On 2016 Aug 17, Lydia Maniatis commented:

Short version:

The authors say: "We hypothesize that how our stimuli are perceived depends on spatial frequency. Our hypothesis is confirmed. (P.S. The effect is size invariant.)"

Translation: "We hypothesize that how our stimuli are perceived depends on spatial frequency. How our stimuli are perceived DOES NOT depend on spatial frequency. Our spatial frequency hypothesis is falsified."

a + (-a) = 0

(The discrete addition of the phrase "at one size" into the abstract, as though this did not upend the entire position, is really quite extraordinary.)

On 2016 Apr 11, Lydia Maniatis commented:

It is worth asking why, on the one hand, Solomon and Pelli (1994) seem to be asking questions about initial steps in vision - asking about "visual filters," and, on the other, referring, not to "perception" of letters, but to "identification" or "recognition." In order to identify or recognise, we have to first perceive, but this step is skipped over not only by these researchers but by those contemporary investigators who consider themselves "psychophysicists" in general.

The reason is this: These investigators insist on treating perception as a problem of detection, such that they can recycle the signal detection models first developed by radar researchers.

Thus, they assume that identification/recognition would be perfect if it were not for “internal noise” in the system, which produces uncertainty in the outcome. This noise is the random firing of neurons, base rates etc. This claim about internal noise and perception has never been tested, nor (to my knowledge) have these researchers proposed a way to test or measure the claim, but it is assumed.

It is also assumed that we can add “external noise” to a stimulus and thus counterbalance the “internal noise.” This “external noise” can take the form of variance of the stimulus pool, splotches superimposed on the stimulus (as here), etc. Anything that can be described as “unpredictable.”

If we do this, then there is an equation that we can use that allows us to measure “performance” regardless of the task, as long as we have a value to plug in for our “noise” term.

If we can assume that the “external noise” is much bigger than the “internal noise” then we can use this equation whose output is labelled “E” for efficiency, by plugging in our “external noise” value.

Clear so far?

What is wrong with this story is also the reason researchers adopting it don't refer to “perception.”

The stimulus, which consists of retinal photoreceptors excited by disconnected photons pouring in from the environment does not contain “letters” or any other objects to be “detected.” Only photons are detected by the system. It is the response of the nervous system that organizes these points, based on rules (instantiated in the system) that have been selected for over evolutionary time to produce a good enough match between the objects out there and the objects we perceive. We can understand the importance of sophisticated rules if we consider how easy it is for humans to read the letters of “captchas” used by websites to ensure you're not a robot, but which programmers have a hard time beating. (This is actually a crude and somewhat misleading example, because even seeing the entire captcha without being able to recognize the letters it “contains” is a feat of perceptual organization). So the visual system creates a percept by adding structure to structureless data, and the number of these potential structurings is infinite. The limiting factors are the rules.

So to see a letter, even as a physical shape, implies an enormously complicated process of perceptual organization, based on inferences that constitute the output of the system. What we “identify” or “recognize” is the product of these processes. We are not “detecting” something that is there a priori, in the stimulus, the disconnected photons. Seeing splotches, or any other feature of the percept, is also the result of the same organizing processes. There's no principle justifying calling part of the percept “noise” and another part “signal” if what we are interested in are basic perceptual processes. It is relatedly a problem to overlook that some kinds of “noise” will, for organizational reasons, interfere more with the “signal” that we have in mind than other kinds, something that the makers of “captchas” know well.

But all of these facts are silently overlooked by investigators who choose to treat perception as a detection problem involving imaginary spatial filters and arbitrarily defined “noise”. Which is truly extraordinary.

On 2016 Mar 12, Lydia Maniatis commented:

Continuation of previous:

Translation: Forget everything we've assumed so far about spatial filters. We don't have special detectors sensitive to the frequency of light bands falling across the retina! Visual angle doesn't matter. The visual system has special detectors for detecting the number of light/dark cycles per (familiar) "letter" - (about 3), actually a range of cycles, since some letters are quite wide (e.g. m's) and others quite narrow (e.g. l's). The processes implicit in organizing points of light into letter shapes so that we can recruit the relevant letter-sensitive channels are not relevant. Our data plus assumptions prove it.

