VSS 2015 Note

21.21 Long et al. developed a size stroop effect paradigm is which the real world size and the image size may have a different relation among test objects. They had observers to select one of the two objects with bigger sizes. The object with smaller image may have a bigger real world size. The RT for the congruent condition is smaller than the incongruent condition. For most objects, Size classification accuracy cannot predict identification performance, but stroop effect ER change. Basic level recognition is not necessary to activate real world size.

21.22 The authors explored the cortical representation in different areas in the ventral stream of the object categorical boundaries. They Focused on V1, V2, hV4 and LOC. The computed the correlation of voxel activations for different objects and used it as a measurement of distance. The within-category difference decreased while between-category distance increased from V1 to LOC. The less typical exemplar tend to have a greater distance from the centroid of the category.

22.24 They use fMRI adaptation paradigm. The stimuli were colored ring modulated in Red/green or luminance. The observers were adapted to blank, R/G or ACH ring and tested with R/G or ACH ring. Vo is identified by contrasting colored and ACH contrast. Without adaptation, V4,V3V and VO all shows a bias toward color rings. Adapted to R/G reduced R/G test activation everywhere, but ACH adaptor produced a significant reduction effect for R/G activation only in VO. The ACH adaptor produced more effect than R/G on ACH  test only in dorsal areas.

[I am not sure what is new about this study that we have not known already from previous literature. ]

22.26 Tyler demonstrated a new color induction effect in which one can perceive color of dot inducers on a blank field.

23.4072 The authors used a block design fMRI to measure BOLD response to images of six individuals in six expressions. They compute the correlation of voxel responses in FFA, OFA and STS. They then correlated the fMRI responses with extracted internal and external images. They found internal features correlated well with expression and external, identity.

[We can really use this paradigm for our ORE on facial expression experiment. We were able to extract various images features. Instead component analysis, we can just correlate these features with fMRI activation to observe which feature can predict BOLD activation to ORE.] 

26.4005 Using nominal or subjective equilluminance S-cone modulation colors has no effect on VEP.

26.4007 In V1, the BOLD activation for Glaucoma patients for ACH,RG or BY patterns are lower than normal. However, in LGN, glaucoma patients actually have greater activations.

26.4008 This paper from Kingdon's lab measured the fusion threshold for dichoptic colors. They showed that the fusion threshold increased with luminance contrast.

26.4080 The authors showed a novel Lemon illusion, in which the observers may perceive curvature on straight lines (thus, shape like a lemon) if the line is bound by curves.

26.4081  The authors suggested that Ebbinghaus illusion is not due to size contrast. The difference between target and background size is not sufficient (it is possible to have size contrast without illusion) nor necessary (one can produced the illusion by a squared contour surrounding the target).

33.3016 This study from Webster's lab measured the average color of a patch of random dots in which the color of a pixel was sample from a distribution along a color axis. They found that the variability of white setting increased significantly with any added random variability in the image.    

33.3018 The authors showed that the afterimage for induced color is not complementary colors but biased toward S-cone contribution.

33.3049 The authors changed the viewpoint and expression of facial images and measured the classification of BOLD activations for those images.  They found that IT can classify both viewpoint and expression. However, when there is  a joined change in viewpoint and expression. The expression difference dominate classification.

33.4037 Boynton used M-sequence to measure fMRI pRF. They found that pRF ignored the scotoma in stimulus presentation, suggesting a filling-in process is involved. 

33.4041  Mulligan used a titration method to calibrate a display without photometer. The stimulus including stripes of a checkerboard of two end points to be bisected and a gray. It alternating with gray-upper-end and gray-lower-end checkerboards. If the gray is too bright, it would move to an opposite direction form when it is too dark. Thus, we can use minimum motion as a criterion to determine the gray level that bisect the range.

33.4087 The authors found that the fMRI activation in PPA and surrounding areas can classify building of different styles, but not different architects of the same style.

36.3001 The authors made a review on several aspects of brain responses to symmetry patterns.

36.4001 McCourt et al. manipulated the luminance of the collinear or the context bars in the White illusion display. They found that the illusion mainly changed with the collinear luminance and much less with context luminance.

