1 Introduction

Classical psychedelics—including psilocybin, mescaline, DMT, and LSD—are a family of hallucinogenic compounds with a common mechanism of action: they are agonists for the 5-HT2a serotonin receptor commonly expressed on the apical dendrites of cortical pyramidal neurons [1]. These drugs induce numerous effects in human subjects, including: complex visual, auditory, and tactile hallucinations; intense spiritual experiences; long-lasting alterations in mood; changes in personality; and increases in synaptic plasticity [2, 3, 4]. They have further been used for millennia as medicine and in religious rituals [5]; more recently, they have been explored clinically as potential treatments for depression and anxiety [6], as well as PTSD [7].

The 5-HT2a receptor plays a critical role in psychedelic-induced hallucinations. Indeed, perceptual effects are largely eliminated by blocking these receptors in the cortex [8, 9]. However, very little is understood about why highly structured hallucinations and changes in synaptic plasticity emerge from activating cortical 5-HT2a receptors: to explain this, it is necessary to develop mechanistic theories that are capable of linking changes in neuron-level properties (receptor agonism) to changes in perception and behavior. Psychedelic drug users and therapists have long noted the ‘dream-like’ qualities of psychedelic drug hallucinations, which are realistic but untethered from the external world; this observation leads naturally to speculation that these drugs are ‘oneirogens,’ or dream-manifesting compounds [10]. However, beyond perceptual phenomenology (and some evidence pointing to the effects of psychedelics on sleep cycles [11, 12, 13]), we lack a mechanistic proposal that could explain the similarity between dreams and psychedelic drug experiences. Here, we articulate the ‘oneirogen hypothesis’, which describes one such potential mechanistic explanation. We propose that classical psychedelics induce a dream-like state by shifting the balance between bottom-up pathways transmitting sensory information and top-down pathways ordinarily used to create replay sequences in the brain. Replay sequences have been shown to be important for learning during sleep [14, 15, 16, 17, 18]: we propose that mechanisms supporting replaydependent learning during sleep are key to explaining the increases in plasticity caused by psychedelic drug administration. In total, our model of the functional effect of psychedelics on pyramidal neurons could provide a explanation for the perceptual psychedelic experience in terms of learning mechanisms for consolidation during sleep [19], and cortical ‘replay’ phenomena [20, 21, 22, 23, 24, 25, 26, 27, 28].

To explore the oneirogen hypothesis concretely, we use the aptly named Wake-Sleep algorithm [29], which has historically been used to train artificial neural networks (ANNs) that possess both a bottomup “recognition” pathway and a top-down “generative” pathway to learn a representation of incoming sensory data. It enables unsupervised learning in ANNs by alternating between periods of “waking perception” (wherein bottom-up recognition pathways drive activity) and “dreaming sequences” (wherein top-down generative pathways drive activity). With these alternate periods of distinct activity, connectivity parameters in each pathway are adjusted to match the activity of the opposite pathway. This way, the top-down pathway learns to generate activity consistent with that induced by sensory inputs, and the bottom-up pathway learns better representations thanks to generated activity.

In this work, we show that within a neural network trained via Wake-Sleep, it is possible to model the action of classical psychedelics (i.e. 5-HT2a receptor agonism) by shifting the balance during the wake state from the bottom-up pathways to the top-down pathways, thereby making the ‘wake’ network states more ‘dream-like’. Specifically, we model the effects of classical psychedelics by manipulating the relative influence of top-down and bottom-up connections in neural networks trained with the Wake-Sleep algorithm on images. Doing so, we capture a number of effects observed in experiments on individuals under the influence of psychedelics, including: the emergence of closed-eye hallucinations, increases in stimulus-conditioned variability, and large increases in synaptic plasticity. This data suggests that the oneirogen hypothesis may indeed help to explain why 5-HT2a agonists have the functional effects that they do. We subsequently identify several testable predictions that could be used to further validate the oneirogen hypothesis.

2 Results

Mapping the Wake-Sleep algorithm onto cortical architecture

The Wake-Sleep algorithm allows ANNs to optimize a global, unsupervised objective function for sensory representation learning— the Evidence Lower Bound (ELBO)—through local synaptic modifications to a bottom-up recognition pathway and a top-down generative pathway. As a precursor to the variational autoencoder [30, 31], the Wake-Sleep algorithm provides a mechanism for learning a probabilistic latent representation r responding to incoming sensory stimuli s, which obeys representational characteristics that are ideal for a neural system (e.g. sparsity and metabolic efficiency [32], compression and coding efficiency [33, 34], or disentanglement [35, 36]). To do this, Wake-Sleep optimizes the ELBO through an approximation of the Expectation Maximization (EM) algorithm [37] to train the two pathways (Figure 1a). For readers who are unfamiliar with the Wake-Sleep algorithm, a tutorial can be found here [38].