I think this front-page Nature article has zero credible content or face-validity.

On 2016 Mar 12, Lydia Maniatis commented:

The claims the authors make in this publication lack both empirical and logical support. To make what is perhaps the most clear point, the very tenuous assumptions underlying the project are themselves contradicted by the authors own statements. This is represented as a refinement of the fundamental assumption, but in fact it pulls the rug out from under it. Each step of the way, definite assertions are made which citations provided do not support.

The train of thought (with comments) is this:

Authors: “We hear periodic sounds, or tones, by means of parallel auditory filters, each tuned to a band of temporal frequency (Fletcher, 1940)”

From MIT online Encyclopedia of Cognitive Sciences: To account for the obtained data, Fletcher proposed a “critical band” that likened the ear to a bandpass filter (or, to encompass the entire frequency range, a set of contiguous, overlapping bandpass filters). The proposed “auditory filter” is a theoretical construct that reflects frequency selectivity present in the auditory system...”

Translation: The auditory system exhibits frequency selectivity. We could find no more convincing citations with respect to auditory filters between 1940 and 1994.

Authors: “...and we see periodic patterns, or gratings, by means of parallel visual filters (Campbell and Robson, 1968).”

From Robson and Campbell, 1968: “Thus, it seems that we cannot satisfactorily model the over-all visual system by a simple peak detector following a spatial filter...As a modification of this theory, we may assume...Thus, we may suppose that the visual system behaves not as a single detector mechanism preceded by a single broadband spatial filter but as a number of independent detector mechanisms each preceded by a relatively narrowband filter 'tuned' to a different frequency...Such a model could account for our findings...”

Translation: Robson and Cambell's (1968) results did not bear out their original predictions, and they speculated as to possible alternative accounts. Between 1968 and 1994, no firmer conclusions were apparently available.

Note also that the analogy between frequency of pressure waves and the “frequency of periodic patterns (gratings)” is very loose. What is being claimed, specifically, is that there is a class of visual neurons that are excited by stimuli with a specific, sine-wave-type luminance structure across a region of space (more specifically, across a region of the retina), and that they are sensitive to the frequency of these spatial luminance alterations. But frequency of pressure waves is a fundamental feature of the energy that interacts with our sensory system to initiate perception. Grating patterns, on the other hand, are just one of an infinite number of possible patterns of light impinging on our retinas. If assumptions about vision are to rest on analogy, then frequency of light waves would seem more appropriate.

Authors: “Do these visual filters participate in everyday tasks such as reading and object recognition?”

Translation: We assume the existence of special, “low-level” mechanisms for detecting grating patterns of light on the retina. (Campbell and Robson, 1968 specified visual angle, i.e. extent on the retina, in defining their patterns). We wonder if these exist exclusively to detect grating patterns, or whether they are engaged in seeing other things as well. We're particularly interested in the perception of familiar objects that we can name upon seeing. The possibility that the visual system may possess “low-level” neurons interested only in ecologically irrelevant grating patterns but not in objects seems to us to be not only credible but testable and worth testing...

Authors: “After all, grating visibility only requires the distinguishing of pattern from blank, whereas object recognition, for example letter identification, requires classification by the observer into one of many learned categories.”

Translation: It's possible that letter identification doesn't implicate mechanisms used for seeing “pattern from blank,” or for seeing unfamiliar letters. To the extent that such mechanisms might, in fact, be implicated in letter recognition, we presume they consist of spatial filters.

Authors: “Here, we make use of results from hearing research (Patterson, 1974).”

Translation: An ad hoc model constructed on the basis of hearing experiments is an appropriate model for testing our conceptually confused variant of an assumption about the visual system based on another analogy to audition. (From Patterson, 1974: From Patterson (1974): “Threshold for a pulsed tone was measured as a function of its distance in frequency from the edge of a broad band of noise with very sharp skirts. Tone frequency was held constant at 0.5, 1.0, 2.0, 4.0, or 8.0 kHz while the position of the noise edge was varied about the frequency of the tone. The spectrum level of the noise was 40 dB. As expected, tone threshold decreased as the distance between the tone and the noise edge increased, and the rate of decrease was inversely related to tone frequency. The data were used in conjunction with a simple model of masking to derive an estimate of the shape of the auditory filter. A mathematical expression was found to describe the filter, and subsequently, this expression was used to predict the results reported by several other investigators.”)

