Numero 35

fredag 15 november 1996, årgång 26

Veckoblad om forskning, undervisning och administration m.m.
NADA - Institutionen för numerisk analys och datalogi, KTH


Aktuellt kalendarium (inte kalendariet som hör specifikt till denna Numeroutgåva).

Teoretisk datalogi-seminarium: Parallel repetition

Alex Russell, NADA, håller seminarium i teoretisk datalogi under ovanstående rubrik på torsdag (21 nov.) kl. 15.15-16.15 i seminarierum 1537 (Lindstedtsv. 3, plan 5).


The notion of a two-prover proof system, introduced by Ben-Or, et. al. in 1988, is a cornerstone in the remarkably successful theory of interactive and probabilistically checkable proofs. Although this theory provides startling new models for many known complexity classes, most celebrated is it's unexpected connection to approximation algorithms which has provided the first general framework for proving lower bounds on approximability.

The behavior of two-prover proof systems under "parallel repetition", a natural method for amplifying such systems, remained poorly understood until 1995 when Ran Raz fairly closed the area.

This talk has two goals. First, we shall introduce two-prover proof systems and give some indication of their connection with approximation lower bounds. Second, we shall discuss the ramifications of parallel repetition for these applications and survey Raz's proof the parallel repetition theorem.

Johan Håstad

SANS-seminarium: connectionist modelling of generalization in conditioning

Christian Balkenius, Dept. of Cognitive Science, Lunds Universitet, håller SANS-seminarium under ovanstående rubrik på tisdag (19 nov.) kl. 13.15-14.30 i seminarierum 4523 (Lindstedtsv. 5, plan 5).


While it was once believed that conditioning in some form could account for all learning it is now widely recognized that this is not the case. Many types of learning does not fall into the simple pairing of a neutral stimulus with a subsequent significant event. There are however a number of learning situations which can be adequately described as instances of conditioning, and any complete model of learning must include mechanisms which can handle such learning. However, the simplicity of the experimental procedure does not mean that the processes and representations required for conditioning are simple too.

I will describe my research on how a representation at multiple scales can be used for generalization in conditioning. The use of such representations is an efficient way to code similarities and differences within a stimulus dimension, and allows a learning system to generalize easily between various situations. A neural network based model of classical and instrumental conditioning is presented and its ability to generalize using multi-scale representations is demonstrated in a number of simulations.

Mikael Djurfeldt, Anders Lansner

I korthet

Disputation: Techniques for Efficient Simulation of Computer Architectures, Luis J. Barriga C., IT, KTH. Torsdagen den 5 december kl. 14.00 i Kollegiesalen, administrationshuset (Valhallavägen 79). Numeropärmen 96.141

Disputation: Data-Parallel Implementation of Prolog, Johan Bevemyr. Onsdagen den 20 november kl. 14.00 i rum 311, hus 1, Polacksbacken, Computing Science Department, Uppsala Universitet. Numeropärmen 96.142

PDC-seminarium: Transverse velocity increments in turbulent flow using the RELIEF technique

Uriel Frisch, Observatoire de Nice, Frankrike, håller PDC-seminarium under ovanstående rubrik på tisdag (19 nov.) kl. 15.15 i seminarierum 1537 (Lindstedtsv. 3, plan 5).


Nonintrusive measurements of the streamwise velocity of turbulent: round jets in air are performed by recording short-time displacements and distorsions of very thin tagging lines written spanwise into the flow. The lines are written by Raman Exciting oxygen molecules and are interrogated by Laser Induced Fluorescence (RELIEF). This gives access to the spatial structure of transverse velocity increments without recourse to the Taylor hypothesis. The resolution is $25mu$, comparable to the Kolmogorov scale $eta$ for the experiments performed (with $R_lambda pprox 300--500$).

The technique is validated by comparison with results obtained from other techniques for longitudinal or transverse structure functions up to order eight. The agreement is consistent with the estimated errors -- about 5 % on exponents determined by Extended-Self-Similarity -- and indicates significant departures from Kolmogorov 1941 scaling.

Probability distribution functions of velocity increments $Delta u$ over separations down to $1.5 eta$ are reported for the first time. Violent events with $Delta u$ comparable to the rms turbulent velocity fluctuation, are found to take place with statistically significant probabilities. The shapes of the corresponding lines suggest the effect of intense slender vortex filaments.

This talk is based on a joint paper by Noullez, Wallace, Lempert, Miles and Frisch, submitted to Journal of Fluid Mechanics.

Erik Aurell

^ Upp till Numeros hemsida.

Senast ändrad 14 nov. 1996 <>