Figure 1.

Notably, the Wake-Sleep algorithm requires two phases of activity (i.e. “Wake” and “Sleep”), where the network phase is controlled by a global state variable α ∈ [0, 1] that regulates the balance between the bottom-up and top-down pathways. In the Wake phase (α = 0), the network processes real sensory stimuli drawn from the environment, and network activity is sampled based on the bottom-up inputs (corresponding to the approximate inference distribution). In the Sleep phase (α = 1), the network internally samples neural activity from its generative model, which then produces generated activity in the stimulus layer s. We use this structure of the Wake-Sleep algorithm as a concrete model to express oneirogen hypothesis. Specifically, we use changes to the value of α as a means of modeling a 5-HT2a agonist-induced shift to a more dream-like state, as we detail below.

Within the Wake-Sleep algorithm, neurons alternate between ‘Wake’ and ‘Sleep’ modes, where activity during each mode is dominated by the bottom-up and top-down pathways, respectively. We can determine the neural activity for a given intermediate layer l with the following equation:

where h(r) defines bottom-up input, μ(r) defines top-down input, f (h, μ, α) is any interpolation function such that f (h, μ, 0) = h and f (h, μ, 1) = μ, σ_b and σ_p define the bottom-up and top-down activity standard deviations, and η ∼ 𝒩(0, 1) adds random noise to the neural activity (see Methods for more detail). Here, for notational conciseness we treat r as a concatenated vector of all r^(l) vectors from each layer. This equation means that α controls whether bottom-up inputs or top-down inputs control the dynamics of individual neural units.

Thus, as α moves from a value of 0 to a value of 1 the activity of the neurons shifts from being driven by the bottom-up recognition pathway to being driven by the top-down generative pathway. How could this occur in the brain? In the cortex, excitatory pyramidal neurons receive inputs from distinct sources: inputs that are from ‘higher order’ cortical areas target the apical dendrites, whereas inputs that are from ‘lower order’ cortical or sensory subcortical areas target the basal dendrites [39]. Thus, we can capture the core idea behind the oneirogen hypothesis using the Wake-Sleep algorithm, by postulating that the bottom-up basal synapses are predominantly driving neural activity during the Wake phase (when α is low), while top-down apical synapses are predominantly driving neural activity during the Sleep phase (when α is high; Figure 1) [40]; this is in agreement with several recent theoretical studies that have proposed that apical dendrites could serve as a site for integrating top-down learning signals [41, 42, 43, 44, 45, 46], particularly those which propose that the top-down signal corresponds to a predictive or generative model of neural activity [47, 48]. The idea that apical dendritic influence increases during sleep (or replay) in cortex has been only partially supported by experiments [49], and is one critical testable prediction of our model. Notably, 5-HT2a receptors are are expressed in the apical dendrites of pyramidal neurons [1] and basal dendrite-targeting parvalbumin inhibitory interneurons [50], and have an excitatory effect that positively modulates glutamatergic transmission [51, 52]. These data suggest that 5-HT2a agonists could have a push-pull effect on cortical pyramidal neurons, increasing the relative influence of apical dendrites and decreasing the relative influence of basal dendrites. Hence, we can model these effects by changing the α value in a Wake-Sleep trained network, and then ask whether the networks exhibit other phenomena that match the known impact of classical psychedelics on neural activity. We note that with this mapping of the Wake-Sleep algorithm to models of basal and apical processing, synaptic modifications at both apical and basal synapses correspond to minimizing a local prediction-error between top-down and bottom-up inputs (see Methods).

Modeling Hallucinations

To see whether a transition from waking to a more dream-like state would induce hallucinatory effects in our model, we trained multilayer neural networks with branched dendritic arbors (see Methods) on the MNIST digits dataset [53] using the Wake-Sleep algorithm and subsequently simulated hallucinatory activity by varying α (see Methods; Eq. (8)). We could visualize the effects of our simulated psychedelic with snapshots of the stimulus layer s at a fixed point in time for various values of α (Figure 2; see Supplemental Materials for videos). As α increased, we observed that network activity gradually deformed away from the ground-truth stimulus in a highly structured way, adding strokes to the original digit that were not originally present. At the highest values of α tested, we found that network states were wholly divorced from the ground-truth stimulus, but retained many characteristics of the MNIST digits on which the network was trained (e.g. smooth strokes and the rough form of digits). These results emphasize that hallucinations induced by a shift to a more dream-like state in these models are heavily influenced by the training dataset, which for an animal would correspond to the statistics of the sensory environment in which it learns its sensory representation. To emphasize this point, we further trained our networks on the CIFAR10 natural images dataset [54] (Figure 2c), to provide an example of a more naturalistic training dataset. In this case, our model was not powerful enough to reproduce realistic natural images—instead, we found that our modeled hallucinatory activity corresponded to ‘ripple’ effects, which are similar to the ‘breathing’ and ‘rippling’ phenomena reported by psychedelic drug users at low doses [2].