35.24 This study from Gallent's lab decomposed images into features (SF, depth, surface orientation). Send the observers to fMRI scanner and estimated the weighting of each feature per voxel. Low level features are from converting input images Gabor wavelets. The depth were from depth map (the test images are rendered from a 3D model) and the orientation is from surface norms. [CC: this is clever. Using the rendered images, they completely avoid the tedious task of image understanding). The local features predicts V1 responses. But the orientation and depth combined predicts PPA responses. They can then construct the voxel receptive field, with particular depth and surface orientation selectivity.    

35.28 The authors compared human behavior data with electrophysiological response to nature scene. The behavior task is simple 2AFC detection task. The target was Gabor patch on a patch of a natural scene background. The result is a typical no-dipper TvC function (for background was a patch of nature scene and thus provided little excitation to the detector). They then measured the voltage sensitive dye imaging on monkey brain to the target. The brain area for the target is used as ROI. They then measured the response to background in target ROI in each trial and got a histogram of response distribution. They then computed neurometrics function for various background and estimated threshold for the target contrast that gave d'=1 for different background contrasts. The response, when appropriately scheduled, fit human performance data well.

52.13 The authors use emotional faces and perceptual similarity task (whether two faces had the same emotion). They showed that there is culture difference between Chinese and British in perceptual similarity task. There is a difference in categorization. They then repeated the experiment with lower and higher half faces. Again, there is no difference between culture. The categorization showed a difference in the lower face region, not in the upper face regions.    They then tried two databases (Chinese and Caucasian). Again, a own-group advantage in lower face regions.

52.15 The authors found that there is a low correlation between different holistic face tasks (inversion, composite and part-whole). Only face inversions correlated with face perception (0.39). There is no correlation between composite face task and face perception,

53.4002 The authors used typical phase tagging paradigm but swap 2D spatial change with depth change. In this case, they claimed to be able to map depth tuning in cortex. [Their result seems to be a very preliminary stage. It remains to be seen whether their method holds].

53.4007 The authors measured the fMRI bold response to rings of different orientation. They used MVPC to identify the voxels that contribute most to classification and use population receptive field to identify the pRF of these voxels. They found that the ones that contributed most to classification are those tune to the edges of the rings.

61.13 The authors uses MEG to measure broadband field response, that is normally acquired with Ecog. The signal is a broadband increase related to baseline, other than the stimulus driven response. It is a proxy for mean local spiking activity and is correlated with fMRI BOLD activation. The MEG field potential is weaker and more pronounced at high frequency. It is spatially localized by low in SNR. They used a denoising algorithm. The noise is acquired from PCA of channels that do not respond to the stimuli (e.g., frontal electrodes in visual stimulation).   In that way, the claim to be able to measure reliable field potential.

61.14 Following the previous presentation from the same lab, they compared MEG measured local field potential and BOLD. They extract several frequency band from MEG signal and correlated it with BOLD and found individual bands do not correlate with BOLD well, but broadband field potential.

61.15 They modeled ganglion cell with normalized response then predict the ganglion responses to nature images and compute the statistics of the natural images and retinal signals. The retinal off-neurons dominated low frequency response and there is an interaction in contrast gain control between P- and M-cells.  

63.4066 This paper from Daniel Baker's lab is also identical to YiChen's thesis. We need to get it published ASAP.

SFN 2014 Note

155.09 Smith et al., from Fitzpatrick's lab, used optical image to test the development of cortical maps of ferrets at different stages of development. They found that the on-off map was developed first before the orientation map after eye opening.

155.18. Hermes et al used Kay's model for BOLD response to model ECoG activation. Their idea was that ECoG measured the field potential which is highly correlated with BOLD in fMRI. Thus, they can use a model for BOLD for ECoG.

236.13-16 A series of posters on T. Lee's lab used images with scene context to measure the RF properties of V2 neurons and model their distributed representation.

236.18 Nandy et al. showed that the sharpness of shape tuning is of a V4 cell is inversely proportional to its location tuning. Both tuning function changes with time after stimulus onset.     

332.11 In this curious study, Lafer-Sousa et al. measured BOLD activation on human and monkey with either black-and-white or colored pictures of scene and faces. They showed that in both species, color area is sandwiched between face area and scene area. Hence, the ventral stream of both species has a similar organization.

