[lagudragu]

I'm honestly a bit surprised that such an article is placed in Nature, as there seem to be a couple questionable details in the article regarding the experiment:

- Only 3 volunteers on which the experiment has been tested

- The volunteers were left "in total darkness for around 40 minutes" before the actual experiment

- "In many cases, they got it wrong; this is to be expected, given that more than 90% of photons that enter the front of the eye never even reach a rod cell, because they are absorbed or reflected by other parts of the eye. Still, participants were able to answer correctly more frequently than would be expected if they had guessed at random and their confidence level was higher when they were right."

- The three volunteers sat through a total of more than 2,400 trials in which a single photon was emitted (and many more in which it was not).

- That high volume of testing, the researchers say, gives them strong statistical evidence of single-photon detection

- The participants had to say on which occasion they thought they saw a photon, and how confident they were (on a scale of 1 to 3) about their sighting.

In summation, they put 3 volunteers for 40 minutes in total darkness, then performed more than 2400 trials in total (i assume 1000+ trials per person, taking into account the tests in which no photon was emmitted). "Many" times the volunteers guessed wrong, but when they guessed right they were pretty confident of it (using a rather simplistic scale of 1-3).

What I'm seeing with this experiment is a result which is more confirmed due to psychological bias rather than actual results by putting these volunteers in a straining test. How would you, as a volunteer react, if you were left in the darkness for such a long test duration and would asked over a 1000 times whether you have seen a flash.

Edit: After having read the actual paper on the methodology, I do retract my comment.

[suchow]

In the typical psychophysical study on perceptual thresholds, participants are given a two-alternative forced-choice test: a stimulus (here, a photon) is placed in one of two intervals at random, and the participant guesses which interval had the photon. Using this design, a bias that systematically selects one response more often than the other will lead to performance below chance. Finding that performance is significantly better than chance is evidence that the observer can detect a single photon. Demonstrating that performance is reliably above chance requires many trials because the level of performance is close to chance. Compare this to a yes/no design, where a bias can create the appearance of an ability to detect a single photon where there is none.

The reason for waiting 40 minutes is that the eye and brain adapt to darkness. If this experiment were performed outside in daylight, it would be impossible to detect a difference of 1 photon. Only when the observer sits in a room without light can the brain adapt and have its greatest sensitivity to light.

Finally, note that the logic of this study is an existence proof that people can detect single photons. Selecting three normal observers and finding that all of them have this capability is reasonable evidence that most normal observers can do the same, unless you have some specific reason to believe that these observers are unrepresentative of the population (as the researcher in the press release did re gender differences).

[suchow]

I have a slight correction to my earlier comment, in which I said "Using this design, a bias that systematically selects one response more often than the other will lead to performance below chance." A bias of this kind will drive performance closer to chance. Performance that is significantly below chance would be evidence that observers can see single photons, but that they have the response labels reversed in their minds.

[lagudragu]

I do understand the procedure of this experiment. However doesn't the experiment strain mentally (thus possible generating incorrect results)?

Taking the following from the paper:(http://www.nature.com/ncomms/2016/160719/ncomms12172/full/nc...):

Averaging across subjects’ responses and ratings from a total of 30,767 trials, 2,420 single-photon events passed post-selection and we found the averaged probability of correct response to be 0.516±0.010 (P=0.0545; Fig. 2a).

Edit: Having (quickly) read the paper, it does seem the psychology of the volunteers was taken into proper consideration, I stand corrected:

Before collecting data, subjects were extensively trained using a classical light source with photon number between 1 and 15 photons at the cornea (Supplementary Fig. 4a). The improved performance with experience is clearly and quantitatively visible (Supplementary Fig. 4b,c). Subjects typically required 6–8 sessions, performing one session a day, to reach their optimal performance level (Supplementary Fig. 4b,c). Each session took~2 h, when including dark adaption.

During data acquisition, each subject went through up to 20 sessions. Still this high amount of sessions was not enough to obtain statistically significant performance for individual subjects, and therefore we pooled the data together to increase significance (Fig. 2a–d). As subject’s sensitivity and criteria used to assign the confidence ratings might vary in psychophysics trials28, we aimed to minimize or normalize possible factors causing variability to achieve maximum sensitivity and similarity across subjects by using extensive training of the subjects and using our 2AFC paradigm.

Additionally, the research does have interesting observations:

Surprisingly, a strong dependence on the temporal separation of the two events was observed peaking at ~3.5 s, with a decay time on the order of seconds (Fig. 2c). Such a long timescale phenomenon represents more than an order of magnitude disparity with the known integration time of the visual system4. This result directly shows that the probability of correctly reporting a single photon is highly enhanced by the presence of an earlier photon within ~5 s time interval.

Thus, consistent with both observations, we suggest that the detection of a single photon – or equally a photon-like noise event (that is, spontaneous isomerization) – temporarily increases the effective gain of the visual system under extreme low-light conditions, such that a second temporally coinciding photon (or photon-like noise event) can be behaviourally detected with a higher probability.

[IshKebab]

What do you mean it is "confirmed due to psychological bias"? If they answered better than chance, and it was statistically significant then what's the problem?

[Retr0spectrum]

I haven't read the full paper, so this is only speculation.

I expect they would have counted the number of times the volunteers claimed to have seen a photon, both when a photon was emitted, and when they only said that one was emitted.

If the detection rate is significantly higher for "real" emissions than for the "fake" emissions, then you can draw a statistically significant conclusion (aside from the fact that there were only 3 test subjects).

[saalweachter]

Just a note, when the number of volunteers is less than the number of authors, you're probably looking at self-experimentation.

It's not uncommon when you have a cool idea you want to test that doesn't rely on naive subjects and it would take longer to get permission to & recruit outside volunteers -- just grab a few people from around the lab and try it out.

[zerohp]

Kwiat was already recruiting undergrads in 2013 for this experiment. If I recall, he did not yet have a reliable single photon source but was expecting to have it very soon.

I was in his undergrad quantum physics (for engineers) course at the time.

[yyyuuu]

If they had not detected even once, then your points hold. But they did detect, so the point is that they were able to see? What more do you want? It simply shows that people can sense photons. Maybe not all the time, but sometimes for sure.