Figure 2.

These simulations were produced with a complex, multicompartmental neuron model; however, we found similar results with two alternative network architectures, one with within-layer recurrence (Supplemental Figure S1a) and one which used a simpler single compartment neuron model (Supplemental Figure S1b). We found that our single compartment model produced qualitatively less realistic generated images than the multicompartment and recurrent models, justifying our use of the more complex models (Supplemental Figure S2). To demonstrate the importance of a learned top-down pathway to produce complex, structured hallucinations in the earliest layers of our network, we generated model hallucinations from two control networks: an untrained model and a trained network where psychedelic activity was alternatively modeled by a simple increase in the variance of individual neurons (we will refer to this latter control as the noise-based hallucination protocol). We found that hallucinations under these control conditions resembled additive white noise, rather than structured digit-like shapes (Supplemental Figure S1c-d).

Psychedelic drug users also report observing the emergence of hallucinations while their eyes are closed [2]. Interestingly, we found that our model recapitulated these phenomena: as α increased, networks trained on MNIST gradually began revealing increasingly complex and digit-like patterns (Figure 2b), whereas CIFAR10-trained networks again predominantly produced ‘ripple’ hallucinations (Figure 2d).

Effects of psychedelics on single neurons

Having recapitulated hallucinatory phenomena in stimulus space, we next explored how our proposed mechanism affected neural activity in our network model, in order to establish markers that could be used to experimentally validate or invalidate the oneirogen hypothesis. To start, we investigated the effects of learning and psychedelic drug administration on the activity of single neurons in the model. As noted previously, the learning algorithm used here trains synapses so that top-down inputs to apical dendritic compartments match bottom-up inputs to basal dendritic compartments. As a consequence, we observed that after training, inputs to apical and basal dendritic compartments were much more correlated on the same neuron than they were for random neurons (Figure 3a), which was not observed in untrained models (Supplemental Figure S3a). This form of strongly correlated tuning has been observed in both cortex and the hippocampus [55, 56].

Figure 3.

There are many indicators that psychedelic drug administration in humans and animals can induce marked, long-lasting changes in behavior, as well as large increases in synaptic plasticity [3, 57, 58, 9, 4]. In Wake-Sleep learning, apical synapses learn during the Wake phase, whereas basal synapses learn during the Sleep phase—thus, plasticity at apical synapses is gated by (1− α), whereas plasticity at basal synapses is gated by α (see Methods). However, learning is still theoretically possible without this explicit gating, though it may be noisier and less efficient; furthermore, it is conceivable that classical psychedelics could increase the relative influence of apical inputs on the activity of a neuron without affecting this gating mechanism. As a consequence, we modeled the dose-dependent effects of psychedelics on plasticity both with and without gating (Figure 3b). Consistent with recent experimental results [3], for intermediate doses we found large increases in plasticity at both apical and basal synapses under both conditions, where plasticity was measured as a mean change in normalized synaptic strength across weight parameters in our network (see Methods). In our model, we found that the total evoked plasticity peaked at roughly α = 0.5; we further found that if gating was affected by psychedelics, apical plasticity would eventually be quenched at very high drug doses. We also found that plasticity induced by psychedelic drug administration gradually became unaligned from the weight updates that would have occurred in the absence of the drug (Figure 3c), indicating that these results were not simply due to modulation of the effective learning rate of the underlying plasticity. Rather, as has been suggested by other theoretical studies [59], plasticity in the model likely increased because aberrant hallucinatory activity pulled the learning mechanism out of a local optimum in which plasticity was minimal, producing much more plasticity across the network. Importantly, we observed these increases in plasticity in all network architectures and training datasets we explored, including for our noise-based hallucination protocol (Supplemental S4), demonstrating that changes in apical dendritic influence within a Wake-Sleep learning framework are sufficient, but not necessary to induce increases in synaptic plasticity: for trained networks, it would seem that even simple increases in neural variability can have similar effects.