333.03 Emmerling et al. use 7T MRI to get 0.7x0.7 voxel size. They measured BOLD activation for images with central square that is either far or near from the background. Then, they did a MVPA on MT response for disparity tuning. They only see some weak effect of disparity tuning.

Their technology is impressive, but I think that they are on the wrong track. MVPA is a bad tool in this case. MT should contain all sorts of disparity selectivity cell. They may not favor one disparity over the other.

333.05 Zeater et al. should that there are many binocular cells in LGN. A result that is going to rewrite many textbooks.

333.07 Cooling caudal intraparietal area reduced the activation in many 3D related areas in the monkey brain and impair their 3D discrimination. It reveals a huge network for 3D scene processing.

489.08 Nassi et al. inject optogenetic virus into V1. They then stimulated V1 neurons either directly with laser (optogenetic stimulation) or with visual stimuli. Both drove excitatory and inhibitory responses. The interaction of the two effects can be explained with a normalization model.

582.04 Meyer et al. showed faces with different identity and viewpoints to monkeys. Then, measure single cell responses in different face patches. They then used pattern classification method to show what information is decoded in each patch. 

674.03 Felleman et al. used optical imaging to study color and orientation tuning, They showed a pinwheel and quasi-linear organization for both hus and orientation in V4. However, only 5% of V4 neuron show selectivity for both properties. Most V4 neurons are selective to either one of two.

674.04 There are two types of activation for texture segregation: Feedfoward: Feature extraction, boundary detection, feedback: region filling. The authors measured the V1 response while microstimulate the V4 neurons with the same receptive field. The microstimulation causes a brief excitation followed by inhibition at V4, and slightly drive V1 response. 

674.08 Jonas et al. from Rossion lab measured the intracranial responses to faces. They presented nonface object in 1.2Hz and face in 6Hz. From the FFT analysis, they showed that the energy ratio between face freq. and nonface Freq. increased gradually from V1 to FFA.

674.09 The excitatory and inhibitory features of V4 neurons have orthogonal orientation.  This produced curvature selectivity in V4

674.10 Lateral categorical regions biased toward lower visual field while ventral categorical regions biased toward upper visual fields. The authors then estimated pRF of voxels in several object areas used windowed scene images. They showed PPA pRF is biased toward upper visual field, while TOS, lower visual field.

772.01 Dumoulin group asked where visual cortex responding to retinal position of image location. They measured the pRF for images either the left and the right eye are of the same image or left and the right eye images has different location, but not at the same time. They then fit pRF model for either eye. The brain is able to respond to the offset. They then present images with small disparity. They showed V1 respond to retinal position while V2, visual field location. V3 did not respond to offset for all voxel. But a subset of voxel showing difference in left and right eye conditions, then, all areas from V2 to LO responding to visual field position, only V1 to retinal position.  

This study is striking similar to the one carried out in my lab.

772.05 Allen et al. asked whether amblyopia associated with anatomic abnormalities in white mater. The did DTI with CSE (to resolve crossing fiber issue, the result is similar to DSI). The ROIs are Pulvinar, V1, LGN and MT. They measure mean diffusivity and fractional anisotropy. Amblyopes has a greater diffusivity than the normal control in LGN-V1, V1-Mt and PLN-MT., ie., all pathways between thalamo-cortical connection, but not cortico-cortical connections. Is that More myelin degeneration for amblyopes?

Again, this study is almost identical to the research going on in my lab.

772.06  They used functionally defined White matter, which is functional MRI ROI diluted by 6mm . And then, used those area as seed for DTI. Separate white matter tracts for face and scene regions.  The normal and DP patients has the same tracts, but differ in diffusitivity.  

APCV2014 Note

Winawer

fMRI spatial summation is sublinear and pRF nonlinearity has a compressive nonlinearity.

Second order contrast (Kay et al. 2013, PLOS Computational Biology): V1 front end -> second order contrast (RMS contrast) -> decompose to low and high 2nd order contrast component -> predicted response.  The same model fits fMRI and eCog data.

Kreigeskorte

Cichy et al (2014, Nature neuroscience) showed MEG representation dynamics in which the correlation between the MEG response to different types of objects emerges at different time points.

Current computational models for vision cannot explain IT. They are all less categorical than IT.  He proposed a forward multiple convolution model to explain it.

Ziemba

V2 model: Nature image=> V1 model<==> iterative synthesis with filtered noise -> V2 model (correlation across position, orientation an scale <==> iterative synthesis with white  noise.

