From Towards a Science of Consciousness 3         Section 8: The Timing of Conscious Experience       CogNet Proceedings

1. Libet's Experiment on Backward Referral

The critical findings of Libet's experiments on time anomalies can be summarized as:

(i) A cortical or thalamic stimulus requires a duration of more than 250 msec to be felt, whereas a skin stimulus of 20 msec is adequate. The stimulus duration needed to generate a feeling is called the neuronal adequacy time (NA). It can differ from the perceived time of the feeling.

(ii) A cortical stimulus whose onset is within 250 msec after the skin stimulus can suppress the skin response if it has an overlapping felt location.

(iii) For a skin stimulus to be felt as synchronous with a nonoverlapping brain stimulus, the skin stimulus must be delayed 250 msec relative to a cortical stimulus or delayed 0 msec relative to a thalamic stimulus. [Section 3 asks whether Libet's raw data is adequate for this claim.]

(iv) Both the skin and the thalamic stimulation, but not the cortical stimulus, generate an evoked cortical potential (E) shortly after stimulus onset.

Instead of NA 250 msec that appears in item (i), Wolf and Penrose quote a 500 msec delay (however, see footnote 15 of Wolf 1998). The 500 msec value is only found when stimulating the brain with very weak, near threshold, signals. In all the experiments involving backward referral Libet used a much stronger stimulus, with NA ( 200-300 msec, see tables 29.2 and 29.3 (Libet 1979). It is strange that Libet himself (1991) says NA 500 msec when referring to his 1979 data even though his original article is clear that NA 300 in those experiments.

Items (i), (ii), and (iv) are summarized in panel A of figure 29.1. The XX in connection with the skin stimulation indicates when the skin feeling can be canceled by a cortical stimulus (item (ii)). Item (iii) is summarized in panels B or C by the circled Q's that indicate the timing of the qualia.

For this illustration the thalamic, skin and cortical stimuli are assumed to have the same onsets.

Libet argues that his data on cortical stimulation implies a backward referral in time since item (ii) seems to imply that the peripheral stimulus doesn't become conscious for about 250 msec in order that it can be canceled by the cortical masker. Backward referral is then needed to account for the combination of items (ii) and (iii). The backward referral argument is simpler in the thalamic case since there is a decoupling between the duration for neuronal sufficiency and the time of subjective equality of skin vs. the thalamic stimulation. It is unfortunate that Penrose focuses his discussion on the cortical experiments since the "time anomaly" in the thalamic data is simpler and clearer. Libet argues that since the thalamic sensation doesn't reach neuronal adequacy until 250 msec after the pulse train begins (item i) a back referral mechanism is needed with thalamic stimulation to produce the synchrony sensation (item iii) that is perceived.

[Place figure 29.1 about here; figures not yet available]

2 Multiple Explanations of Libet's Data

Dennett points out that our brain has adopted numerous tricks to make sense out of its tumultuous neural activity. He calls some of these tricks Orwellian and some Stalinesque. A Stalinesque account involves hallucinating (misperceiving) whereby the brain generates images not in agreement with the input information (similar to Stalin's staged trials). An Orwellian account involves rewriting history (now called false memories). For example, item (ii) above would be an Orwellian mechanism. Dennett spends the main part of his chapter 6 ("Time and Experience") on Libet's experiments. Further clarification on the application of the Stalinesque vs Orwellian means of dealing with Libet's time anomalies can be found in Libet's (1993) article in the CIBA Foundation Symposium volume. At the end of that article is a discussion by a number of philosophers including Dennett and Searle together with Libet's responses and several counter-responses.

Three explanations of the subjective data are shown diagramatically in panels B, C, and D of figure 29.1. These explanations make use of the notion of a time marker that is either the evoked potential, E, or the point of neuronal adequacy, NA, whichever is earlier (E for thalamus and skin, NA for cortex).

(panel B) A backward causation explanation, whereby the stimulus is perceived before thalamic neuronal adequacy is reached. The perception is referred back to the time of the evoked potential (E) if an evoked potential is present. Note that since all the time markers and qualia follow stimulus onset there is no violation of physical causality in the data.

(panel C) A Stalinesque explanation, whereby the time of the sensation is a fixed delay from the time marker. This is also Glynn's (1990) account.