Effects of psychedelics on neural variability

Having observed that increasing our modeled drug dosage caused heightened fluctuations and deviations from the ground-truth stimulus in the sensory layer of our network (Figure 2), we next investigated whether variability was affected at the level of individual neurons in higher layers of the model. Indeed, we found that for a fixed stimulus, neural variability increased markedly as the simulated psychedelic drug dose increased (Figure 4a). This result is consistent with the data supporting the Entropic Brain Theory [60, 61, 62, 63], in which neural activity in resting state fMRI recordings becomes increasingly ‘entropic’ (i.e. variable) under the influence of psychedelics; however, it is important to note that our noise-based hallucination protocol also produced these effects (Supplemental Figure S5a). Though most experimental data supporting the Entropic Brain Theory is taken from recordings with relatively poor spatial resolution, averaging activity over large cortical areas, our model predicts that this increase in variability should be reflected at the level of individual neurons; this increase in variability after psychedelic administration has been recently observed in auditory cortical neurons for active mice [64], but whether this phenomenon is general across tasks and cortical areas remains to be seen. We further found that this increase in variability corresponded to a decrease in ability to identify the stimulus being presented to the network: we trained a classifier to identify which MNIST digit was presented to our networks on Wake neural activity (see Methods), and found that the accuracy of our classifier decreased (Figure 4b) while the output variability of the classifier increased (Figure 4c) in response to drug administration.

Figure 4.

Within our model, this increase in variability is quite sensible: in the ordinary Wake state, neural activity is constrained to correspond to the singular sensory stimulus being presented, whereas during Sleep states, neural activity is completely unconstrained by any particular sensory stimulus, reflecting instead the full distribution of possible sensory stimuli. As increasing α in our model interpolates between Wake and Sleep states, we can expect intermediate values of α to produce network states which are less constrained by the particular sensory stimulus being presented, reflected in increased neural variability.

Network-level effects of psychedelics

We next investigated the effects of psychedelics on network-level and inter-areal dynamics within our model. We first identified an important negative result: the pairwise correlation structure between neurons was largely preserved across psychedelic doses (Figure 5a-b), as was the effective dimensionality of population activity (Figure 5c). This was sensible, because a network that has been well-trained with the Wake-Sleep algorithm will have the same marginal distribution of network states in the Wake mode as in the Sleep mode—thus, pairwise correlations between neurons should also not differ (as measures of the second order moments of the marginal distribution). We found empirically that even for intermediate values of α in which activity is a mixture of Wake and Sleep modes, these correlations are largely unchanged; in contrast, we observed large changes in correlation structure for untrained networks, and increases in effective dimensionality for both untrained networks and for our simple noise-based hallucination protocol, suggesting that these results are more specific to our trained models in which hallucinations are caused by an increase in apical dendritic influence (Supplemental Figure S6a-b). Interestingly, these results are consistent with a recent study that has shown only minimal functional connectivity and effective dimensionality changes in task-engaged humans being presented audiovisual stimuli under the influence of psilocybin [63].

Figure 5.

However, though the pairwise correlations between single neurons are largely preserved, the causal influence between lower and higher layers of our model network changes considerably both during hallucination and Sleep modes. Because psychedelic drug administration increases influence of apical dendritic inputs on neural activity in our model, we found that silencing apical dendritic activity reduced across-stimulus neural variability more as the psychedelic drug dose increases (Figure 5d). Further, we found that as α increased, inactivating the deepest network layer induced a large reduction in variability in the stimulus layer relative to baseline (Figure 5e), revealing that within our model, increases in top-down influence are responsible for much of the observed stimulus-conditioned variability at larger drug doses. These inactivations had no impact on neural variability in our noise-based hallucination protocol, but were observed for all network architectures and datasets that we tested in which hallucinations were caused by an increase in apical dendritic influence (Supplemental Figure S6), suggesting that these results are quite specific to our model. Further, these inactivations have not yet been performed in animals, and consequently constitute a critical testable prediction of our model.

3 Discussion

Experimental results captured by our model

In this study, we have examined a hypothetical mechanism explaining how the 5-HT2a receptor agonism of classical psychedelics could induce the highly structured hallucinations reported by people who have consumed these drugs. Specifically, we have explored the ‘oneirogen hypothesis’, which postulates that 5-HT2a agonists have the effects that they do because they shift the neocortex to a more dream-like state, wherein activity is more strongly driven by top-down inputs to apical dendrites than normally occurs during waking. To provide a concrete model to explore the ‘oneirogen hypothesis’ we used the classic Wake-Sleep algorithm, which learns by toggling between a Wake phase, where activity is driven by bottom-up sensory inputs, and a Sleep phase, where activity is driven by top-down generative signals. We modeled the ‘oneirogen hypothesis’ by simulating psychedelic administration as an increase in a neuronal state variable that switches neural activity between these two phases (α), such that the simulated psychedelic caused the network to enter a state somewhere between the Wake and Sleep phases, making activity during the Wake phase less tied to actual sensory inputs by increasing the relative influence of the top-down, apical compartment in the models (depending on the “dosage”). This formulation is consistent with anatomical wiring data [39], as well as several recent theoretical studies which propose a specialized learning role for top-down projections to the apical dendrites of pyramidal neurons [41, 42, 43, 44, 45, 46]. It is also consistent with the known cellular mechanism of action of classical psychedelics [1, 51, 52, 8]. Using this model, we were able to produce both stimulus-conditioned and “closed-eye” hallucinations that are consistent with the low-level effects reported by psychedelic drug users [2], and we were also able to recapitulate the large increases in plasticity observed at both apical and basal synapses at moderate psychedelic doses [3].

