Hysteresis in the Perception of Visual Unity

Confirmation of a Neural Network Model Prediction

Ian Bright

and Raymond Pavloski

Departments of Psychology and Mathematics, Indiana University of Pennsylvania,

1020 Oakland Avenue, 15701, Indiana, PA, U.S.A.

Department of Psychology, Indiana University of Pennsylvania,

1020 Oakland Avenue, 15701, Indiana, PA, U.S.A.

Keywords: Hard Problem, Neurotechnology, Qualia, Recurrent Neural Network, Tolerance Space, Topological Vision,

Visual Object Unity.

Abstract: In response to a simulated retinal image of an object, the recurrent input to a richly connected artificial

neural network organizes into a connected open set (COS) of ionic conductance values, which models the

continuity and unity of a visual object. As the density of light dots on a dark background increases and then

decreases, a COS appears at a density that is higher than that at which it disappears (hysteresis). This

experiment tested the hypothesis that humans will show hysteresis similar to that of the simulation. In

addition, the effect of dot lightness on the perception of a unified visual object was also tested.

1 INTRODUCTION

It has been suggested by the work of Chen (2005)

and Zhang (2005) that the perceived unity of a visual

object is a primitive of the visual system and is

organized in terms of topological properties. While

there is strong experimental data to support this view

of the early visual system, an explanation of how

neural networks process and sort topological

features remains elusive (Pomerantz, 2003). This is

an important problem: understanding how the brain

organizes and separates visual objects would open

new possibilities in the field of neurotechnology

(e.g., a neural prosthesis that restores vision for a

stroke victim with damage to a portion of V1.)

The topological approach to perception advanced

by Chen (2005) has been implemented in work on

computational vision. In this approach, a tolerance is

defined as a range within the image within which

variations are ignored for the subsequent purpose of

computing connectivity (Huang, Huang, Tan & Tao,

2010). While this work demonstrates how the global

property of connectivity can be computed from

discrete elements, it does not address the issue of the

perception of connected elements as visual unity.

This latter issue can be investigated by

determining if there exist natural sources of

tolerance to small differences between the ionic

conductances produced by closely positioned

neurons in a recurrent neural network (RNN). Using

this approach, Pavloski (2015) constructed a RNN

consisting of a 33 x 33 lattice of excitatory neurons

interconnected with an 11 x 11 lattice of inhibitory

neurons for stability. The excitatory neurons were

stimulated by a 33 x 33 lattice of simulated retina

cells. Results showed that a connected open set

(COS) of ionic conductance vectors quickly emerges

from the input of a simulated object image.

Furthermore, COSs mimic several visual phenomena

including just noticeable differences, grouping by

proximity, human V1 fMRI data for real and

apparent motion, and object constancy over rotation

and changes of size and orientation of an image.

Dynamical effects of sequential presentations of

stimuli are well documented (e.g., Tuller, Case,

Ding and Kelso, 1994), and Pavloski (2015) tested

the RNN using sequential presentations of visual

images to determine those image parameters at

which a COS first appears and subsequently

disappears. The images used consist of 140 one-

pixel spots. One hundred of the spots are

repositioned randomly within the 33 x 33 pixel

image area on each iteration (update of all network

neurons). The remaining 40 spots are positioned

randomly within a square window the sides of which

are reduced from a length of 33 pixels to a length of

seven over the first 31 iterations; the sides increase

Bright, I. and Pavloski, R..

Hysteresis in the Perception of Visual Unity - Conﬁrmation of a Neural Network Model Prediction.

In Proceedings of the 3rd International Congress on Neurotechnology, Electronics and Informatics (NEUROTECHNIX 2015), pages 49-53

ISBN: 978-989-758-161-8

over the remaining 30 iterations to the original

length of 33. The inset in Figure 1 shows images

from iterations 1, 10, 20 (top row), 30, 31, 32

(middle row), and 42, 52, and 61 (bottom row).

Hysteresis is present; a COS first appears when the

additional dots are enclosed in a square with sides of

length 13 during initial decreases in window size

(→)

and disappears only when the sides are subsequently

increased (←) to a length of 19.

Figure 1: Examples of retinal images presented to the

RNN (left) and the number of conductance vectors in the

COS are plotted against the length of the sides of the

smaller window (Pavloski, 2015).

This result motivates the present studies, which

tested the hypothesis that human participants would

demonstrate hysteresis in the perception of object

unity when observing visual images similar to those

used in the above simulation.

2 EXPERIMENT 1

In an initial attempt to test the predictions of

Pavloski’s (2015) neural network, a pilot study was

conducted to determine parameters at which

hysteresis could be expected to occur in human

participants.

