Increasing neural network robustness improves match to macaque V1 eigenspectrum, spatial frequency preference and predictivity

Task-optimized convolutional neural networks (CNNs) show striking similarities to the ventral visual stream. However, human-imperceptible image perturbations can cause a CNN to make incorrect predictions. Here we provide insight into this brittleness by investigating the representations of models that are either robust or not robust to image perturbations. Theory suggests that the robustness of a system to these perturbations could be related to the power law exponent of the eigenspectrum of its set of neural responses, where power law exponents closer to and larger than one would indicate a system that is less susceptible to input perturbations. We show that neural responses in mouse and macaque primary visual cortex (V1) obey the predictions of this theory, where their eigenspectra have power law exponents of at least one. We also find that the eigenspectra of model representations decay slowly relative to those observed in neurophysiology and that robust models have eigenspectra that decay slightly faster and have higher power law exponents than those of non-robust models. The slow decay of the eigenspectra suggests that substantial variance in the model responses is related to the encoding of fine stimulus features. We therefore investigated the spatial frequency tuning of artificial neurons and found that a large proportion of them preferred high spatial frequencies and that robust models had preferred spatial frequency distributions more aligned with the measured spatial frequency distribution of macaque V1 cells. Furthermore, robust models were quantitatively better models of V1 than non-robust models. Our results are consistent with other findings that there is a misalignment between human and machine perception. They also suggest that it may be useful to penalize slow-decaying eigenspectra or to bias models to extract features of lower spatial frequencies during task-optimization in order to improve robustness and V1 neural response predictivity.

Potsdam Eye-Movement Corpus for Scene Memorization and Search with Color and Spatial-Frequency Filtering

The availability of large eye-movement corpora has become increasingly important over the past years. In scene viewing, scan-path analyses of time-ordered fixations, for example, allow for investigating individual differences in spatial correlations between fixation locations, or for predicting individual viewing behavior in the context of computational models. However, time-dependent analyses require many fixations per scene, and only few large eye-movement corpora are publicly available. This manuscript presents a new corpus with eye-movement data from two hundred participants. Viewers memorized or searched either color or grayscale scenes while high or low spatial frequencies were filtered in central or peripheral vision. Our database provides the scenes from the experiment with corresponding object annotations, preprocessed eye-movement data, and heatmaps and fixation clusters based on empirical fixation locations. Besides time-dependent analyses, the corpus data allow for investigating questions that have received little attention in scene-viewing research so far: (i) eye-movement behavior under different task instructions, (ii) the importance of color and spatial frequencies when performing these tasks, and (iii) the individual roles and interaction of central and peripheral vision during scene viewing. Furthermore, the corpus allows for validation of computational models of attention and eye-movement control, and finally, analyses on an object- or cluster-based level.

Improved playback uniformity of random-phase free holograms by pixel separation method

Random-phase free computer-generated holograms offer excellent quality of virtually noise-free playback of low-frequency images, but have limited efficiency in the case of highly contrast binary images with dominant high spatial frequencies. Introduction of weak random phase allows the partial suppression of this problem, but causes strong noise in the outcome. Here we present the influence of pixel separation technique on the uniformity of far field reconstructions from such random-phase free holograms. We show the improved image quality with no additional speckle noise. Full Text: PDF ReferencesJ.W. Goodman, Roberts and Company (2005). DirectLink R.W. Gerchberg, W.O. Saxton, "A practical algorithm for the determination of phase from image and diffraction plane pictures", Optik 35, 237 (1972). DirectLink M. Makowski, "Minimized speckle noise in lens-less holographic projection by pixel separation", Opt. Express 21, 29205 (2013). CrossRef I. Ducin, T. Shimobaba, M. Makowski, K. Kakarenko, A. Kowalczyk, Jaroslaw Suszek, M. Bieda, A. Kolodziejczyk, M. Sypek, "Holographic projection of images with step-less zoom and noise suppression by pixel separation", Opt. Comm. 340, 131 (2015). CrossRef T. Shimobaba, T. Ito, "Random phase-free computer-generated hologram", Opt. Express 23, 9549 (2015). CrossRef T. Shimobaba, T. Kakue, Y. Endo, R. Hirayama, D. Hiyama, S. Hasegawa, Y. Nagahama, M. Sano, M. Oikawa, T. Sugie, T. Ito, "Random phase-free kinoform for large objects", Opt. Express 23, 17269 (2015). CrossRef M. Sypek, "Light propagation in the Fresnel region. New numerical approach", Opt. Comm. 116, 43 (1995). CrossRef K. Matsushima, T. Shimobaba, "Band-Limited Angular Spectrum Method for Numerical Simulation of Free-Space Propagation in Far and Near Fields", Opt. Express 17, 19662 (2009). CrossRef

