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

2022 ◽  
Vol 18 (1) ◽  
pp. e1009739
Author(s):  
Nathan C. L. Kong ◽  
Eshed Margalit ◽  
Justin L. Gardner ◽  
Anthony M. Norcia
Keyword(s):  
Spatial Frequency ◽  
Power Law ◽  
Frequency Tuning ◽  
Neural Responses ◽  
Visual Stream ◽  
Frequency Distributions ◽  
Machine Perception ◽  
Spatial Frequencies ◽  
Power Law Exponent ◽  
Stimulus Features

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.

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

2021 ◽  
Author(s):  
Nathan C. L. Kong ◽  
Eshed Margalit ◽  
Justin L. Gardner ◽  
Anthony M. Norcia
Keyword(s):  
Spatial Frequency ◽  
Power Law ◽  
Frequency Tuning ◽  
Neural Responses ◽  
Visual Stream ◽  
Frequency Distributions ◽  
Machine Perception ◽  
Spatial Frequencies ◽  
Power Law Exponent ◽  
Stimulus Features

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.

Spatial-frequency and orientation tuning in psychophysical end-stopping

1998 ◽  
Vol 15 (4) ◽  
pp. 585-595 ◽  
Author(s):  
CONG YU ◽  
DENNIS M. LEVI
Keyword(s):  
Spatial Frequency ◽  
Frequency Tuning ◽  
Spatial Filter ◽  
Orientation Tuning ◽  
Maximal Strength ◽  
Facilitation Effect ◽  
Spatial Filters ◽  
Spatial Frequency Tuning ◽  
Spatial Frequencies ◽  
Wide Range

A psychophysical analog to cortical receptive-field end-stopping has been demonstrated previously in spatial filters tuned to a wide range of spatial frequencies (Yu & Levi, 1997a). The current study investigated tuning characteristics in psychophysical spatial filter end-stopping. When a D6 (the sixth derivative of a Gaussian) target is masked by a center mask (placed in the putative spatial filter center), two end-zone masks (placed in the filter end-zones) reduce thresholds. This “end-stopping” effect (the reduction of masking induced by end-zone masks) was measured at various spatial frequencies and orientations of end-zone masks. End-stopping reached its maximal strength when the spatial frequency and/or orientation of the end-zone masks matched the spatial frequency and/or orientation of the target and center mask, showing spatial-frequency tuning and orientation tuning. The bandwidths of spatial-frequency and orientation tuning functions decreased with increasing target spatial frequency. At larger orientation differences, however, end-zone masks induced a secondary facilitation effect, which was maximal when the spatial frequency of end-zone masks equated the target spatial frequency. This facilitation effect might be related to certain types of contour and texture perception, such as perceptual pop-out.

scholarly journals The Order of Information Transfer into Short- Term Memory from Visual Pathways with Different Spatial-Frequency Tunings

2020 ◽  
Vol 13 (2) ◽  
pp. 72-89
Author(s):  
D.S. Alekseeva ◽  
V.V. Babenko ◽  
D.V. Yavna
Keyword(s):  
Spatial Frequency ◽  
Information Transfer ◽  
Short Term Memory ◽  
Frequency Tuning ◽  
Visual Pathways ◽  
Input Image ◽  
Target Face ◽  
Short Term ◽  
Term Memory ◽  
Spatial Frequencies

Visual perceptual representations are formed from the results of processing the input image in parallel pathways with different spatial-frequency tunings. It is known that these representations are created gradually, starting from low spatial frequencies. However, the order of information transfer from the perceptual representation to short-term memory has not yet been determined. The purpose of our study is to determine the principle of entering information of different spatial frequencies in the short-term memory. We used the task of unfamiliar faces matching. Digitized photographs of faces were filtered by six filters with a frequency tuning step of 1 octave. These filters reproduced the spatial-frequency characteristics of the human visual pathways. In the experiment, the target face was shown first. Its duration was variable and limited by a mask. Then four test faces were presented. Their presentation was not limited in time. The observer had to determine the face that corresponds to the target one. The dependence of the accuracy of the solution of the task on the target face duration for different ranges of spatial frequencies was determined. When the target stimuli were unfiltered (broadband) faces, the filtered faces were the test ones, and vice versa. It was found that the short-term memory gets information about an unfamiliar face in a certain order, starting from the medium spatial frequencies, and this sequence does not depend on the processing method (holistic or featural).