The measured naturalness of random noise pattern from different texture family. They found that V2 response is affected by naturalness of the texture  (between 150-250ms after stimulus onset). fMRI also showed V2 and V3 response correlated strongly  with naturalness. There is no correlation in V1.  V2 also decode different texture types better.

  

Zaidi suggested that unique colors have no psychophysical color or physiological uniqueness.  Instead, it may related to language. For instance, if one ask an observer to identify a color that is half way between unique colors A and B and between C and D, and find another color that is half way between the two just picked. The match is not the same as picking and C and B and D first. That is (A+B) + (C + D) ~= (A+C) + (B+D). The law of distribution holds only if the observer is instructed to use opponent colors, rather than unique colors.   

Koida et al. demonstrated a glaze illusion in which a white disk appears to be self-illuminated when it is surround by a luminance gradient, but not when it is surrounded by uniform gray. See below

[I am wondering how the effect will hold if the disk and the surround are isoluminance colors].

Isherwood, Schira & Spehar showed that BOLD activation in V1 is a U-shape function of 1/f slope of the (noise) image spectrum. There is an activation in the medial frontal cortex that is correlated with the aesthetic preference to the slope.

Schira constructed a software that can simulate and model BOLD response from an image.

[Schira's model has a good performance to small stimuli. Hence, it may resolve some problems we encountered in the lab on BOLD activation for small stimulus difference]

Beer & Greenlee studied the network of multiple sensory network by combining probabilistic fiber tracing and functional MRI. They seeded a voxel by fMRI, traced fibers from that voxel, and observed the terminations of the fibers. The termination was compared with the fMRI activation and showed a high correlation between the two measurements.

VSS2014 Note

• Kay et al. proposed a scheme to model fMRI BOLD response. He model includes a pRF, followed by contrast normalization and second order summation. This model can fit V1-V3 BOLD activation.

Many experiments in my lab should be able to use this model. Also, once we know the pRF, we can construct our own BOLD activation model.

• Freeman et al. used designer stimuli to explore the BOLD activation in the ventral stream. He used a lot stimuli to probe brain activation in many, many areas. Then, used simocelli's synthetic texture method to model the stimulus optimal for a brain region. Then, use that stimuli to explore the function of the target brain area.
• 24.25 Webster and Fine used a data driven approach to explore the brain function. They basically just use a huge bank of stimulus. Then get a correlation between voxels.  They then arrange the correlation matrix to find clusters of voxels with high correlations and define that as an ROI. This provides a way to detect brain regions without making assumption about the function of a particular area in advance.
•  26.11 - 26.14 many authors reported that grouping flankers with other stimuli reduces crowding effect.
• 26.16 BOLD response to target with iso-orientation collinear flankers is contrast dependent but not orthogonal flankers. This study is almost identical to mine submitted earlier this year.
• 26.573 Palmer believed the 1/3 rule of composition, or that the object should put at 1/3 position of the display for the best aesthetic value is not the whole story. This rule changes with other picture properties, such as inward bias or the position of background objects.
• 32.13 Yuvol-Greenberg & Heeger studied continuous flash suppression on the BOLD activation in early visual cortex. Their stimuli has the mask either at the same eye or different eye as the target and varied target contrast. They showed that the mask changes the gain to the target and suggested that the CFS may be just a kind of monocular noise masking.
• 32.14 Metzger used ERP to study the brain activation of a probe in the suppressed eye in binocular rivalry. They showed that the suppressed or unsuppressed probe had the same waveform in early ERP waveforms and only differed after 300ms later than probe onset. Hence, they conclude that the response to probe in binocular rivalry is due to top-down attentional process.
• 61.14 Marlow  et al. from Anderson's lab reported that not only the depth percept and  lighting condition perceived from a sinewave grating can be influenced by the shape of its contour, but also the perceived material (matt or metallic) can be affected by the contour if the grating has more black than white.
•  

SFN 2013 note

120.4 Ress & Greene used a back projection method and sweeping method to provide a model free method to estimate pRF. That is, they used a wenier kernel to estimate HRF from data, and then used back projection to decompose BOLD data into pRF. They then fit the pRF with an ellipse. They showed that pRF is only mildly elongated (aspect ration <1.3) and has weak suppression zone. Hence, the traditional Gaussian model for pRF is basically right.

