Past Special Colloquia IRCS/Computational Neuroscience Speaker Series
IRCS is collaborating with Penn's Computational Neuroscience group to present a series of lectures during the 2009-2010 academic year.IRCS/SILC Joint Talks
IRCS is partnering with SILC, the Spatial Intelligence Learning Center at Temple University, to host lectures during the academic year. Visitors for this lecture series will present their research in Spatial Cognition.
September 26, 2008:
Veronique Bohbot
Faculty of Medicine, McGill University
Modulating grey matter in the hippocampus and striatum in mice and humans
October 3, 2008:
David Waller
Department of Psychology, Miami University of Ohio
Two systems of spatial knowledge and their application to virtual environment interfaces
January 30, 2009:
A. David Redish
Department of Neuroscience, University of Minnesota
Deliberative and automatic spatial decision making in the rat
Anna Nesterova, University of Pennsylvania
Tuesday, May 27, 2008
Orientation in a Crowded Environment: Can King Penguin Chicks Find Their Crèches?
In colonial birds, the problem of short-range orientation is especially challenging. The presence of many conspecifics can obstruct local cues that are helpful for accurate orientation. This problem is especially apparent in the case of king penguins (Aptenodytes patagonicus), that need to find their small residence areas in a large and crowded colony. This study investigated short-range orientation in the one-year old king penguin chicks. The three objectives were to determine 1) if chicks can orient towards the colony; 2) if they can return to their specific place within a colony (crèche); and 3) whether they rely on visual or non-visual cues for orientation when displaced to a novel location away from the colony. To address these questions, a circular arena was constructed 100 m away from the colony edge, at a location where the colony was not visible. Sixty four chicks were released in the arena under different conditions. Animals were tested during the day (visual cues present) and night (visual cues absent). In addition, the availability of visual cues was further manipulated during the day. A low arena barrier provided animals with a full view of the landscape. A high arena barrier allowed animals to see only a general outline of the landscape, but obscured details. Chicks were captured at their crèches and released in the arena. During 15 min period we noted the amount of time animals spent in each half of the arena (half closer to the colony and the opposite half). Then the barrier was lowered completely and animals were allowed to home. During day trials with a high or low barrier, more chicks preferred the half of the arena that was closer to the colony. Such preference was less pronounced during night trials. However, at night birds spent more time on ‘the colony half’ of the arena if the wind blew from the direction of the colony. When animals were allowed to leave the arena, 98% of chicks homed during the day, but only 62% of chicks returned to their crèche at night. Chicks that homed at night took longer to find their crèche. The experiments revealed that one-year old king penguin chicks could find their crèche from a novel location. In addition, visual cues appeared important to a chicks’ ability to home. However, they were not essential, and some animals were able to home in the absence of visual cues. When visual cues were not present, animals apparently attended to other information carried by the wind, for example, auditory or olfactory cues.
Mark Johnson, Brown University
November 5, 2003
A TAG Noisy Channel Model of Speech Repairs
Speech errors are one of the many difficulties encountered when analysing naturally
occuring speech. In earlier work Charniak and Johnson (2001) used a word-by-word
classifier to detect and correct speech repairs before parsing. While this provided
good performance, it had a number of practical and theoretical deficiencies. This talk
describes on-going work on addressing these problems using a noisy channel model of
speech repairs. Since speech repairs involve "rough copy" dependencies, I suggest that
a stochastic TAG transducer is a natural way to model the noisy channel of speech
repairs.
Jont Allen, University of Illinois, Urbana Champaign
May 5, 2003
From Lord Rayleigh to Shannon: How do humans decode speech?In 1908 Lord Rayleigh reported on his speech perception studies using the ``acousticon'' (a commercial sound system produced in 1905), demonstrating that he was well aware of the importance of the speech bandwidth and blind speech testing in speech perception. It was the creation of the telephone that both allowed and pushed mathematicians and physicists to develop the science of speech perception. Critical to this development was probability theory. One of their main tools was the confusion matrix which estimates the probability of hearing phoneme P_i when speaking phoneme P_j. From 1910 to 1950 speech perception was extensively studied in telephone research departments. However it was the work of Harvey Fletcher in 1921 that made the first major breakthroughs. By 1930 millions of dollars were being spent every year on speech perception research at the newly created Bell Labs. The key was his quantification of the transmission of information, as characterized by phone error patterns. Fletcher's full and final theory was not published until 1950, following his AT&T retirement. The next breakthroughs were provided by George Miller and his colleagues at the Harvard Acoustics Lab during and following WWII. Miller used concepts from information theory, developed at Bell Labs by Claude Shannon, to quantify speech entropy. While these studies provide key insight into speech perception, they do not take the final elusive step that would allow us to build robust automatic speech recognition (ASR) machines. Regardless of what you read in the popular press, ASR is still an unsolved problem. I will attempt to pass along some wisdom I have learned over the years on what we now know about human speech recognition (HSR). It is hoped that by learning more about HSR we might make ASR robust to noise and filtering. Today ASR is based on language models which have not, and can not, give ASR the basic robustness to noise and filtering found in HSR. I will summarize important results from the 30 years of work by Fletcher and his colleagues, which resulted in the ``articulation index,'' a widely recognized method for characterizing the information bearing frequency regions of speech. Next I summarize the speech work of George Miller. Miller showed the importance of source entropy (randomness) in speech perception. He did this by controlling for both the cardinality (size of the test corpus) and the signal to noise ratio of the speech samples. Finally I briefly describe recent experimental work important to robust ASR. The goal is to make a system that works as well as human listeners at decoding degraded (filtering plus noise) nonsense speech sounds. Two important conclusions I draw from these studies are that:
1) The robustness to noise and filtering for HSR is independent of language context effects. This is good news for the speech recognition scientist as it allows us to uncouple these two complex systems.
2) Psychophysical studies can also provide us with the insight we need to replicate HSR's robustness in ASR.