Binocular Summation Is Intact in Intermittent Exotropia After Surgery

Purpose: To determine binocular summation of surgically treated intermittent exotropia (IXT) patients by measuring the contrast threshold.Methods: We recruited 38 surgically treated IXT patients aged 8–24 years and 20 age-matched healthy controls. All participants had normal or corrected-to-normal visual acuity (Snellen ≥ 20/20) in both eyes. The IXT patients had undergone the surgery at least a year prior to the study. Twenty-one of them obtained good alignment and 17 experienced a recurrence of exotropia. We measured the observers' monocular and binocular contrast sensitivities (CS) at six spatial frequencies (1.5, 3, 6, 12, 18, 24 cycles/degree) as an index of visual information processing at the threshold level. Binocular summation was evaluated against a baseline model of simple probability summation based on the CS at each spatial frequency and the area under the log contrast sensitivity function (AULCSF).Results: The exo-deviation of IXTs with good alignment was −6.38 ± 3.61 prism diopters (pd) at 33 cm and −5.14 ± 4.07 pd at 5 m. For the patients with recurrence, it was −23.47 ± 5.53 pd and −21.12 ± 4.28 pd, respectively. There was no significant difference in the binocular summation ratio (BSR) between the surgically treated IXT patients, including those with good alignment and recurrence, and normal controls at each spatial frequency [F(2,55) = 0.416, P = 0.662] and AULCSF [F(2,55) = 0.469, P = 0.628]. In addition, the BSR was not associated with stereopsis (r = −0.151, P = 0.365).Conclusion: Our findings of normal contrast sensitivity binocular summation ratio in IXT after surgical treatment suggest that the ability of the visual cortex in processing binocular information is intact at the contrast threshold level.

Relationship Between Facial Areas With the Greatest Increase in Non-local Contrast and Gaze Fixations in Recognizing Emotional Expressions

The aim of our study was to analyze gaze fixations in recognizing facial emotional expressions in comparison with o the spatial distribution of the areas with the greatest increase in the total (nonlocal) luminance contrast. It is hypothesized that the most informative areas of the image that getting more of the observer’s attention are the areas with the greatest increase in nonlocal contrast. The study involved 100 university students aged 19-21 with normal vision. 490 full-face photo images were used as stimuli. The images displayed faces of 6 basic emotions (Ekman’s Big Six) as well as neutral (emotionless) expressions. Observer’s eye movements were recorded while they were the recognizing expressions of the shown faces. Then, using a developed software, the areas with the highest (max), lowest (min), and intermediate (med) increases in the total contrast in comparison with the surroundings were identified in the stimulus images at different spatial frequencies. Comparative analysis of the gaze maps with the maps of the areas with min, med, and max increases in the total contrast showed that the gaze fixations in facial emotion classification tasks significantly coincide with the areas characterized by the greatest increase in nonlocal contrast. Obtained results indicate that facial image areas with the greatest increase in the total contrast, which preattentively detected by second-order visual mechanisms, can be the prime targets of the attention.

Inferring the DNA replication origin density landscape in human cells from replication timing and fork directionality data

The determinants of the locations and firing times of the multiple replication origins are still elusive in human and other metazoan organisms. Experiments can independently profile mean replication timing (MRT) and replication fork directionality (RFD) genome-wide. In the hypothesis of a constant replication fork speed, MRT and RFD are related to each other by an analytical formula so are a priori equivalent. However, we show here that experimental noises result in MRT and RFD profiles containing information at different spatial frequencies. We further demonstrate that one can compute an origin density landscape that, when inserted in an appropriate simulation framework, jointly predicts experimental MRT and RFD profiles with an unprecedented precision. We also extract an analytical formula linking intrinsic origin efficiency with observed origin efficiency and MRT. We then compare the computed origin density landscape with experimental distributions of potential origins (ORC, MCM) or actual initiation events (Bubble-seq, SNS-seq, OK-seq). The results indicate that MRT and RFD data are highly consistent with each other, that our simple model suffices to capture the replication dynamics during S phase given an appropriate initiation probability landscape, but that the density of potential origins is not the sole determinant of this landscape.