Temporal and Spatial Frequency Tuning of the Flicker Motion Aftereffect

Perception ◽  
1996 ◽  
Vol 25 (1_suppl) ◽  
pp. 12-12
Author(s):  
P J Bex ◽  
F A J Verstraten ◽  
I Mareschal
Keyword(s):  
Spatial Frequency ◽  
Frequency Tuning ◽  
Temporal Frequency ◽  
Motion Aftereffect ◽  
Sine Wave ◽  
Test Grating ◽  
Spatial Frequency Tuning ◽  
Spatial Frequencies ◽  
Low Pass ◽  
Sine Wave Gratings

The motion aftereffect (MAE) was used to study the temporal-frequency and spatial-frequency selectivity of the visual system at suprathreshold contrasts. Observers adapted to drifting sine-wave gratings of a range of spatial and temporal frequencies. The magnitude of the MAE induced by the adaptation was measured with counterphasing test gratings of a variety of spatial and temporal frequencies. Independently of the spatial or temporal frequency of the adapting grating, the largest MAE was found with slowly counterphasing test gratings (∼0.125 – 0.25 Hz). For slowly counterphasing test gratings (<∼2 Hz), the largest MAEs were found when the test grating was of similar spatial frequency to that of the adapting grating, even at very low spatial frequencies (0.125 cycle deg−1). However, such narrow spatial frequency tuning was lost when the temporal frequency of the test grating was increased. The data suggest that MAEs are dominated by a single, low-pass temporal-frequency mechanism and by a series of band-pass spatial-frequency mechanisms at low temporal frequencies. At higher test temporal frequencies, the loss of spatial-frequency tuning implicates separate mechanisms with broader spatial frequency tuning.

How Reliable is the Pattern Adaptation Technique? A Modeling Study

2009 ◽  
Vol 102 (4) ◽  
pp. 2245-2252 ◽  
Author(s):  
Jay Hegdé
Keyword(s):  
Spatial Frequency ◽  
Frequency Tuning ◽  
Spatial Vision ◽  
Neural Mechanisms ◽  
System Level ◽  
Sinusoidal Grating ◽  
Tuning Parameters ◽  
Neuronal Ensemble ◽  
Spatial Frequency Tuning ◽  
Spatial Frequencies

Upon prolonged viewing of a sinusoidal grating, the visual system is selectively desensitized to the spatial frequency of the grating, while the sensitivity to other spatial frequencies remains largely unaffected. This technique, known as pattern adaptation, has been so central to the psychophysical study of the mechanisms of spatial vision that it is sometimes referred to as the “psychologist's microelectrode.” While this approach implicitly assumes that the adaptation behavior of the system is diagnostic of the corresponding underlying neural mechanisms, this assumption has never been explicitly tested. We tested this assumption using adaptation bandwidth, or the range of spatial frequencies affected by adaptation, as a representative measure of adaptation. We constructed an intentionally simple neuronal ensemble model of spatial frequency processing and examined the extent to which the adaptation bandwidth at the system level reflected the bandwidth at the neuronal level. We find that the adaptation bandwidth could vary widely even when all spatial frequency tuning parameters were held constant. Conversely, different spatial frequency tuning parameters were able to elicit similar adaptation bandwidths from the neuronal ensemble. Thus, the tuning properties of the underlying units did not reliably reflect the adaptation bandwidth at the system level, and vice versa. Furthermore, depending on the noisiness of adaptation at the neural level, the same neuronal ensemble was able to produce selective or nonselective adaptation at the system level, indicating that a lack of selective adaptation at the system level cannot be taken to mean a lack of tuned mechanisms at the neural level. Together, our results indicate that pattern adaptation cannot be used to reliably estimate the tuning properties of the underlying units, and imply, more generally, that pattern adaptation is not a reliable tool for studying the neural mechanisms of pattern analysis.

scholarly journals Dynamics of spatial frequency tuning in mouse visual cortex

2012 ◽  
Vol 107 (11) ◽  
pp. 2937-2949 ◽  
Author(s):  
Samme Vreysen ◽  
Bin Zhang ◽  
Yuzo M. Chino ◽  
Lutgarde Arckens ◽  
Gert Van den Bergh
Keyword(s):  
Visual Cortex ◽  
Spatial Frequency ◽  
Visual Information ◽  
Frequency Tuning ◽  
Neuronal Response ◽  
Correlation Technique ◽  
Low Frequencies ◽  
Temporal Shift ◽  
Spatial Frequency Tuning ◽  
Spatial Frequencies