Authors: We find that letter-identification and grating detection filters are identical, showing that the recognition of these objects at one size is mediated and constrained by a single visual filter, or 'channel.'

Translation: Our “single visual filter” conclusions contradict the speculations of Campbell and Robson (1968) (the authors we cited as support for the existence of filters to begin with) i.e. that “we may suppose that the visual system behaves not as a single detector mechanism preceded by a single broadband spatial filter but as a number of independent detector mechanisms each preceded by a relatively narrowband filter 'tuned' to a different frequency.” We don't know if our results on "letters" apply to other shapes. The varied shapes and proportions of "letters" have no bearing on the issue. For the visual system, "letters" is a special category of its own.

Authors: Surprisingly, our noise-masking paradigm shows that the same channel performs both low-level detection of narrow-band gratings and high-level identification of broad-band letters.

Translation: Uncorroborated assumptions inspired by analogies to electrical circuits and sound systems, which ignore the fact of neural inhibition and the principles of perceptual organisation, allowed us to show that hypothetical grating-detectors are responsible for the processes involved not only in perceiving, but also in recognizing letters. (Clearly a breakthrough for memory research as well as perception.) Also, we are arbitrarily splitting the part of the pattern we are showing observers into “noise” and “signal” terms, and assume that the visual system agrees with our analysis and segregates letters from splotches, without recourse to any organizing constraints (e.g. figure-ground rules). (Also, see again comment above re: single channel.)

Authors: (Blurb to Figure 1): Note that changes in viewing distance, from 3 to 60cm, hardly affect the visibility of any given letter, indicating that the channel scales with letter size. (cont'd in next comment).

On 2016 Mar 12, Lydia Maniatis commented:

The claims the authors make in this publication lack both empirical and logical support. To make what is perhaps the most clear point, the very tenuous assumptions underlying the project are themselves contradicted by the authors own statements. This is represented as a refinement of the fundamental assumption, but in fact it pulls the rug out from under it. Each step of the way, definite assertions are made which citations provided do not support.

The train of thought (with comments) is this:

Authors: “We hear periodic sounds, or tones, by means of parallel auditory filters, each tuned to a band of temporal frequency (Fletcher, 1940)”

From MIT online Encyclopedia of Cognitive Sciences: To account for the obtained data, Fletcher proposed a “critical band” that likened the ear to a bandpass filter (or, to encompass the entire frequency range, a set of contiguous, overlapping bandpass filters). The proposed “auditory filter” is a theoretical construct that reflects frequency selectivity present in the auditory system...”

Translation: The auditory system exhibits frequency selectivity. We could find no more convincing citations with respect to auditory filters between 1940 and 1994.

Authors: “...and we see periodic patterns, or gratings, by means of parallel visual filters (Campbell and Robson, 1968).”

From Robson and Campbell, 1968: “Thus, it seems that we cannot satisfactorily model the over-all visual system by a simple peak detector following a spatial filter...As a modification of this theory, we may assume...Thus, we may suppose that the visual system behaves not as a single detector mechanism preceded by a single broadband spatial filter but as a number of independent detector mechanisms each preceded by a relatively narrowband filter 'tuned' to a different frequency...Such a model could account for our findings...”

Translation: Robson and Cambell's (1968) results did not bear out their original predictions, and they speculated as to possible alternative accounts. Between 1968 and 1994, no firmer conclusions were apparently available.

