In the year 2000, the biologists Amish Dave and Daniel Margoliash published a report in the journal Science describing heir research on zebra finches (Taeniopygia guttata), which are passerine birds native to Australia. One of the evolutionary challenges these birds face is that they must learn their song from their parents and siblings, since it is not innate.

Research on birdsong has historically focused on what these animals do while awake to imitate and memorize their song, but Dave and Margoliash wondered whether sleep might also play a role in song acquisition. Could sleep help juvenile finches internalize the acoustic patterns they hear from their family members and commit them to long-term memory? Could these birds learn their song at least in part by rehearsing it in their minds while asleep?

To test this possibility, Dave and Margoliash performed an experiment in which they mapped the patterns of neural activation elicited in the “birdsong system” (the forebrain nucleus robustus archistriatalis) of a group of juvenile finches while they slept. By analyzing these patterns, they discovered that the brains of zebra finches oscillate between two states during sleep: a state of constant but low-level neural activity in which nothing remarkable happens, and a state marked by spontaneous bursts of high-level neural activity recurring at regular intervals.

In itself, this discovery was not particularly groundbreaking, as it merely confirmed previously published work on avian sleep reporting a sleep cycle divided into phases of low and high neural activity (just like mammalian sleep). However, Dave and Margoliash then decided to map the neural pattern that emerged in the same brain region when the finches practiced their song while awake and to compare it to the patterns elicited during sleep. Their findings were astonishing.

They discovered that the pattern elicited by the act of singing while awake was an exact structural replica of the pattern elicited during the period of sleep marked by sudden bursts of high-level neural activity, which was all they needed to realize that the brains of the zebra finches were doing the exact same thing—their neurons were firing in the same organized manner—when the birds sang their song in the middle of the day for the world to hear as when they entered a period of high neural activation during sleep.

The match was so perfect that the authors realized they could map these patterns onto one another syllable-by-syllable, nay, note-by-note. From this, they concluded that zebra finches learn their song not only by practicing it out loud while awake (“play”) but also by mentally replaying it while asleep without making a chirp (“replay”). “Replay,” they write, “generates coherent activity throughout the song system that is similar to singing in the absence of actual sound production and perception.”

One might argue that the activation of the birdsong system during sleep is evidence that zebra finches were dreaming of singing their song. Curiously, Dave and Margoliash resist this interpretation. Instead, they argue that the replay they observed in the finches represents nothing more than the execution of a computational process (in their words, an “algorithmic implementation”) that unfolds without the finches’ conscious awareness.

In their view, finches do not experience replay any more than my laptop experiences running Adobe Reader or Microsoft Word, since replay itself is a brain state wholly devoid of what Ned Block calls “experiential properties.” Bluntly put, it has no accompanying phenomenology.

Zebra finches memorize their song not merely by sleeping on it, but by dreaming about it as well.

In my view, this algorithmic interpretation is not entailed by the data so much as superimposed upon it—and, I should add, without much in the way of justification. It is not at all clear why Dave and Margoliash interpret replay algorithmically, especially when their own findings suggest that it represents a lived reality that zebra finches experience from a first-person perspective during sleep.

Two pieces of evidence, I argue, tip the balance against computationalism in favor of phenomenology. One is temporality. Aside from finding structural parallels between play and replay, Dave and Margoliash also found temporal ones. It took the finches roughly the same amount of time to sing their song while awake as it did to rehearse it during sleep.

This is important since there is no prima facie reason why a computational process would need to run on the same timescale as the subjective experience it mechanically reiterates. This temporal parallelism could be the result of a shared underlying phenomenology linked to the animals’ experience of lived time. If it takes these animals the same amount to time to play their song while awake as to replay it in their sleep, this could be because play and replay instantiate a similar subjective experience.

The other piece of evidence here is embodiment. Dave and Margoliash noticed that it was not only the brain that was conscripted into the performance of replay, but the body as well—especially the throat. During replay, the birds’ vocal cords expanded and contracted exactly as they did during play, which can only mean one thing: as the birds went through the steps of mentally rehearsing their song during sleep, they also practiced the bodily skills needed to actualize this song.

Granted, the movements of the vocal cords did not yield any sound, but the fact that they occurred at all indicates that the memories the animals were retrieving during replay were inexorably embodied. During replay, the animals were not remembering that; they were remembering how. And it is highly likely that in the process of remembering how, the birds had a genuine auditory experience since the auditory regions of their brains also lit up like a Christmas tree. The sleeping finches, it seems, “heard” their own song in the heavy silence of sleep.

Thus, I agree with Dave and Margoliash’s observation that replay takes place in the absence of “sound production,” but not with their assertion that it also occurs in the absence of “perception.” There is a massive difference between the two. Sound production is an objective state of affairs: did the animals sing? Did they produce sound waves? Perception is about a subjective state: did they hear a song? Did they experience sounds? My reading of the data is that the birds did not sing for the simple reason that they emitted no sound, but they did hear a song—what I am calling a “soundless song.” They heard it silently, much like we hear the clamoring soundscapes of our own dreams—the voice of a lover, the rustling in the trees, the sound of a church bell in the distance.

