Seminars in January and February, 1997
- 02/25/97 - Generalized Functions and Hybrid Control, Professor Bob Hermann, MIT, NASA Ames.
- 02/25/97 - Hierarchical, Logic and Hybrid Control (I): Hierarchical Control of Finite Machines, Peter E. Caines, McGill University and CIAR.
- 02/18/97 - Prediction, diagnosis and control in dense belief networks, Michael I. Jordan, MIT.
- 02/14/97 - MURI Visualization Seminar: Recent Developments in 3D Laser Scanning, Marc Rioux and Luc Cournoyer, National Research Council, Canada.
- 02/11/97 - CNN Technology: a new computational platform for dynamic spatiotemporal array computing, Tamas Roska, Analogical and Neural Computing Laboratory in Budapest (visiting UC Berkeley).
- 01/21/97 - Motion Analysis, Perceptual Organization and the EM algorithm, Yair Weiss, MIT
Seminar Abstracts
Tuesday, February 25th, 1997
5-6 pm *
306 Soda Hall
Generalized Functions and Hybrid Control
Professor Bob Hermann
MIT, NASA Ames
* this will be an informal talk taking place during the discussion period of Peter Caines' talk advertised earlier.
Tuesday, February 25th, 1997
4-5pm*
306 Soda Hall
Hierarchical, Logic and Hybrid Control: I
Hierarchical Control of Finite Machines
Peter E. Caines
McGill University and CIAR
* Discussion will follow from 5-6pm in 306 Soda. This is the first in a three-talk series by Professor Caines: the second and third talks will take place Wed. Feb 26th, 4-5pm, and Fri. Feb 28th, 4-5pm.
Abstract
1. Complexity and HierarchyThe great complexity of many natural and designed systems limits the ability of humans and machines to directly describe, regulate and manage them. Consequently, hierarchically structured information and control systems are frequently employed for these tasks. Furthermore, complex artificial systems are often hierarchically configured at the design stage so as to facilitate their control and supervision.
2. State Aggregation and High Level Dynamics
A notion of state aggregation with an associated controlled dynamics is introduced via the dynamical consistency (DC) relation between the sets of states in the cells (or blocks) of an arbitrary partition of the state space of a finite machine M. A DC relation between a partition member Xi and a member Xj corresponds to the existence, for every state x in Xi, of some (x dependent) control sequence which takes that state into Xj without excursions into other blocks. This formulation results in a definition of high level dynamics on the finite partition machine whose states correspond to the given partition elements. The DC relation is then similarly defined on any further partitions (i.e. aggregations) of the state set of the finite partition machine, and so on. In a management or operations research context such system representations could be called aggregated-enabled networks. We analyse what are termed the set of in-block controllable (IBC(M)) and between-block controllable (BBC(M)) partition machines. It is shown that any partition machine in IBC(M) is between-block controllable if and only if the original machine M is controllable.
3. The Lattice of Hierarchical Control Systems
The class of in-block controllable partition machines has a (non-distributive) lattice structure and this permits a natural construction and classification of all hierarchical control systems for a given machine M. This setting allows the feedback control of trajectories of M to be decomposed into a chain of control laws in a hierarchy of control systems.
4. Source-Target Systems
In the analysis and design of hierarchical control systems, one is often interested in the reachability of some set of target states from some initial set of states. The theory of hierarchical control described above has a direct generalization to such ST-systems and applications to buffer-machine production lines will be given.
5. Supervisory Control
Finally, the expression of this theory in terms of the supervisory control theory of Ramadge-Wonham will be outlined; this will include the construction of the dynamically consistent (DC) sub-languages of a given language and the associated controllable DC languages.