(panel D) An Orwellian explanation, whereby the qualia for any stimulus are delayed, but the timing of those perceptions is not recorded. What gets into memory is based on the time marker information as in panel B. This explanation actually involves backward referral of memories laid down later, pointing to the time markers, so there is no need for the quantum mechanism suggested by Wolf (1998, 1999) or the causality violating "time symmetric" mechanism suggested by Bierman and Radin (1999).

Dennett does a lucid job of applying the Stalinesque and Orwellian explanations to the Libet data without any need for quantum mechanisms. Dennett goes on to argue that there is no difference between an Orwellian and a Stalinesque account, and concludes that qualia are illusory. I think these claims go too far, and that we will be able to distinguish the two sorts of accounts when the neural correlates of consciousness are discovered. One hopes that discovery will clarify the many issues regarding qualia.

Physicists like Penrose and Wolf tend to find the Stalinesque and Orwellian explanations as ad hoc and distasteful. These explanations lack the elegance of the grand principles underlying many physical phenomena. Psychologists and biologists, on the other hand, are quite familiar with the inelegant contrivances that evolution invents again and again.

Wolf (1998) asks "what purpose could evolution have in allowing such a strange and confused temporal ordering of conscious experiences?" To me the answer seems obvious. The brain has evolved to compensate for delays in timing. It makes ecological sense that the perceived time of touch should be referred back to the actual time of touch to minimize confusions of time ordering. It is known that temporal compensation mechanisms are found throughout the nervous system. Let me give two examples from vision. In the retina there are time delays in the ganglion cell (the output cell of the retina) signals depending on the distance of the ganglion cell from the optic nerve. But by the time the neural activity reaches cortex these delays have been compensated. The second example is from my own research (Baldo and Klein 1995). We used a spatial alignment task to compare the subjective timing of a flashed dot relative to a moving dot (Nijhawan 1994) and found that subjects align the flashed dot to where the moving dot will be about 100 msec later. This predictive mechanism would be a clever evolutionary adaptation for creatures trying to catch flying bugs (among other challenges with moving stimuli). Forward referral has selection advantages when dealing with moving objects, just as backward referral has advantages for designing the motor control system of hand-eye-world coordination. It is not surprising that evolution would come up with the strategy of using time markers to synchronize events across the brain. Time anomalies will occur when an evoked potential time marker is missing, as happens with cortical stimulation. Typically, illusions are a byproduct of sensory mechanisms that were developed by evolution for one purpose being now used in an unusual context (Gregory 1970).

3 A Reanalysis of Libet's Raw Data

In order to clarify Libet's data it is helpful to take a detour and examine a simpler task of comparing the relative location of two arrowheads (upper left of figure 29.2 panels). The lower arrow is always 3.0 mm in length and the length of the upper arrow is randomly varied from trial to trial with the choices being (2.9, 2.95, 3.0, 3.05, 3.1) mm. Suppose each length is shown 50 times and the observer is asked to judge whether the upper arrowhead is to the right or left (longer or shorter) of the reference. Table 29.1 shows a possible dataset. The probability that the upper arrow is judged to be longer is plotted as asterisks in figure 29.2A. This plot of frequency of seeing as a function of stimulus strength is called

psychometric function.

Table 29.1

Data for figure 29.2A. The number of occurrences of each outcome.

[Place figure 29.2 about here; figures not yet available]

The horizontal axis is the length difference of the upper and lower arrows in microns. The vertical axis is the fraction of times that the observer says the upper arrow is longer. The curve that is fit to the data is a cumulative normal distribution, a form commonly used for fitting psychometric functions (Levi, Klein, and Aitsebaomo 1984). Two numbers can be extracted from the psychometric function: the point of subjective equality (PSE) and the threshold.

Parameter 1, the PSE, is the length difference corresponding to the 50 percent point. It would be expected that the PSE should be near zero, but often small biases are found in the perceived location of the upper or lower arrow (or perceived time of brain vs. skin stimulation in Libet's case). For example, the lower arrow might seem smaller because of the "moon illusion" caused by perspective cues that make it seem closer, together with a "size constancy" mechanism. Another possible source of the bias is related to what the subject does in cases when the two lengths appear equal. The subject could either: 1) randomly distribute those responses to the two categories (but this introduces noise and gives the psychometric function a shallower slope), or 2) put those responses into one category (this introduces a PSE bias, but the psychometric function remains steep). The latter strategy is the one typically chosen by psychophysicists since the PSE is easily biased, while thresholds (slopes) are more trustworthy.