Unfortunately, Dave and Margoliash cannot see this because they are committed to a computationalist interpretation of replay, which I believe causes them to miss the phenomenological significance of their own findings. Zebra finches memorize their song not merely by sleeping on it, but by dreaming about it as well. As the seventeenth century poet John Dryden wrote in his 1665 play, The Indian Emperor: “The little birds in dreams their songs repeat.”


Although Dave and Margoliash’s comments about the absence of perception during bird sleep can prompt skepticism about animal dreaming, we find a refreshing alternative to their algorithmic interpretation of replay in a study conducted in 2001 by Kenway Louie and Matthew Wilson from the Massachusetts Institute of Technology (MIT). While working at MIT’s Center for Learning and Memory, Louie and Wilson sought to gain a better understanding of how sleep affects memory and spatial  reasoning in rats. To do so, they decided to study how rats would mentally deal with a spatial task while awake and while asleep.

They selected a spatial task because rats, like humans, have a sophisticated space-mapping system in the hippocampus made up of CA1 pyramidal cells that map the animal’s physical environment and fire differentially depending on the animal’s position in space. When a rat occupies position X in the mapped environment, a specific set of CA1 cells fire; but as soon as the rat moves to position Y, a different set of CA1 cells begin firing. Critically, if the rat then returns to position X, the exact same set of CA1 cells that originally fired will fire again.

As long as the rat’s physical environment remains relatively constant, researchers can pinpoint the rat’s physical location with extreme precision based solely on hippocampal activation information. By tracking hippocampal activity, then, Louie and Wilson could track the physical location of the rats while they were awake, as well as the location the rats thought they occupied while asleep.

The experiment began by acclimating a group of rats to an elevated circular track and training them to run “from a start location to a goal location for a food reward.” Once the rats learned this route, Louie and Wilson tracked single-cell activity in the hippocampus and recorded the specific pattern of CA1 pyramidal cell activation that resulted from their movement, thereby mapping “the sequence in which the animal’s behavior [took] it through the task environment.” They called this pattern “RUN” since it was produced by the act of running towards the reward.

Then, they wondered whether the pattern of pyramidal activation associated with RUN might re-appear during REM sleep, so they let the rats take a nap after running the track and recorded hippocampal activity while they slept. They called this second pattern “REM” since it occurred while the rats were in REM sleep. Thus, in this context RUN and REM refer to neural patterns: one connected to physically running a track while awake, and the other connected to mentally replaying this act during REM sleep.

So, what did Louie and Wilson learn? Echoing Dave and Margoliash’s findings about birdsong, they discovered that RUN and REM are mirror images of one another, meaning that when the rats fell asleep, they probably dreamed about the spatial test they had just completed. Furthermore, Louie and Wilson found that RUN and REM unfolded “at approximately the same speed.” Just as zebra finches replayed their song while asleep in the same amount of time that it took them to play it while awake so, too, rats performed RUN and REM on a comparable timescale, on the order of minutes to seconds. Structurally and temporally, “REM recapitulates RUN.”

This is where things get interesting. Technically, Louie and Wilson merely replicated Dave and Margoliash’s findings about the structural and temporal parallels that connect waking and sleeping states, but their interpretation of these parallels could not be more different. In contrast to Dave and Margoliash, who interpreted replay in zebra finches as an unconscious algorithmic process without experiential properties, Louie and Wilson interpret replay in rats as a phenomenologically rich experience—in other words, as a dream. Replay must be a lived reality for the rats, they say, because it depends on a past experience and because it mirrors this experience structurally and temporally.

Instead of a phenomenologically hollow state, replay is a genuine subjective experience “despite the absence in REM of the explicit sensorimotor cues that drive distinct neural patterns during RUN.” Even without any of the sensorimotor cues that they had while running the track in the waking state (such as visual information about the environment, the feeling of the ground beneath their feet, and the smell of the food reward at the end of the track), the sleeping rats experienced running toward the reward. They generated an internal simulation that “re-activated” or “re-constructed” a waking behavior. And, of course, to reconstruct a waking behavior during sleep simply is to dream about that behavior, so this amounts to saying that the rats were dreaming of running the track.

Let’s be clear about this: the dispute between Louie and Wilson on the one hand and Dave and Margoliash on the other is not scientific. It is philosophical. At stake in it is the question of the kinds of beings that rats are. Are they furry little computers that implement algorithms? Or are they conscious subjects with an inner phenomenology, subjects who perceive, feel, and think?

This is not a question that can be answered on purely empirical grounds, which is why these researchers can agree about the facts yet disagree about what these facts ultimately mean. The crux of their disagreement is reflected in the terms they use to describe what they see: “algorithmic implementation” versus “internal simulation.”

The first places replay squarely within in the ambit of computationalist theories of mind, whereas the second portrays it less as a program that the rat brain runs unconsciously and more as an experience that rats live through while asleep with the full thrust of their being.


Text from WHEN ANIMALS DREAM: The Hidden World of Animal Consciousness by David M. Peña-Guzmán. Copyright © 2022 by Princeton University Press. Reprinted by permission of Princeton University Press.

Source link