 2011 8:00 AM - 9:30 AM

Conscious experience and the observing ego: A dynamic global workspace hypothesis
	Bernard Baars The Neurosciences Institute San Diego, CA
Global Workspace (GW) theory aims to explain the differences between conscious and unconscious brain activities, such as the striking limited capacity of conscious contents vs. the vast capacities of unconscious memory storage, automatic skills, implicit knowledge and subcortical computations. Like any theory, this one must explain its own observable indices. The most widely used behavioral index of consciousness is accurate reportability of brain events attributed to a stable executive perceiver. A global workspace is a domain of signal integration and propagation in a set of parallel-interacting processors that combine to resolve ambiguous or unpredictable signals. In nature, animals encounter such signals very often, and the ability to resolve them in a timely way can be a matter of life or death. The contents of consciousness are supported by reentrant cortical and thalamic signaling, regulated by waking state modulation. Anatomically the cortico-thalamic (C-T) system is by far the largest parallel-interactive structure in the brain. Major C-T hubs are well suited for global signal integration and dissemination. Dynamic signals in the C-T core are interpreted in the egocentric/allocentric maps of the parietal and medial temporal lobes. MTL-parietal maps enable a stable egocentric platform for visual conscious input and voluntary control (Milner & Goodale, 2004). In the case of conscious vision, a dynamic Global Workspace (dGW) integrates occipital, temporal and parietal oscillations into a single gestalt, emerging as a preconscious P3 waveform in the visual event-related potential (ERP). After achieving equilibrium, the P3 propagates a burst of phase-locked gamma/theta oscillations to multiple receiving populations in frontoparietal regions, appearing in the ERP as a conscious vs. unconscious difference wave between 400-600 ms (Del Cul et al, 2004; Revonsuo et al 2006). Receiving populations resonate to match the global signal. When a widespread match is achieved the global event fades from consciousness. A second integrative wave of dGW activity combines in the frontal lobes for voluntary control of the vocal tract and cranioskeletal muscles. This second wave enables voluntary actions with respect to conscious visual events with very high accuracy. Recurrence of the posterior dGW may be triggered by a control cycle using a frontal-basal ganglia-thalamic loop. Conscious experiences also evoke longterm coding of novel gestalts by way of MTL-neocortical theta oscillations.

The Making of Mind through the Action-Perception Cycle
	Walter J. Freeman Department of Molecular & Cell Biology University of California at Berkeley Berkeley, California, USA
Phylogenetically speaking, mind emerged through the need for food. Primitive vertebrates flourished using their abilities to experience the need, predict odorant chemicals in the environment that signified acceptable foods, search through the environment with tactile, visual and auditory guidance, categorize multiple modality-specific sensory inputs, synthesize a multisensory gestalt, create a cognitive map with which to label each sample giving the time and place of each acquisition, and store and recall significant memories for personal use in future searches. Vertebrates invented cognitive codes with which to make, store and recall their memories as categories of what foods to look for, where, and when. The code for memories is a landscape of chaotic attractors in each cortex. The sensory code is microscopic. Samples are taken by search (sniffing, looking, whisking), conveyed by spatial patterns of action potentials, and mapped in each cortex by topographically organized axons. The perceptual code is macroscopic. Each cortex creates a burst of amplitude-modulated gamma oscillation by a phase transition that resembles the condensation of a gas to a liquid. Every cortical neuron participates in every percept by time multiplexing in a feature vector. The vectors combine in the entorhinal cortex in a multisensory gamma burst. Recursion of the high-dimensional vector through the hippocampal loop incorporates time and place. The entorhinal cortex broadcasts each new vector in a global burst of beta oscillation, as demonstrated in 64-channel EEGs (Ruiz et al., 2011). The neural-mental search image closes the action-perception cycle and is integrated into the personal memory bank. For more see http://soma.berkeley.edu Ruiz Y, Pockett S, Freeman WJ, Gonzales E, Li Guang (2010) A method to study global spatial patterns related to sensory perception in scalp EEG. J Neuroscience Methods 191: 110-118. Freeman WJ, Kozma R [2010] Freeman's mass action. Scholarpedia, 5(1): 8040. http://www.scholarpedia.org/article/Freeman%27s_mass_action

How can children (and robots) learn to follow gaze, share joint attention, imitate their teachers, and use tools during social interactions?
	Stephen Grossberg Center for Adaptive Systems Boston University Boston, Massachusetts, USA

The emergence of a mind depends upon many factors, notably the ability to learn from teachers who see the world through a different perspective. How can an infant, or robot, incrementally learn through visual experience to imitate actions of adult teachers, despite the fact that the infant and adult view one another and the world from different perspectives? To accomplish this, an infant needs to learn how to share joint attention with adult teachers and to follow their gaze towards valued goal objects. The infant also needs to be capable of view-invariant object learning and recognition whereby it can carry out goal-directed behaviors, such as the use of tools, using different object views than the ones that its teachers use. Such capabilities are often attributed to "mirror neurons". This attribution does not, however, explain the brain processes whereby these competences arise. The CRIB (Circular Reactions for Imitative Behavior) model suggests how a child's brain may achieve these goals through inter-personal circular reactions. Inter-personal circular reactions generalize the intra-personal circular reactions of Piaget, which clarify how infants learn from their own babbled arm movements and reactive eye movements how to carry out volitional reaches, with or without tools, towards valued goal objects. The CRIB model proposes how intra-personal circular reactions create a foundation for inter-personal circular reactions when infants and other learners interact with external teachers in space. Both types of circular reactions involve learned coordinate transformations between body-centered arm movement commands and retinotopic visual feedback, and coordination of processes within and between the What and Where cortical processing streams. Specific breakdowns of model processes generate formal symptoms similar to clinical symptoms of autism. Supported in part by the DARPA SyNAPSE program and the NSF Science of Learning program. References (see http://.cns.bu.edu/~steve) Cao, Y., Grossberg, S., & Markowitz, J. (2011). How does the brain rapidly learn and reorganize view- and positionally-invariant object representations in inferior temporal cortex? Neural Networks, in press. Fazl, A., Grossberg, S., and Mingolla, E. (2009). View-invariant object category learning, recognition, and search: How spatial and object attention are coordinated using surface-based attentional shrouds. Cognitive Psychology, 58, 1-48. Grossberg, S. and Seidman, D. (2006). Neural dynamics of autistic behaviors: Cognitive, emotional, and timing substrates. Psychological Review, 113, 483-525. Grossberg, S., and Vladusich, T. (2010). How do children learn to follow gaze, share joint attention, imitate their teachers, and use tools during social interactions? Neural Networks, 23, 940-965. Grossberg, S., Markowitz, J., and Cao, Y. (2011). On the road to invariant recognition: Explaining tradeoff and morph properties of cells in inferotemporal cortex using multiple-scale task-sensitive attentive learning. Neural Networks, in press. Huang, T.-R., and Grossberg, S. (2010). Cortical dynamics of contextually cued attentive visual learning and search: Spatial and object evidence accumulation. Psychological Review, 117, 1080-1112. ;