Header logo is ei


2003


no image
Study of Human Classification using Psychophysics and Machine Learning

Graf, A., Wichmann, F., Bülthoff, H., Schölkopf, B.

6, pages: 149, (Editors: H.H. Bülthoff, K.R. Gegenfurtner, H.A. Mallot, R. Ulrich, F.A. Wichmann), 6. T{\"u}binger Wahrnehmungskonferenz (TWK), Febuary 2003 (poster)

Abstract
We attempt to reach a better understanding of classi cation in humans using both psychophysical and machine learning techniques. In our psychophysical paradigm the stimuli presented to the human subjects are modi ed using machine learning algorithms according to their responses. Frontal views of human faces taken from a processed version of the MPI face database are employed for a gender classi cation task. The processing assures that all heads have same mean intensity, same pixel-surface area and are centered. This processing stage is followed by a smoothing of the database in order to eliminate, as much as possible, scanning artifacts. Principal Component Analysis is used to obtain a low-dimensional representation of the faces in the database. A subject is asked to classify the faces and experimental parameters such as class (i.e. female/male), con dence ratings and reaction times are recorded. A mean classi cation error of 14.5% is measured and, on average, 0.5 males are classi ed as females and 21.3females as males. The mean reaction time for the correctly classi ed faces is 1229 +- 252 [ms] whereas the incorrectly classi ed faces have a mean reaction time of 1769 +- 304 [ms] showing that the reaction times increase with the subject's classi- cation error. Reaction times are also shown to decrease with increasing con dence, both for the correct and incorrect classi cations. Classi cation errors, reaction times and con dence ratings are then correlated to concepts of machine learning such as separating hyperplane obtained when considering Support Vector Machines, Relevance Vector Machines, boosted Prototype and K-means Learners. Elements near the separating hyperplane are found to be classi ed with more errors than those away from it. In addition, the subject's con dence increases when moving away from the hyperplane. A preliminary analysis on the available small number of subjects indicates that K-means classi cation seems to re ect the subject's classi cation behavior best. The above learnersare then used to generate \special" elements, or representations, of the low-dimensional database according to the labels given by the subject. A memory experiment follows where the representations are shown together with faces seen or unseen during the classi cation experiment. This experiment aims to assess the representations by investigating whether some representations, or special elements, are classi ed as \seen before" despite that they never appeared in the classi cation experiment, possibly hinting at their use during human classi cation.

PDF Web [BibTex]

2003

PDF Web [BibTex]


no image
A Representation of Complex Movement Sequences Based on Hierarchical Spatio-Temporal Correspondence for Imitation Learning in Robotics

Ilg, W., Bakir, GH., Franz, MO., Giese, M.

6, pages: 74, (Editors: H.H. Bülthoff, K.R. Gegenfurtner, H.A. Mallot, R. Ulrich, F.A. Wichmann), 6. T{\"u}binger Wahrnehmungskonferenz (TWK), February 2003 (poster)

Abstract
Imitation learning of complex movements has become a popular topic in neuroscience, as well as in robotics. A number of conceptual as well as practical problems are still unsolved. One example is the determination of the aspects of movements which are relevant for imitation. Problems concerning the movement representation are twofold: (1) The movement characteristics of observed movements have to be transferred from the perceptual level to the level of generated actions. (2) Continuous spaces of movements with variable styles have to be approximated based on a limited number of learned example sequences. Therefore, one has to use representation with a high generalisation capability. We present methods for the representation of complex movement sequences that addresses these questions in the context of the imitation learning of writing movements using a robot arm with human-like geometry. For the transfer of complex movements from perception to action we exploit a learning-based method that represents complex action sequences by linear combination of prototypical examples (Ilg and Giese, BMCV 2002). The method of hierarchical spatio-temporal morphable models (HSTMM) decomposes action sequences automatically into movement primitives. These primitives are modeled by linear combinations of a small number of learned example trajectories. The learned spatio-temporal models are suitable for the analysis and synthesis of long action sequences, which consist of movement primitives with varying style parameters. The proposed method is illustrated by imitation learning of complex writing movements. Human trajectories were recorded using a commercial motion capture system (VICON). In the rst step the recorded writing sequences are decomposed into movement primitives. These movement primitives can be analyzed and changed in style by de ning linear combinations of prototypes with di erent linear weight combinations. Our system can imitate writing movements of di erent actors, synthesize new writing styles and can even exaggerate the writing movements of individual actors. Words and writing movements of the robot look very natural, and closely match the natural styles. These preliminary results makes the proposed method promising for further applications in learning-based robotics. In this poster we focus on the acquisition of the movement representation (identi cation and segmentation of movement primitives, generation of new writing styles by spatio-temporal morphing). The transfer of the generated writing movements to the robot considering the given kinematic and dynamic constraints is discussed in Bakir et al (this volume).

PDF Web [BibTex]

PDF Web [BibTex]


no image
Models of contrast transfer as a function of presentation time and spatial frequency.

Wichmann, F.