2.1 Background

It has been demonstrated by Poltoraski and Tong

(2014) that hysteresis occurs in the perception of

scenes and objects. For example, when shown a

series of pictures of a living room, participants who

were shown a close up of a table first and then

shown images that gradually zoomed out considered

more of the pictures to be of the table, rather than a

living room, when compared to those who began

with a broader picture of the living room and

gradually zoomed in.

Hysteresis has also been demonstrated in the

auditory system. In a series of experiments

conducted by Tuller, Case, Ding, and Kelso (1994),

participants were presented with a male utterance of

a /s/ sound followed by a silent gap lasting between

0 and 76 ms (increased in increments of 4 ms) and

an electronically generated /ay/ sound for which the

first formant had an onset frequency of either 230

Hz (biased toward the perception of stay) or 430 Hz

(biased toward the perception of say). Tuller et al.

used these two types of stimuli so that participants

could not simply count stimuli in order to report

their perceptions of either say or stay. At a shorter

silent gap it was found that participants heard the

word say, whereas participants heard the word stay

when a longer gap was present. During ordered

presentations, Tuller et al. found that perception of

the word stay occurred at a far longer silence gap

than the gap at which it disappeared. As expected,

the 230 Hz /ay/ first formant onset frequency was

found to be more likely to lead to the perception of

the word stay than was the 430 Hz first formant

onset frequency.

2.2 Method

Five undergraduate participants volunteered to take

part in the experiment. All participants had normal

or corrected to normal vision. Participants were

placed in a booth constructed to block out

extraneous light and viewed images (created using

Mathematica Version 9) projected onto a screen.

Each image consisted of 1500 dots of gray level

0.335 randomly placed on a gray level 0.1, 10x10

square. An additional 100 gray level 0.335 dots were

randomly positioned within an area the size of which

changed from one image to the next. The length of

each individual dot was .2 percent of the width of the

video screen. A sample image is shown in Figure 2.

Participants took part in two trials consisting of

four runs of images (i.e. increasing, decreasing, and

random change in area containing the additional 100

dots). The sequence of the runs in each trial was

randomly assigned with the constraint that each

ordered run would be followed by a random run and

vice-versa. Each run in the trials contained the same

100 images.

Images were displayed for 500 ms, and were

immediately replaced by a plain white screen.

Participants were instructed verbally respond yes if a

unified object was perceived and respond no if no

unified object was perceived. The next image would

not appear until a response was recorded. Timing

and recording of answers was handled by E-Prime

software. In between the two trials, participants were

NEUROTECHNIX 2015 - International Congress on Neurotechnology, Electronics and Informatics

given up to a five-minute break.

Figure 2: An example of a high density/low area image is

shown. Contrast is enhanced for ease of viewing.

2.3 Results

A one-way within-subjects analysis of variance was

performed on the data. As shown in Figure 3,

hysteresis was clearly present. As expected, the

order in which the images were presented was found

to be significant F(1,4) = 109.394, p < .01. The

decreasing density trials were more likely to lead to

the experience of a unified object than the increasing

density trials.

Figure 3: A graph of the number of times each image was

reported to contain an object during the runs.

2.4 Discussion

The results of the pilot study were consistent with

the prediction of the neural network with respect to

the presence of hysteresis. In looking at Figure 3, it

is easy to see that in the middle of the runs

participants were less likely to see a unified object as

the density increased when compared to the

decreasing density runs. This is in comparison to the

high and low densities, where the two runs

demonstrated no difference.

3 EXPERIMENT 2

Building on the results of Experiment 1, a more

robust study was performed to determine if the

perception of a unified object is characterized by

hysteresis in human participants.

3.1 Changes from Experiment 1 and

Rationale for Changes

While having subjects report their response verbally

was successful in minimizing errors, it did have its

drawbacks. Participants frequently reported fatigue

from having to give hundreds of verbal responses.

Additionally, having trials consist of 400 images

also resulted in fatigue. Participants would

sometimes take unscheduled breaks during the

experiment by delaying a response as an attempt to

recover.

In an attempt to decrease fatigue, a wireless

keyboard was used to allow the participants to input

their responses. The highest and lowest dot densities

were also moved closer together, as the original

maximum and minimum demonstrated no

difference. In addition, the change in density

between images was increased to get the number of

images contained in each run down to 20. This

decreased the number of images in each trial to 80.

The number of trials were increased to 10, however

resulting in the same number of total images shown.

Participants were given a two-minute break between

each pair of trials also to decrease fatigue.

Another frequently reported issue was directed

towards the white screen that flashed following the

presentation of each image. Participants often

complained that the bright light of the image caused

discomfort for those who were inside the otherwise

completely darkened booth. In an attempt to

decrease this discomfort a dark gray image was

presented between each pair of images.