Tuesday, February 18th, 1997
4-5:30pm
306 Soda Hall
Prediction, diagnosis and control
in dense belief networks
Michael I. Jordan
MIT
Abstract:
Belief networks provide an elegant formalism for managing uncertainty that unifies much of the literature on stochastic modeling. For sparse belief networks (e.g., networks in the form of chains or trees, such as Kalman filters, hidden Markov models, and probabilistic decision trees), belief network algorithms are exact, efficient and practical. For dense belief networks, however, the exact algorithms are often (hopelessly) inefficient, and this fact has hindered the application of this richer class of models to real-life problems. I discuss variational methodology, which provides a general framework for approximate belief network inference. The variational methods I present are efficient; moreover, they tend to be more accurate for dense networks than for sparse networks. They can readily be combined with exact techniques to yield a class of algorithms that perform well for a wide variety of network architectures. I illustrate these ideas with examples of applications of dense belief networks to problems in prediction, diagnosis and control.
Friday, February 14, 4:00pm, 306 Soda Hall
Marc Rioux and Luc Cournoyer
National Research Council (Canada)
Recent Developments in 3D Laser Scanning
Abstract:
A presentation of the current research activities in 3D digitizing will be made. Emphasis will be on the digitizing of environment for applications such as documentation and simulation. We will show experimental results related to industrial, museum and space interests. Fundamentals of laser properties for 3D digitizing will also be discussed, in the context of future developments.After the talk we will adjourn to room 544 Soda for some real-time demos.
Tuesday, February 11th, 1997
4-5:30pm
306 Soda Hall
CNN Technology: a new computational platform for dynamic spatiotemporal array computing
Tamas Roska
Analogical and Neural Computing Laboratory in Budapest
(visiting UC Berkeley)
Abstract:
The CNN Technology is based on a new computing paradigm and architecture: the CNN Universal Machine. Its elementary instruction is defined on an array computer where each processor is a dynamical system with continuous signals and containing some additional logic. This unconventional architecture is very close to the anatomy and physiology of many parts of the brain, hence, it is called neuromorphic. It is also close to many dynamic models of physics, chemistry, population dynamics, etc. In the Cellular Nonlinear/neural Network (CNN), each of the processors placed on a two or three dimensional grid are dynamic cells, and they are interacting mainly with their local neighbor cells/processors.Unlike the cellular automaton or systolic array, here the signals and the time is continuous. The CNN paradigm was invented by Prof. Leon Chua and his graduate student L. Yang in 1987/88. If we extend the CNN cell by analog and logic memory as well as additional interface units , and the whole array is driven by a special Global Analogic Program Unit (GAPU), we have all the key elements of the CNN Universal Machine (CNN-UM) architecture. Since the invention of the CNN-UM (Roska-Chua, 1992), several VLSI chips were designed and recent experiments show close to trillion operations per second computing power on 1 cm2 silicon area. In this lecture, we will brifly review the three main aspects of recent works, with special reference to works made in the cooperating Laboratory in Budapest:Some case studies will be shown to explain why this new computing platform is so efficient in solving complex image processing problems like mathematical morphology, image analysis, or propagating spatiotemporal dynamics on images. Measured figures, including equivalent DSP computing times, will be presented.
- chips, chip sets and engine boards with accompanying high level language and operating systems
- key application areas and examples
- biological relevance.
Tuesday, January 21, 1997
4-5:30pm
306 Soda Hall
Motion Analysis, Perceptual Organization and the EM algorithm
Yair Weiss
MIT
Abstract: Estimating motion in scenes containing multiple, complex motions remains a difficult problem in computer vision yet is solved effortlessly by humans. I will briefly discuss experiments we have conducted to understand the representation underlying this remarkable performance in human vision. The results implicate an early perceptual organization process as crucial to human motion perception. I will present a new Expectation-Maximization (EM) algorithm that analyzes motion based on these principles. Distinguishing features of our algorithm include automatic estimation of the number of models, the use of spatial coherence to aid the segmentation, and the estimation of a nonparametric smooth flow field for each model. The performance of the algorithm will be illustrated with synthetic and real image sequences.Joint work with Ted Adelson.
[Top|Home]