Our model uses a particular functional form of synaptic plasticity at both apical and basal synapses, reminiscent of the classical delta rule [65], which seeks to minimize a prediction error between inputs in apical and basal synapses. There are many theoretical models of learning that propose similar forms of plasticity [42, 43, 47], so while this plasticity is a necessary prediction of our model, it is not sufficient to validate it. Experimentally, plasticity dynamics which could, theoretically, minimize such a prediction error have been observed in cortex [66, 67], and it has also been proposed that behavioral timescale plasticity in the hippocampus could subserve a similar function [68]. We found that plasticity rules of this kind induce strong correlations between inputs to the apical and basal dendritic compartments of pyramidal neurons, which has been observed in the hippocampus and cortex [55, 56].

Interestingly, we also found that increasing the influence of apical dendrites in the model increased stimulus-conditioned variability in our individual neurons. In cortex, this effect has recently been shown at the level of single auditory neurons [64]; further, there have been numerous studies reporting similar increases in variability (or, analogously, entropy) in resting-state human brain recordings, previously modeled using Entropic Brain Theory [62]. This theory proposes that many of the effects of classical psychedelics on perception and learning can be explained in terms of increases in variability induced by drug administration (e.g. the increase in variability could introduce novel patterns of thinking, or perturb learning to allow it to break out of ‘local minima’). Our results are broadly consistent with this perspective, to which we have added explanatory layers that are both normative and mechanistic [69, 70]: namely, we speculate that this variability under ordinary conditions results from an ethologically important mechanism underlying generative replay for unsupervised learning during sleep or quiescence, and we propose that mechanistically this increase in variability is caused by the increased influence of top-down synapses that are not tied to incoming sensory stimuli.

Testable predictions

While our results are broadly consistent with existing experimental evidence, there are many unconfirmed aspects of our model which could be tested to validate or invalidate it (summarized in Table 1). As mentioned in the previous section, our model predicts that single neurons should increase variability in response to psychedelic drug administration in any cortical area affected by psychedelic drugs. Second, we propose that psychedelic drugs should not push network dynamics into wildly different operating regimes than normal wakefulness, beyond any differences observed between wakefulness and replay (dream) during sleep. In particular, we found that our simulated psychedelic drug administration did not perturb pairwise correlations between neurons within local circuits when averaged across an ecologically representative set of stimuli.

Table 1:

Within our model, psychedelic drug administration is expected to increase the relative influence of top-down projections. This could be explored experimentally in several ways: first, we have shown that apical dendrite-targeted silencing experiments can identify the amount of influence apical dendritic inputs exert on neuronal dynamics; second, we have shown that increases in top-down influence can in principle be identified with interareal silencing experiments. We caution that interpreting results in this second vein may be difficult, as establishing a clean distinction between a ‘higher order’ and ‘lower order’ cortical area may be much more difficult in a densely recurrent system such as the brain, compared to our simplified and fully observable network model. Interestingly, if psychedelic drugs are genuinely exploiting circuitry ordinarily reserved for generative replay during periods of offline quiescence or sleep, we would expect that the same changes in functional connectivity observed during psychedelic drug administration would also occur during periods of replay. In the hippocampus, periods of replay have been tied to increases in apical dendritic influence [71], and increases in the strength of apical inputs have also been observed during NREM cortical updates in prefrontal cortex [49], though in this latter case it is unclear whether apical inputs affected the firing properties of the cell: thus, though there is some circumstantial evidence supporting the idea, much work remains to assess whether apical dendritic inputs dominate pyramidal neuron activity during replay or dreaming in cortex.

Though we have provided a candidate explanation for several of the effects of psychedelic drugs, our model rests on a number of testable assumptions. Our goal has been to articulate these assumptions as clearly as possible, to facilitate experimental efforts to test them.

Comparisons to alternative models

Aside from our model, there are two prominent existing hypotheses as to how psychedelic drugs could induce hallucinations in neural networks. The first proposed that incredibly complex, geometric patterns formed by DMT administration could be attributed to pattern-formation effects in visual cortex caused by a disruption of the balance between excitation and inhibition in locally-coupled topographic recurrent neural networks [72, 73]. Our work differs from this approach in several respects. First, rather than disrupting E-I balance, we propose that psychedelics increase the relative influence of apical dendrites and top-down projections on the dynamics of neural activity. Second, though their model is able to generate geometric patterns, it is not able to generate patterns that are statistically related to the features of the sensory environment (e.g. MNIST digits). Lastly, for simplicity we avoided including topographic (or convolutional) recurrent connectivity in our model; however, it would be a very fruitful direction for future research to extend our work to generative modeling of temporal video sequences, as in [74, 75]. With such a development, it is conceivable that our model could directly generalize these pattern formation-based approaches.