2003 (poster)

Abstract
Understanding contrast transduction is essential for understanding spatial vision. Using standard 2AFC contrast discrimination experiments conducted using a carefully calibrated display we previously showed that the shape of the threshold versus (pedestal) contrast (TvC) curve changes with presentation time and the performance level defined as threshold (Wichmann, 1999; Wichmann & Henning, 1999). Additional experiments looked at the change of the TvC curve with spatial frequency (Bird, Henning & Wichmann, 2002), and at how to constrain the parameters of models of contrast processing (Wichmann, 2002). Here I report modelling results both across spatial frequency and presentation time. An extensive model-selection exploration was performed using Bayesian confidence regions for the fitted parameters as well as cross-validation methods. Bird, C.M., G.B. Henning and F.A. Wichmann (2002). Contrast discrimination with sinusoidal gratings of different spatial frequency. Journal of the Optical Society of America A, 19, 1267-1273. Wichmann, F.A. (1999). Some aspects of modelling human spatial vision: contrast discrimination. Unpublished doctoral dissertation, The University of Oxford. Wichmann, F.A. & Henning, G.B. (1999). Implications of the Pedestal Effect for Models of Contrast-Processing and Gain-Control. OSA Annual Meeting Program, 62. Wichmann, F.A. (2002). Modelling Contrast Transfer in Spatial Vision [Abstract]. Journal of Vision, 2, 7a.

[BibTex]

1998


no image
Masking by plaid patterns: effects of presentation time and mask contrast

Wichmann, F., Henning, G.

pages: 115, 1. T{\"u}binger Wahrnehmungskonferenz (TWK 98), February 1998 (poster)

Abstract
Most current models of early spatial vision comprise of sets of orientation- and spatial-frequency selective filters with our without limited non-linear interactions amongst different subsets of the filters. The performance of human observers and of such models for human spatial vision were compared in experiments using maskers with two spatial frequencies (plaid masks). The detectability of horizontally orientated sinusoidal signals at 3.02 c/deg was measured in standard 2AFC-tasks in the presence of plaid patterns with two-components at the same spatial frequency as the signal but at different orientations (+/- 15, 30, 45, and 75 deg from the signal) and with varying contrasts (1.0, 6.25 and 25.0% contrast). In addition, the temporal envelope of the stimulus presentation was either a rectangular pulse of 19.7 msec duration, or a temporal Hanning window of 1497 msec.Threshold elevation varied with plaid component orientation, peaked +/- 30 deg from the signal where nearly a log unit threshold elevation for the 25.0% contrast plaid was observed. For plaids with 1.0% contrast we observed significant facilitation even with plaids whose components were 75 deg from that of the signal. Elevation factors were somewhat lower for the short stimulus presentation time but were still significant (up to a factor of 5 or 6). Despite of the simple nature of the stimuli employed in this study-sinusoidal signal and plaid masks comprised of only two sinusoids-none of the current models of early spatial vision can fully account for all the data gathered.

Web [BibTex]

1998

Web [BibTex]


no image
A bootstrap method for testing hypotheses concerning psychometric functions

Hill, N., Wichmann, F.

1998 (poster)

Abstract
Whenever psychometric functions are used to evaluate human performance on some task, it is valuable to examine not only the threshold and slope values estimated from the original data, but also the expected variability in those measures. This allows psychometric functions obtained in two experimental conditions to be compared statistically. We present a method for estimating the variability of thresholds and slopes of psychometric functions. This involves a maximum-likelihood fit to the data using a three-parameter mathematical function, followed by Monte Carlo simulation using the first fit as a generating function for the simulations. The variability of the function's parameters can then be estimated (as shown by Maloney, 1990), as can the variability of the threshold value (Foster & Bischof, 1997). We will show how a simple development of this procedure can be used to test the significance of differences between (a) the thresholds, and (b) the slopes of two psychometric functions. Further, our method can be used to assess the assumptions underlying the original fit, by examining how goodness-of-fit differs in simulation from its original value. In this way data sets can be identified as being either too noisy to be generated by a binomial observer, or significantly "too good to be true." All software is written in MATLAB and is therefore compatible across platforms, with the option of accelerating performance using MATLAB's plug-in binaries, or "MEX" files.

[BibTex]


no image
Support Vector methods in learning and feature extraction

Schölkopf, B., Smola, A., Müller, K., Burges, C., Vapnik, V.

Ninth Australian Conference on Neural Networks, pages: 72-78, (Editors: T. Downs, M. Frean and M. Gallagher), 1998 (talk)

[BibTex]

[BibTex]


no image
Nonlinearities and the pedestal effect

Wichmann, F., Henning, G., Ploghaus, A.

Perception, 27, pages: S86, 1998 (poster)

Abstract
Psychophysical and physiological evidence suggests that luminance patterns are independently analysed in "channels" responding to different bands of spatial frequency. There are, however, interactions among stimuli falling well outside the usual estimates of channels' bandwidths (Henning, Hertz, and Broadbent, (1975). Vision Res., 15, 887-899). We examined whether the masking results of Henning et al. are consistent with independent channels. We postulated, before the channels, a point non-linearity which would introduce distortion products that might produce the observed interactions between stimuli two octaves apart in spatial frequency. Standard 2-AFC masking experiments determined whether possible distortion products of a 4.185 c/deg masking sinusoid revealed their presence through effects on the detection of a sinusoidal signal at the frequency of the second harmonic of the masker-8.37 c/deg. The signal and masker were horizontally orientated and the signal was in-phase, out-of-phase, or in quadrature with the putative second-order distortion product of the masker. Significant interactions between signal and masker were observed: for a wide range of masker contrasts, signal detection was facilitated by the masking stimulus. However, the shapes of the functions relating detection performance to masker contrast, as well as the effects of relative phase, were inconsistent with the notion that distortion products were responsible for the interactions observed.

[BibTex]

[BibTex]