Note also that the analogy between frequency of pressure waves and the “frequency of periodic patterns (gratings)” is very loose. What is being claimed, specifically, is that there is a class of visual neurons that are excited by stimuli with a specific, sine-wave-type luminance structure across a region of space (more specifically, across a region of the retina), and that they are sensitive to the frequency of these spatial luminance alterations. But frequency of pressure waves is a fundamental feature of the energy that interacts with our sensory system to initiate perception. Grating patterns, on the other hand, are just one of an infinite number of possible patterns of light impinging on our retinas. If assumptions about vision are to rest on analogy, then frequency of light waves would seem more appropriate.

Authors: “Do these visual filters participate in everyday tasks such as reading and object recognition?”

Translation: We assume the existence of special, “low-level” mechanisms for detecting grating patterns of light on the retina. (Campbell and Robson, 1968 specified visual angle, i.e. extent on the retina, in defining their patterns). We wonder if these exist exclusively to detect grating patterns, or whether they are engaged in seeing other things as well. We're particularly interested in the perception of familiar objects that we can name upon seeing. The possibility that the visual system may possess “low-level” neurons interested only in ecologically irrelevant grating patterns but not in objects seems to us to be not only credible but testable and worth testing...

Authors: “After all, grating visibility only requires the distinguishing of pattern from blank, whereas object recognition, for example letter identification, requires classification by the observer into one of many learned categories.”

Translation: It's possible that letter identification doesn't implicate mechanisms used for seeing “pattern from blank,” or for seeing unfamiliar letters. To the extent that such mechanisms might, in fact, be implicated in letter recognition, we presume they consist of spatial filters.

Authors: “Here, we make use of results from hearing research (Patterson, 1974).”

Translation: An ad hoc model constructed on the basis of hearing experiments is an appropriate model for testing our conceptually confused variant of an assumption about the visual system based on another analogy to audition. (From Patterson, 1974: From Patterson (1974): “Threshold for a pulsed tone was measured as a function of its distance in frequency from the edge of a broad band of noise with very sharp skirts. Tone frequency was held constant at 0.5, 1.0, 2.0, 4.0, or 8.0 kHz while the position of the noise edge was varied about the frequency of the tone. The spectrum level of the noise was 40 dB. As expected, tone threshold decreased as the distance between the tone and the noise edge increased, and the rate of decrease was inversely related to tone frequency. The data were used in conjunction with a simple model of masking to derive an estimate of the shape of the auditory filter. A mathematical expression was found to describe the filter, and subsequently, this expression was used to predict the results reported by several other investigators.”)

Authors: We find that letter-identification and grating detection filters are identical, showing that the recognition of these objects at one size is mediated and constrained by a single visual filter, or 'channel.'

Translation: Our “single visual filter” conclusions contradict the speculations of Campbell and Robson (1968) (the authors we cited as support for the existence of filters to begin with) i.e. that “we may suppose that the visual system behaves not as a single detector mechanism preceded by a single broadband spatial filter but as a number of independent detector mechanisms each preceded by a relatively narrowband filter 'tuned' to a different frequency.” We don't know if our results on "letters" apply to other shapes. The varied shapes and proportions of "letters" have no bearing on the issue. For the visual system, "letters" is a special category of its own.

Authors: Surprisingly, our noise-masking paradigm shows that the same channel performs both low-level detection of narrow-band gratings and high-level identification of broad-band letters.

Translation: Uncorroborated assumptions inspired by analogies to electrical circuits and sound systems, which ignore the fact of neural inhibition and the principles of perceptual organisation, allowed us to show that hypothetical grating-detectors are responsible for the processes involved not only in perceiving, but also in recognizing letters. (Clearly a breakthrough for memory research as well as perception.) Also, we are arbitrarily splitting the part of the pattern we are showing observers into “noise” and “signal” terms, and assume that the visual system agrees with our analysis and segregates letters from splotches, without recourse to any organizing constraints (e.g. figure-ground rules). (Also, see again comment above re: single channel.)

Authors: (Blurb to Figure 1): Note that changes in viewing distance, from 3 to 60cm, hardly affect the visibility of any given letter, indicating that the channel scales with letter size. (cont'd in next comment).