Perhaps more closely related to our model is the ‘relaxed beliefs under psychedelics’ (REBUS) model, which proposes to explain the effects of classical psychedelics in terms of predictive coding theory [60]. Similar to the Wake-Sleep algorithm, predictive coding theory [76] models sensory representation learning with neural dynamics and local synaptic modifications that collectively optimize an ELBO objective function. However, at a mechanistic level, there are numerous differences, the most easily distinguishable feature being that the Wake-Sleep algorithm requires periods of offline ‘generative replay’ to train bottom-up synapses in its network, whereas predictive coding learning occurs concomitantly with stimulus presentation. Furthermore, the REBUS model of psychedelic effects is described at a computational level, in terms of a decrease in the ‘precision-weighting of top-down priors.’ While it is more difficult to map the REBUS model directly onto cortical microcircuitry, and the hallucinatory effects of such a model have, to our knowledge, not been directly analyzed, it has been shown that the proposed mechanism causes an increase in bottom-up information flow between cortical areas [77], in direct contrast to the effects that we have shown in our model (Figure 5c-d). However, because genuinely causal interareal information flow can be difficult to analyze due to dense recurrent connectivity, we stress that it would be easier to distinguish between the REBUS model and our ‘oneirogen hypothesis’ by exploring whether psychedelic drugs affect the same circuitry that induces ‘generative replay’ during periods of sleep and quiescence.

Lastly, it should be noted that the Wake-Sleep algorithm and our choice of network architecture constitute one particular model within a family of related models, all of which satisfy our key criteria for a good model of the ‘oneirogen hypothesis,’ namely that 1) the model has well-defined top-down and bottom-up pathways, 2) it learns a generative model of incoming sensory inputs, and 3) it uses periods of offline replay for learning through local synaptic plasticity. For example, in the Supplemental Materials, we have replicated all of our essential results for two alternative network architectures, also learned via the Wake-Sleep algorithm: one model uses within-layer recurrence to improve generative performance, while the other model uses a simpler single compartment neuron model. Furthermore, the closely related Contrastive Divergence learning algorithm for Boltzmann Machines [78] also involves alternations between Wake and generative Sleep phases and learns through local synaptic plasticity, though Boltmzann machines are computationally more cumbersome to train and require more non-biological network features than the Wake-Sleep algorithm. We feel as though it is important to recognize that models that satisfy these three criteria are more similar than they are different, and that it may be quite difficult to experimentally distinguish between them.

Limitations

While our model is capable of capturing several effects of classical psychedelics, it also has several clear limitations. First, and most notably, our top-down generative model does not have sufficient expressive power to induce complex hallucinations of naturalistic stimuli (though it does a better job of modeling MNIST digits), producing instead ‘ripples,’ or ‘breathing’ effects. While psychedelic drug users do report these phenomena, they also report observing much more complex figures, including people, animals, and complex scenes [79, 80]. Generative models trained through backpropagation have been much more successful in producing more complex generated sensory stimuli [31, 30, 81], and even model hallucinations [82], but, as is well known, backpropagation is itself not a good model of learning in the brain [83]. This suggests that while it is quite possible for generative modeling approaches to produce complex hallucinations through non-biological means, algorithmic or architectural improvements may be necessary in order to make the performance of the more plausible Wake-Sleep algorithm closer to that achieved by backpropagation.

Our model also oversimplifies several aspects of biology. In particular, we do not use neurons that respect Dale’s law [84, 85], and the majority of our efforts to map the Wake-Sleep algorithm onto biology focus on excitatory pyramidal neurons. Furthermore, though we do observe that neural dynamics can tolerate a significant amount of top-down input before disrupting perception, experiments and theoretical studies have shown that inputs to apical dendrites of pyramidal neurons do play an important role in waking perception [39, 86, 87], and are not just learning signals. We focused on clear distinctions between basally-driven Wake modes and apically-driven Sleep modes during training for computational efficiency reasons, and also due to the fact that parameter sharing across inference and generative networks in the Wake-Sleep algorithm is theoretically under-explored (though it is supported in closely related predictive coding approaches [76] and Boltzmann machines [78]).

Lastly, our modeling focus has been exclusively on cortical plasticity and hallucination effects: it should be noted that our model has little bearing on other important features of the psychedelic experience of potential therapeutic relevance, because we have not included the effects of psychedelics on subcortical structures including the serotonergic system [88], which plays an important role regulating mood and may be where psychedelics exert some of their antidepressant effects.