Parameter 2, the threshold, measures the shallowness (slope) of the psychometric function. In this article, threshold is defined as the abscissa range needed to go from 10 percent to 90 percent correct (2.6 times the standard deviation for an underlying normal distribution).

Figure 29.2B is a psychometric function similar to figure 29.2A except that now the test pattern involves comparing an arrow with a regular arrowhead to an arrow with a backward arrowhead. The threshold is the same but now the PSE has increased to 300 microns. In this case the PSE is substantially greater than the threshold and our judgment is called an optical illusion (the Muller-Lyer illusion). When the PSE shift is less than threshold (as in figure 29.2A) the shift is called a response bias rather than an illusion (a point to be remembered for when we look at Libet's data) and one doesn't look for special mechanisms. Only when a true illusion is present as in figure 29.2B does a major industry among perceptual psychologists (or quantum physicists) develop to understand the brain mechanisms that can account for the data.

Figure 29.2C shows two psychometric functions collected using a procedure similar to Libet's where the number of response categories is increased from two to three: a) upper arrow is left of lower, b) upper arrow equal to lower, and c) upper arrow right of lower. A sample dataset using three response categories is given in table 29.2.

Table 29.2

Raw observations for figure 29.2C. Three categories, similar to Libet.

The "upper is rightward" category of table 29.2 is the same as in table 29.1, implying that in table 29.1 the "upper is equal" and the "upper is shorter" categories had been combined. The psychometric function for that criterion are given in the bottom row of table 29.3. A second criterion groups the "upper is equal" with the "upper is rightward" as shown in the upper row of table 29.3.

Table 29.3

Probabilities for figure 29.2C. The two psychometric functions from table 29.2 data.

The results of fitting thess data are the three parameters: PSE₁ = -49.4±5.0 microns, PSE₂ = +49.4±5.0 microns, and threshold = 2.6 = 92 ± 10 microns.

I was first inspired to look closely at Libet's raw data by reading Wolf's (1998) article. He reproduced a portion of Libet's table 2a in order to test a hypothesis about a 20 msec delay between thalamic stimulation vs. skin stimulation. Being skeptical about whether the data was good enough to make 20 msec reliable distinctions, I decided to plot and refit Libet's raw data.

Figure 29.3 plots the data from table 2a of Libet (1979). This is just like figure 29.2C except that now the abscissa is the temporal difference rather than spatial difference between two percepts. The vertical axis is the percent of occurrences that a stimulus was perceived to be later than the reference stimulus. The three response categories are: skin first, a tie, and thalamus first (for the skin-thalamus experiment with skin as a reference). The total responses in the three categories were approximately equal (54, 66, 52 trials respectively). The lower curve (diamonds) is the probability that the skin was first. For the upper curve (asterisks) the "equal" votes were combined with the "skin first" results corresponding to a looser criterion for the skin being first. For the upper curve (asterisks) the "equal" votes were combined with the "skin first" results corresponding to a looser criterion for the skin being first. It corresponds to the probability that the thalamus was chosen as last. The upper pair of panels are for observer HS; the lower are for GS. The right pair of panels compares the perceived timing of a thalamic stimulus relative to a nonoverlapping skin stimulus; the left compares skin to skin.

The two continuous curves in each panel are the best fits to the data. The fitting function is the standard cumulative normal distribution given by:

          prob(z_i) = g + (1-2g)exp(-(x-PSE_i)²/2²) dx                 (1)

where ( is the standard deviation (stimulus threshold defined as the stimulus range for 10% to 90% correct is 2.6). A guessing parameter, g, has been included (Manny and Klein 1984) to allow for the possibility of making absent-minded errors even when the stimulus delay is large; g was constrained to equal 0.025, a small number. The chi square goodness of fit is much worse and the thresholds are elevated if g = 0. Three free parameters were used to fit the data for each panel: two criteria, PSE_i, (with i = 1 or 2) to locate the horizontal position of each curve, and one parameter, s, to determine the slopes (threshold) of the curves. The two curves were assumed to have the same slopes.

[Place figure 29.3 about here; figures not yet available]

In order to help make the connection between figure 29.3 and the raw data in Libet's table 2a we have reorganized the table so that it is similar to our previous tables. Table 29.4 shows the data for the thalamic vs. skin comparison for subject GS. We have switched the sign of the abscissa so that the delay is relative to the skin stimulus reference in all cases. That makes comparing the delays less confusing.

Table 29.4