Neuronal spatial frequency tuning in primary visual cortex (V1) substantially changes over time. In both primates and cats, a shift of the neuron's preferred spatial frequency has been observed from low frequencies early in the response to higher frequencies later in the response. In most cases, this shift is accompanied by a decreased tuning bandwidth. Recently, the mouse has gained attention as a suitable animal model to study the basic mechanisms of visual information processing, demonstrating similarities in basic neuronal response properties between rodents and highly visual mammals. Here we report the results of extracellular single-unit recordings in the anesthetized mouse where we analyzed the dynamics of spatial frequency tuning in V1 and the lateromedial area LM within the lateral extrastriate area V2L. We used a reverse-correlation technique to demonstrate that, as in monkeys and cats, the preferred spatial frequency of mouse V1 neurons shifted from low to higher frequencies later in the response. However, this was not correlated with a clear selectivity increase or enhanced suppression of responses to low spatial frequencies. These results suggest that the neuronal connections responsible for the temporal shift in spatial frequency tuning may considerably differ between mice and monkeys.

Monocular and binocular response properties of cells in the striate-recipient zone of the cat's lateral posterior-pulvinar complex

1989 ◽  
Vol 62 (2) ◽  
pp. 544-557 ◽  
Author(s):  
C. Casanova ◽  
R. D. Freeman ◽  
J. P. Nordmann
Keyword(s):  
Spatial Frequency ◽  
Frequency Tuning ◽  
Single Cells ◽  
Temporal Frequency ◽  
Orientation Tuning ◽  
Low Frequencies ◽  
Response Properties ◽  
Spatial Frequencies ◽  
Lateral Posterior ◽  
The Mean

1. We have studied response properties of single cells in the striate-recipient zone of the cat's lateral posterior-pulvinar (LP-P) complex. This zone is in the lateral section of the lateral posterior nucleus (LP1). Our purpose was to determine basic response characteristics of these cells and to investigate the possibility that the LP-P complex is a center of integration that is dominated by input from visual cortex. 2. The majority (72%) of cells in the striate-recipient zone respond to drifting sinusoidal gratings with unmodulated discharge. 3. Cells in the LP1 are selective to the orientation of gratings, and tuning functions have a mean bandwidth of 31 degrees. More than one-half of these units are direction-selective. The preferred orientation and the tuning widths for the two eyes are generally well matched. However, a few cells exhibited the interesting property of opposite preferred directions for the two eyes. Orientation tuning for a small group of cells was different for the mean discharge and first harmonic components, suggesting a convergence from different inputs to these cells. 4. Two-thirds of LP1 cells are tuned to low spatial frequencies (less than 0.5 c/deg). The tuning is broad with a mean bandwidth of 2.2 octaves. The remaining one-third of the units are low-pass because they show no attenuation of their responses to low spatial frequencies. Both eyes exhibit the same spatial frequency preference and the same spatial frequency tuning. There is a high correlation between spatial frequency and orientation selectivities. 5. All cells tested are tuned for temporal frequency with a sharp attenuation for low frequencies. The optimal values range between 4 and 8 Hz, and the mean bandwidth is 2.2 octaves. 6. Cells in LP1 are mostly binocular. When monocular, cells are almost always contralaterally driven. Dichoptic presentation of gratings reveals the presence of strong binocular interaction. In almost all cases, these interactions are phase specific. The cell's discharge is facilitated at particular phases and inhibited at phases 180 degrees away. These binocular interactions are orientation dependent. 7. Twenty-five percent of the cells with phase-specific binocular facilitation appear to be monocular when each eye is tested separately. For three cells, we observed a non-phase-specific inhibitory effect of the silent eye. 8. Our findings indicate that LP1 cells form a relatively homogeneous group, suggesting a high degree of integration of multiple cortical inputs.(ABSTRACT TRUNCATED AT 400 WORDS)

Spatial Frequency-Specific Contrast Adaptation Originates in the Primary Visual Cortex