Spatial tuning of translational optic flow responses in hawkmoths of varying body size

To safely navigate their environment, flying insects rely on visual cues, such as optic flow. Which cues insects can extract from their environment depends closely on the spatial and temporal response properties of their visual system. These in turn can vary between individuals that differ in body size. How optic flow-based flight control depends on the spatial structure of visual cues, and how this relationship scales with body size, has previously been investigated in insects with apposition compound eyes. Here, we characterised the visual flight control response limits and their relationship to body size in an insect with superposition compound eyes: the hummingbird hawkmoth Macroglossum stellatarum. We used the hawkmoths’ centring response in a flight tunnel as a readout for their reception of translational optic flow stimuli of different spatial frequencies. We show that their responses cut off at different spatial frequencies when translational optic flow was presented on either one, or both tunnel walls. Combined with differences in flight speed, this suggests that their flight control was primarily limited by their temporal rather than spatial resolution. We also observed strong individual differences in flight performance, but no correlation between the spatial response cutoffs and body or eye size.

Increasing pupil size is associated with improved detection performance in the periphery

Visible light enters our body via the pupil. By changing its size, the pupil shapes visual input. Small apertures increase the resolution of high spatial frequencies, thus allowing discrimination of fine details. Large apertures, in contrast, provide a better signal-to-noise ratio, because more light can enter the eye. This should lead to better detection performance of peripheral stimuli. Experiment 1 shows that the effect can reliably be demonstrated even in a less controlled online setting. In Experiment 2, pupil size was measured in a laboratory using an eye tracker. The findings replicate findings showing that large pupils provide an advantage for peripheral detection of faint stimuli. Moreover, not only pupil size during information intake in the current trial n, but also its interaction with pupil size preceding information intake, i.e., in trial n-1, predicted performance. This suggests that in addition to absolute pupil size, the extent of pupillary change provides a mechanism to modulate perceptual functions. The results are discussed in terms of low-level sensory as well as higher-level arousal-driven changes in stimulus processing.

Comparison of CSV-1000 and Metrovision contrast sensitivity tests in normal eyes

Background: Measuring contrast sensitivity (CS) allows a better understanding of the visual performance of the human eye. This study aimed to examine the correlation and agreement between the results of two sinewave grating-based CS measurement methods, Metrovision and CSV-1000, in normal eyes. Methods: This cross-sectional, comparative study was performed between December 2018 and April 2019, at an optometry clinic. Subjects underwent comprehensive ocular examinations, which included pupil reflexes, subjective refraction, external eye examinations, smooth pursuit eye movement assessment, the cover–uncover test, and detailed slit-lamp examination of the anterior and posterior segments. Metrovision and CSV-1000 were employed to assess CS under photopic conditions. The correlation and agreement of the results of the two tests were evaluated. Results: CS was measured for 104 normal eyes for 3, 6, 12, and 18 cycles per degree (cpd) spatial frequencies (participants’ mean age ± standard deviation: 37.3 ± 26.4 years). The CSV-1000 measurements were significantly higher for the 3 and 6 cpd spatial frequencies (both P = 0.01); however, at higher spatial frequencies, CS scores were similar. The highest and lowest differences between the two tests were recorded for the 3 cpd spatial frequency and 18 cpd spatial frequency, respectively. Except for the 3 cpd spatial frequency, in both eyes, the findings correlated significantly between the CSV-1000 and Metrovision (P < 0.05). The narrowest and widest limits of agreement between the two tests were found for the 12 and 3 cpd spatial frequencies, respectively. Conclusions: The CSV-1000 method estimated CS higher than the Metrovision method, mostly at lower spatial frequencies. Furthermore, the agreement between the two methods was greater at higher spatial frequencies than at lower frequencies. This should be kept in mind when using the two methods interchangeably in visual screening.