120.5 The cortical reorganized itself after scotoma. Such recovery may reflect in the change of pRF in lesion projection zone. Brewer, however, used a nature scotoma to study the effect of scotoma on pRF. That is, the rod is absent in the fovea. She showed that the pRF for Scotopic was shifted compared with that for the photopic. The activity for scotopic is also weaker. 

** 120.10 Petridou at al. used a methodology derived from pRF to show that numerosity is topographically mapped in the parietal cortex. Their model is that the neural response in the parietal cortex increased with the number of elements. Hence,  They then convolved this linear function with HRF to derived the model prediction and fit the model to the data. In this way, they can calculate the "numerosity" tuning of each voxel, and thus determined the topogrphic map.

120.12 & 120.13 These two studies investigated what MVPA can do and cannot do. The MVPA is best to use when the response is relatively homogeneous in the ROI. MVPA cannot "access sparse, non-clustered neuronal responses".

259.17 Curtis Baker found three types of simple cells. One without orientation tuning but respond nonlinearly to contrast. The other two are orientation tuned.  One of the oriented cell has an expansive while the other one, compressive. They (359.15) further suggested that the non-oriented cell may be responsible for second-order processing. 

311.07 Use the voltage sensitive dye technique, Slovin et al. is able to image the population neural response with a much better spatial resolution than ordinary optical imaging methods. In Gilad & Solvin, they compared the V1 response to two lines joined together as one feature by an extra line and to two separated lines. They showed that the v1 response to one of the two separate lines are greater than the same line that belong to one big feature. Thus, they concluded that V1 encoded different features by activation strength.

The same group also compared the imaged response to a black and a white squared. They showed a response bias toward dark. They proposed a linear-nonlinear-linear model with on, off, and edge receptive fields with a weighting 0.1:0.22:1. They were able to fit the response map with such model.

358.14 the position of v1 rf shift to reflect perceived object size in corridor illusion.

555.02 Kay et al from Wandell's group showed a two-stage model to fit fmri response to a stimulus. Their model is an LNL model. The BOLD activation is proportional to the second order response.

555.21 Komatsu et al. measured color selectivity at low and high luminance in V4, AIT and PIT. They showed that the color selectivity in V4 changes with luminance while that in PIT is invariant to luminance. Thus, V4 neurons should carry both luminance and color information. Only neurons in AIT are true color selective and is invariant to luminance.

The Presidential lecture was given by Dorothy Tsao, who showed that there are six areas in the monkey's temporal cortex for face processing. Among them, the anterior areas are view point invariant while the posterior one depends on viewpoints.

VSS 2013 note

16.538 Xu et al. used bubble technique to construct adapting stimuli of partial revealed faces. After adaptation, they tested the categorization performance on a series of morphs of two facial expressions. They showed that the shift of the psychometric function was greatest following the whole face adaptation, weaker for mouth-only, followed non-mouth features. This showed that features around mouth is important for facial expression categorization.

21.22 The authors showed that there is an individual difference in the strength of the tilt illusion induced by oriented flankers. They used fMRI and DCM to find effective connectivity between foveal and peripheral regions (i.e., target and surround regions in a tilt illusion display)  in V1. They showed that there is a correlation between periphery-to-fovea connectivity and the strength of tilt illusion.

   

25.14 Starting with a replication of Hansen & Gengenfurtner (2013) noise masking data which showed a very narrowed tuned threshold elevation function, Esker tried to argue that these highly specific masking can be explained by adding one extra channel to the traditional three-channel opponent color model with a nonlinearity rather than a model of sixteen classes proposed by Gengenfurtner.

32.16 In shape-from shading, the change of luminance signals a change in 3D shape. In this study, the authors shows that the 3D percept can be eliminated when there is a color gradient coincident with the luminance gradient. The percept of the 3D shape, however, is not change by a color gradient that is inconsistent with the luminance gradient.

[CC]: My cue combination stimulus may be a better tool to explore this phenomenon. 

33.530

The experiment stimuli were blurred faces. The participants used a mouse to control which part of face will be revealed while doing expression identification task. The authors showed that the participants spend more time 

On the left-side of the face image.