On 2016 Mar 12, Lydia Maniatis commented:

Continuation of previous:

Translation: Forget everything we've assumed so far about spatial filters. We don't have special detectors sensitive to the frequency of light bands falling across the retina! Visual angle doesn't matter. The visual system has special detectors for detecting the number of light/dark cycles per (familiar) "letter" - (about 3), actually a range of cycles, since some letters are quite wide (e.g. m's) and others quite narrow (e.g. l's). The processes implicit in organizing points of light into letter shapes so that we can recruit the relevant letter-sensitive channels are not relevant. Our data plus assumptions prove it.

I think this front-page Nature article has zero credible content or face-validity.

On 2016 Apr 11, Lydia Maniatis commented:

It is worth asking why, on the one hand, Solomon and Pelli (1994) seem to be asking questions about initial steps in vision - asking about "visual filters," and, on the other, referring, not to "perception" of letters, but to "identification" or "recognition." In order to identify or recognise, we have to first perceive, but this step is skipped over not only by these researchers but by those contemporary investigators who consider themselves "psychophysicists" in general.

The reason is this: These investigators insist on treating perception as a problem of detection, such that they can recycle the signal detection models first developed by radar researchers.

Thus, they assume that identification/recognition would be perfect if it were not for “internal noise” in the system, which produces uncertainty in the outcome. This noise is the random firing of neurons, base rates etc. This claim about internal noise and perception has never been tested, nor (to my knowledge) have these researchers proposed a way to test or measure the claim, but it is assumed.

It is also assumed that we can add “external noise” to a stimulus and thus counterbalance the “internal noise.” This “external noise” can take the form of variance of the stimulus pool, splotches superimposed on the stimulus (as here), etc. Anything that can be described as “unpredictable.”

If we do this, then there is an equation that we can use that allows us to measure “performance” regardless of the task, as long as we have a value to plug in for our “noise” term.

If we can assume that the “external noise” is much bigger than the “internal noise” then we can use this equation whose output is labelled “E” for efficiency, by plugging in our “external noise” value.

Clear so far?

What is wrong with this story is also the reason researchers adopting it don't refer to “perception.”

The stimulus, which consists of retinal photoreceptors excited by disconnected photons pouring in from the environment does not contain “letters” or any other objects to be “detected.” Only photons are detected by the system. It is the response of the nervous system that organizes these points, based on rules (instantiated in the system) that have been selected for over evolutionary time to produce a good enough match between the objects out there and the objects we perceive. We can understand the importance of sophisticated rules if we consider how easy it is for humans to read the letters of “captchas” used by websites to ensure you're not a robot, but which programmers have a hard time beating. (This is actually a crude and somewhat misleading example, because even seeing the entire captcha without being able to recognize the letters it “contains” is a feat of perceptual organization). So the visual system creates a percept by adding structure to structureless data, and the number of these potential structurings is infinite. The limiting factors are the rules.

So to see a letter, even as a physical shape, implies an enormously complicated process of perceptual organization, based on inferences that constitute the output of the system. What we “identify” or “recognize” is the product of these processes. We are not “detecting” something that is there a priori, in the stimulus, the disconnected photons. Seeing splotches, or any other feature of the percept, is also the result of the same organizing processes. There's no principle justifying calling part of the percept “noise” and another part “signal” if what we are interested in are basic perceptual processes. It is relatedly a problem to overlook that some kinds of “noise” will, for organizational reasons, interfere more with the “signal” that we have in mind than other kinds, something that the makers of “captchas” know well.

But all of these facts are silently overlooked by investigators who choose to treat perception as a detection problem involving imaginary spatial filters and arbitrarily defined “noise”. Which is truly extraordinary.

On 2016 Aug 17, Lydia Maniatis commented:

Short version:

The authors say: "We hypothesize that how our stimuli are perceived depends on spatial frequency. Our hypothesis is confirmed. (P.S. The effect is size invariant.)"

Translation: "We hypothesize that how our stimuli are perceived depends on spatial frequency. How our stimuli are perceived DOES NOT depend on spatial frequency. Our spatial frequency hypothesis is falsified."

a + (-a) = 0

(The discrete addition of the phrase "at one size" into the abstract, as though this did not upend the entire position, is really quite extraordinary.)