2007 ◽  
Vol 98 (1) ◽  
pp. 187-195 ◽  
Author(s):  
Thang Duong ◽  
Ralph D. Freeman
Keyword(s):  
Visual Cortex ◽  
Spatial Frequency ◽  
Frequency Tuning ◽  
Adaptation Effect ◽  
Tuning Curves ◽  
Contrast Adaptation ◽  
Spatial Frequencies ◽  
Neurophysiological Studies ◽  
High Contrast Grating ◽  
Adaptation Effects

Adaptation to a high-contrast grating stimulus causes reduced sensitivity to subsequent presentation of a visual stimulus with similar spatial characteristics. This behavioral finding has been attributed by neurophysiological studies to processes within the visual cortex. However, some evidence indicates that contrast adaptation phenomena are also found in early visual pathways. Adaptation effects have been reported in retina and lateral geniculation nucleus (LGN). It is possible that these early pathways could be the physiological origin of the cortical adaptation effect. To study this, we recorded from single neurons in the cat's LGN. We find that contrast adaptation in the LGN, unlike that in the visual cortex, is not spatial frequency specific, i.e., adaptation effects apply to a broad range of spatial frequencies. In addition, aside from the amplitude attenuation, the shape of spatial frequency tuning curves of LGN cells is not affected by contrast adaptation. Again, these findings are unlike those found for cells in the visual cortex. Together, these results demonstrate that pattern specific contrast adaptation is a cortical process.

Cooperative Representation of Visual Borders

Perception ◽  
1992 ◽  
Vol 21 (2) ◽  
pp. 185-193 ◽  
Author(s):  
Geoffrey W Stuart ◽  
Terence R J Bossomaier
Keyword(s):  
Spatial Frequency ◽  
Spatial Scales ◽  
Receptive Fields ◽  
Frequency Content ◽  
Cortical Cells ◽  
Firing Rates ◽  
Spatial Frequencies ◽  
Cooperative Coding ◽  
Stimulus Features ◽  
Visual Cortical

Recently it has been reported that the visual cortical cells which are engaged in cooperative coding of global stimulus features, display synchrony in their firing rates when both are stimulated. Alternative models identify global stimulus features with the coarse spatial scales of the image. Versions of the Munsterberg or Café Wall illusions which differ in their low spatial frequency content were used to show that in all cases it was the high spatial frequencies in the image which determined the strength and direction of these illusions. Since cells responsive to high spatial frequencies have small receptive fields, cooperative coding must be involved in the representation of long borders in the image.

scholarly journals Neurons in cat V1 show significant clustering by degree of tuning

2015 ◽  
Vol 113 (7) ◽  
pp. 2555-2581 ◽  
Author(s):  
Avi J. Ziskind ◽  
Al A. Emondi ◽  
Andrei V. Kurgansky ◽  
Sergei P. Rebrik ◽  
Kenneth D. Miller
Keyword(s):  
Spatial Frequency ◽  
Frequency Tuning ◽  
Direction Selectivity ◽  
Preferred Orientation ◽  
Orientation Tuning ◽  
Complex Cells ◽  
Spatial Frequency Tuning ◽  
Spatial Frequencies ◽  
Tuning Width ◽  
Simple And Complex Cells

Neighboring neurons in cat primary visual cortex (V1) have similar preferred orientation, direction, and spatial frequency. How diverse is their degree of tuning for these properties? To address this, we used single-tetrode recordings to simultaneously isolate multiple cells at single recording sites and record their responses to flashed and drifting gratings of multiple orientations, spatial frequencies, and, for drifting gratings, directions. Orientation tuning width, spatial frequency tuning width, and direction selectivity index (DSI) all showed significant clustering: pairs of neurons recorded at a single site were significantly more similar in each of these properties than pairs of neurons from different recording sites. The strength of the clustering was generally modest. The percent decrease in the median difference between pairs from the same site, relative to pairs from different sites, was as follows: for different measures of orientation tuning width, 29–35% (drifting gratings) or 15–25% (flashed gratings); for DSI, 24%; and for spatial frequency tuning width measured in octaves, 8% (drifting gratings). The clusterings of all of these measures were much weaker than for preferred orientation (68% decrease) but comparable to that seen for preferred spatial frequency in response to drifting gratings (26%). For the above properties, little difference in clustering was seen between simple and complex cells. In studies of spatial frequency tuning to flashed gratings, strong clustering was seen among simple-cell pairs for tuning width (70% decrease) and preferred frequency (71% decrease), whereas no clustering was seen for simple-complex or complex-complex cell pairs.

Sign in / Sign up

Export Citation Format

Share Document