Planning Classification Image Experiments

Original Abstract

 

Planning support using classification image estimation results from JOV 1(2) papers (Ahumada, 2002; others, 2002).

 

Outline

Example classification image experiment: Ahumada, 1996 ECVP

JOV 1(2) Results

Planning Decisions (iterative process)

Stimulus representation

Dimensionality of the representation. m

External noise distribution and level. s E

Internal noise level estimation. s I

Number of trials. N s

Confidence interval size or power of tests.

 

We will look at some alternatives and examples for each decision.

 

Introduction

 

Example classification image experiment: Ahumada, 1996 ECVP.

 

JOV 1(2) and Newer* Results

            (Assuming linear decision function)

            Improved estimates of classification image when there is response bias. (Murray, Bennett, and Sekuler).

            Improved estimates of errors in classification image using estimates of internal noise (Murray, Bennett, and Sekuler).

            Improved methods for specific hypothesis testing (Abbey and Eckstein; Solomon).

            Methods for estimating internal noise from the signal-to-noise ratio of the classification image (Nykamp and Ringach).

 

*Simpler formulas for estimating internal noise from repeated noise trials experiments.

            *Simpler formulas for estimating internal noise from the signal-to-noise ratio of the classification image.

 

 

Planning Decisions (iterative process)

 

Stimulus representation where the classification is expected to have a large linear component.

Example problem: Low vs. high spatial frequency; 

Example solutions:

Auditory frequency component amplitude;

Solomon FFT power;

Contrast energy (Neri and Heeger, 2002 Nature Neuroscience).

General nonlinear search spaces:

E. Sutter, Wiener kernel analysis;

S. Klein, Book chapter.

Dimensionality of the representation. m

External noise distribution and level. s E

            Noise correlation matrix: white, pink, narrow-band, low-pass

            Dynamic range considerations

            Masking considerations:

Gain control (internal noise) from linearly irrelevant components.

            Mimicking internal noise can mean more “ecological” validity.

                        Distribution: spectrum (Ahumada and Watson, 1985; Solomon, 2000) ;

                        Level: Beard and Ahumada (1997; 1998) criterion: “no change in performance”.

Internal noise level estimation. s I

            Frozen noise method (Burgess and Colborne, 1988);

            Nykamp and Ringach method.

Number of trials (per stimulus). N S.

Size of classification image pixel confidence intervals or the power for correctly selecting among alternative hypotheses.

            Multiply per trial signal-to-noise ratio in classification image by N S

(The dimensionality of the representation m comes through by reducing the size of the normalized weights)

The number of parameters being estimated controls the necessary sample size, not the number of pixels. 

Testing specific hypotheses (Abbey and Eckstein; Solomon).

           

Acknowledgements