Conclusions

Here we have proposed a hypothesis for the mechanism of action of psychedelic drugs in terms of its excitatory effects on the apical dendrites of pyramidal neurons, which we propose pushes network dynamics into a state normally reserved for offline replay and learning; we have also proposed a number of testable predictions which could be used to validate or invalidate our hypothesis. If validated, our model would describe a mechanism by which the psychedelic experience causes ordinary sensory perception to become literally more dream-like; it further suggests that the plasticity increases observed during both sleep and psychedelic experience could occur via a common mechanism dedicated to sensory representation learning in the brain.

4 Methods

Model architecture and training

To model the effects of psychedelics on neural network dynamics and plasticity, we first constructed a simple model of the early visual system by training neural networks on two different image datasets (MNIST [53] and CIFAR10 [54]). Networks were trained with the Wake-Sleep algorithm [29], which requires, for each layer, two modes of stochastic network activity: a ‘generative mode’, and an ‘inference mode’. For the ‘inference’ mode we must specify a probability distribution b(r^(l) | r^(l−1)), while for the ‘generative’ mode we must specify a separate distribution p(r^(l) | r^(l+1)). As a notational convention, we will use letters when referring to mathematical objects from the generative, top-down distribution, and their vertical reflection when referring to the inference, bottom-up distribution (e.g. p and b). Notice here that activity in ‘inference’ mode is conditioned on ‘bottom-up’ network states (r^(l−1)), while activity in generative mode is conditioned on ‘top-down’ network states (r^(l+1)) (Figure 1a).

The ‘inference mode’ specifies a probability distribution over neural activity, conditioned on the next-lower layer (where the lowest layer is the stimulus layer, i.e. r⁽⁰⁾ = s)—mechanistically it corresponds to activity generated by feedforward projections. To increase the expressive power of our neural units, we use multicompartmental neuron models similar to [89] with N_d dendritic compartments, whose voltages are summed nonlinearly to form the full input to the basal dendrites. For l > 0, layer activity is sampled from the distribution [ineq,where for neuron i in layer l, h_i(r^(l−1)) is given by:

where is a 1× 𝒩 ^(l−1) matrix of synaptic weights onto dendrite n, c_in is the corresponding bias for the nth dendritic compartment, is the strictly positive weight given to the nth dendritic branch (roughly corresponding to a conductance), and is the bias for the entire basal compartment. ϕ_d(·) and ϕ(·) are nonlinearities for the dendritic branches and the total basal compartment, respectively: both are the sequential composition of the tanh nonlinearity, followed by batch normalization [90]. For the dendritic branch nonlinearities, we allow for learnable affine parameters (scale and bias), but for the entire basal dendritic compartment we constrain activity to be zero-mean and unit variance across batches in order to prevent indeterminacy between apical and basal scale parameters. For the final inference layer r^(L), as in the variational autoencoder [30], we parameterize both the mean and a diagonal covariance matrix of the inference distribution: r^(L) ∼ 𝒩 (h(r^(L−1)), diag(h₂(r^(L−1)))), where h₂(·) is also a multicompartmental model, in this case replacing the final batch normalization with an exponential nonlinearity to ensure positivity.

The ‘generative’ mode specifies a probability distribution over neural activity, conditioned on the next-higher layer—it corresponds mechanistically to activity generated by feedback projections. The highest layer, r^(L) is sampled from an N ^(L)-dimensional independent standard normal distribution, r^(L) ∼ N (0, I), and all subsequent layers are sampled from the distribution , where for the ith neuron, μ_i(r^(l+1)) is given by:

where is a 1 × N ^(l+1) matrix of synaptic weights onto apical dendritic branch n, is the corresponding bias for the nth dendritic compartment, is the strictly positive weight given to the nth dendritic branch, and is the bias for the entire apical compartment. Again, ϕ_d(·) and ϕ(·) are nonlinearities, identical to the inference (basal) pathway.

While the neuron model used here is more complicated than is normally used for single-unit neuron models, functions of this kind could feasibly be implemented by nonlinear dendritic computations [89]; we further found that using this nonlinearity qualitatively improved generative performance (Supplemental Figure S2). Given these parameterized probability distributions, we then determined the neural activity for each layer l according to Eq. 1. Our network trained on MNIST was composed of 3 layers, with widths [32, 16, 6], listed in ascending order. A full list of network hyperparameters for both our MNIST and CIFAR10-trained networks can be found in the Supplemental Methods.

All synaptic weights and parameters in our networks were trained via the Wake-Sleep algorithm [29], which is known to produce ‘local’ parameter updates for a wide range of neuron models (and rate or spike-based output distributions), though the specific functional form of the update may vary depending on the neuron model chosen [91]. These updates, for reasonable choices of neural network architecture, can be interpreted as predictions for how synaptic plasticity should look in the brain, if learning were really occurring via the Wake-Sleep algorithm or some approximation thereof.