[CC]:  This study still has not solved the issue whether it is left-side visual field advantage or left-side face advantage.  To resolve this issue, one do need to put the stimuli to the periphery to see whether the effect is still there.

33.536

 Here, the authors applied Dakin & Watt's bar code theory for face identity to facial expression. They found inconsistence results. The horizontal information is essential for some types f expressions but not all. 

34.24 typical likelihood theory of cue combination explains the edge location from disparity and luminance well.

35.25 Crowding occurs only when the target and the flanker are similar to each other.

35.28 Landy suggested that nonlinear pooling from a bunch of first-order filters are important to explain the discrimination performance for orientation -contrast pattern.

[cc]: His conclusion is similar to what we suggested in a recently submitted paper (and Landy is the editor). Stay tuned.

43.406 The authors used the flanker effect stimuli. They showed that the N100 (actually, near 150ms) of P1 ERP reduced when there was a difference between the target and the flanker orientation; or, when a square appears to make the target and the flanker the same group.

42.24 The authors used MVPC to compare fMRI responses to famous faces and the name of these famous people. They found that superior IPS showed identity specific responses while VWFA and FFA respond equally to different identity. 

43.421 It is suggested that person with autism tend to focus on local detail of a visual stimulus and ignore the global pattern. The authors measure the pRF of the individuals with autism. They actually got an opposite effect that the pRF of autism people in V2 and V3 are actually greater than that of normal control.   

[CC:  Actually, not only this presentation, many presentations of this section studied the local/global percept of the people with autism. Since perceptual grouping is the main focus of my lab, many of our research works should be helpful to this topic. ]

55.25 Kriegeskorte showed that there is a regular mapping of face features in OFA just like retinotopic mapping in V1.

55.27 Failure of fixation control, though has little, if any, effect, on behavior measurement, it has a statistically significant effect on N170 of ERP.  

61.26 Ebbinghaus illusion has different effect ondifferent states of Schizophrenia. Some Schizophrenia has a weaker Ebbinghaus illusion, while the other has a stronger illusion than the normal control.

OSA Fall Vision Meeting Note

Zickler used apertures of different chromaticity to take pictures. Since aperture size determines the depth of focus, an image taken with a camera with different apertures will have different depths of focus for different color. This information allows the researchers to estimate the depth of different objects in the image.

Bank investigated different 3D display technologies. He focused on depth distortion. Left-right lateral movement on a stereo display appears forward and backward movement in depth. It may be an artifact caused by the current time-sequential display technique. He used a rotating wheel stimuli (disks of different disparity rotating about the fixation point) to measure the nulling depth and estimate the distortion.

Read showed that much disparity information n the retina is lost when disparity is encoded in V1. She used gratings modulated in depth (cf. Kontsevich & Tyler). V1 has on-off regions for luminance but same-disparity tuning for the whole receptive field. Therefore, some areas must integrate V1 cells to see gratings modulated in depth. The cell response for sinewave grating depends on the depth modulation, but not square wave. It is because that larger disparities is encoded by larger receptive fields. To design video codec, it is better to have high resolution monocular images with low resolution depth maps. The depth percept has even less information than V1. In addition, it is easier to detect horizontal grating than vertical grating at lower spatial frequencies.  She also measured the reverse correlation threshold for joined disparity and motion grating.  The reverse correlation threshold allows her o estimate the receptive size for disparity and motion. The RF tuned t different combinations of motion and disparity. It is unlikely to be MT neurons. So, Information is read out from V1, but it is not clear who read it.

Tsirlin from Wilcox 's lab talked about cross talk on perceived depth in 3D displays. Crosstalk=(leakage/signal)*100%.  Crosstalk reduce image quality, quality of depth and induce discomfort. She asked observer to estimate the depth of two bars with and without cross talk. Increase in cross talk reduced perceived depth. Significant reduction started at 4-8% cross talk.

Kazimi from 3D film innovation consortium taled about Hazardous stereography.  UK constraint disparity to be 4% positive (object in front of the background)  and 2% negative as safe zone. Depth setting changes with display size. 

Rudd. Lightness contrast with two contexts. The target disk is surrounded by an immediate surround and then on a large background. The radius of the surround has an effect on the perceived lightness of the target.  The matched target luminance peaks at certain physical luminance, suggesting a contrast gain control mechanism.