The second experiment also included the use of a

second independent variable that was manipulated

by Tuller et al. (1994), but not employed in our pilot

research. As noted above, the first formant of the

electronically generated /ay/ sound used by Tuller et

al. had a first formant onset frequency of either 230

Hz (biased toward the perception of stay) or 430 Hz

Hysteresis in the Perception of Visual Unity - Conﬁrmation of a Neural Network Model Prediction

(biased toward the perception of say). Tuller et al.

used these two types of stimuli so that participants

could not simply count stimuli in order to report

their perceptions of either say or stay. In order to

include this manipulation in our second experiment,

the gray level of the additional 100 dots was set

either to the gray level of the 1500 background dots

(0.335) or to 0.375. It was anticipated that this dot

brightness manipulation would produce a significant

effect in reports of a unified visual object, with the

higher gray level dots being more likely to be

perceived as unified.

3.2 Method

Nine undergraduate participants volunteered to take

part in the experiment. All participants had normal

or corrected to normal vision. Participants were

placed in a booth constructed to block out

extraneous light and viewed images (created using

Mathematica Version 10) projected onto a screen.

Each image consisted of 1500 dots of gray level

0.335 randomly placed on a gray level 0.1, 10x10

square. An additional 100 dots of either gray level

0.335 or 0.375 were randomly positioned within an

area the size of which changed from one image to

the next. The length of each individual dot was .2

percent of the width of the video screen.

Participants took part in 10 trials consisting of

four runs of images (i.e. increasing area, decreasing

area, and random area). The sequence of the runs in

each trial was randomly assigned with the constraint

that each ordered run would be followed by a

random run and vice-versa. Each run in the trials

contained the same 20 images. Five of the trials

consisted of dots that were brighter than the other

five.

Images were displayed for 500 ms, and were

immediately replaced by a plain gray level 0.1

image. Participants were instructed to press one key

if a unified object was perceived and a second key if

no unified object was perceived. The next image

would not appear until a response was recorded.

Timing and recording of answers was handled by E-

Prime Software. Following each trial, participants

were given up to a two-minute break.

3.3 Results

The results of this experiment are shown in Figure 4.

A two-way within-subjects analysis of variance was

performed on the data. The order in which the

images were presented was found to be significant at

F(1,8) = 13.5467, p < .01. The decreasing density

trials were more

likely to lead to the experience of

an image than the increasing image. Dot brightness

did not produce a significant effect.

Figure 4: This graph plots the estimated probability that an

image was reported to contain an object during increasing

and decreasing runs across both levels of brightness.

4 DISCUSSION AND FUTURE

RESEARCH

While the results of the study were consistent with

the predictions of the neural network with respect to

the presence of hysteresis, the hysteresis was not as

strong as that found in the pilot study. A possible

explanation for this may be the fewer number of

images that were used in the individual runs when

compared to the RNN. Along the same lines, it is

also possible that the increased change in density

between the images may have contributed to the

decreased hysteresis. Additionally the brightness did

not play a significant role in changing the perception

of the images. It is possible that the change in

brightness levels was not large enough to cause any

difference in reports of a unified object. Further tests

will need to be performed to determine what role

brightness may play in the perception of a unified

object.

Looking ahead, further experiments will be run

to continue testing the predictions of the RNN. By

decreasing the change in the area of the dot square

between images as well as changing the low and

high points of the area, it is predicted that hysteresis

will increase. In addition to manipulating the area

the square of dots occurs, experiments in

manipulating the density of the dots with a fixed

area will also be performed. Lastly, different

presentation methods, such as the use of virtual

reality headsets will be utilized in an attempt to

further generalize the results.

NEUROTECHNIX 2015 - International Congress on Neurotechnology, Electronics and Informatics

REFERENCES

Chen, L. (2005). The topological approach to perceptual.

organization. Visual Cognition, 12(4), 553-637.

Huang, Y., Huang, K., Tan, T., & Tao, D. (2010). A novel

visual organization based on topological perception.

Computer Vision – ACCV 2009. Lecture Notes in

Computer Science, Volume 5994, 180-189.

Pavloski, R. (2015). Toward sentient neurotechnology:

Visual object unity may be structured by and constrain

neural interactions. Neurotechnix 2015.

Poleratski, S. & Tong, F. (2014) Hysteresis in the

Dynamic Perception of Scenes and Objects. Journal of

Experimental Psychology: General, 143(5), 1875–

1892.

Pomerantz, J. R. (2003). Wholes, holes, and basic features

in vision. Trends in Cognitive Sciences, 7(11), 471-

473.

Tuller, B., Case, P., Ding, M., & Kelso, J.A.S. (1994). The

nonlinear dynamics of speech categorization. Journal

of Experimental Psychology: Human Perception and

Performance, 20(1), 3-16.

Zhang, J. (2005). Object oneness: The essence of the

topological approach to perception. Visual Cognition,

12(4), 683-690.

Hysteresis in the Perception of Visual Unity - Conﬁrmation of a Neural Network Model Prediction