Consider a generic inference (basal dendrite) parameter for neuron i, . The Wake-Sleep algorithm gives the following update, for a single stimulus presentation:

where η is a learning rate, and the gate α ensures that learning only occurs during sleep mode. Further, for reasons of computational efficiency, we average weight updates across a batch of 512 stimulus presentations; similar results could in principal be obtained with purely online updates [92], but we opted to present stimuli in batches here in order to parallelize computations. changes depending on the parameter θ, reflecting that particular parameter’s contribution to basal dendritic activity. For a dendritic branch weight we have:

where is the total input to the basal dendritic compartment, and is the total input to the nth dendritic branch. This update has the functional form of a classical ‘delta’ learning rule [65], where a compartmental prediction error between local dendritic activity and neuronal firing rate is multiplicatively combined with branch-specific input to provide changes in the conductance for the nth branch. Similarly, for the jth synapse on the nth dendritic branch, , we have:

Unlike for simple one-compartment neuron models, computation of parameter updates for dendritic synapses requires weighting the ‘delta’ error by the conductance of the corresponding dendritic branch (w_in), which could be approximated by the passive diffusion of signaling molecules from the principal basal dendritic compartment back along dendritic branches to individual synapses.

For generative parameters , we have a nearly identical update for a single stimulus presentation:

where now input in the apical dendritic compartment, μ_i(r^(l+1)), is being compared to the activity of the neuron as a whole to determine the magnitude and sign of plasticity. The (1 − α) gate in this case ensures that plasticity only occurs during the Wake mode.

We provide pseudocode (Algorithm 1) for our Wake-Sleep implementation, as well as a full list of algorithm and optimizer hyperparameters (Tables S1 and S2) in the Supplemental Materials. Code for reproducing all results from this study is available here: https://github.com/colinbredenberg/oneirogen-hypothesis.

Modeling Hallucinations

During training, neural network activity is either dominated entirely by bottom-up inputs (Wake, α = 0) or by top-down inputs (Sleep, α = 1). As a consequence, sampling neural activity is computationally low-cost, and can be performed in a single time step. During Wake, one can take a sampled stimulus variable s, determine the activity at layer 1, then 2, and so on until layer L, while during Sleep, one can sample a latent network state in layer L and traverse the layers in reverse order, down to the stimulus layer. However, this is not possible if α ∉ {0, 1}, because activity in each layer l should depend simultaneously on layer l + 1 and layer l − 1. For this reason, we chose to model hallucinatory neural activity dynamically, as follows:

where τ is a time constant that determines how much of the previous network state is retained, and η_t−1 ∼ 𝒩 (0, I). Critically, if we take τ = 1 these dynamics reduce to the sampling procedure used during training (Eq. 1). A priori, the choice of interpolation function f (a, b, α) is arbitrary. We selected the following function:

where κ = 0.35 is a free parameter. This function is equivalent to linear interpolation as κ → ∞, and is equivalent to the maximum function between arguments a and b as κ → 0 if α = 0.5. By selecting κ = 0.35, we are biasing the system towards registering positive inputs from apical or basal sources (in the inclusive sense). We found that this produced ‘hallucinatory’ percepts in stimulus space that did not reduce the intensity of input stimuli as α increased; rather, inputs maintained their intensity and hallucinations were added on top if they were of greater intensity than the ground-truth image. All simulations were run for 800 timesteps, with τ = 0.1. As a control, we compared our results to network dynamics produced purely by increases in noise, without increases in apical dendritic influence (which we refer to as our noise-based hallucination protocol). For these control simulations, we produced network activity time series with the following equation:

so that the standard deviation of the injected noise increased linearly with α.

Apical and Basal Alignment

To measure the alignment between inputs in the apical and basal dendritic compartments of our model neurons, we computed the ‘Wake’ neural responses to the full test dataset and measured the activity in both the basal and apical compartments of our neurons (h(r^(l−1)) and μ(r^(l+1)), respectively). We then calculated the correlation coefficient between apical and basal compartments for the same neuron, compared to the correlation between compartments for two randomly selected neurons.

Quantifying plasticity

To quantify the total amount of plasticity induced in our model system by the administration of psychedelic drugs, we measured the change in relative parameter strength (averaging across all synapses in the network and an ensemble of 512 test images). For each test image, we simulated network dynamics according to Eq. 8. Subsequently, for each parameter θ, we calculated the net amount of plasticity induced by viewing all test images, Δθ. We subsequently reported the relative change:

under conditions in which α values gate plasticity (as in ordinary Wake-Sleep) and under conditions in which psychedelic drug administration does not also affect plasticity gating. Here, we took ϵ = 10⁻² to avoid numerical instabilities.

Classifier training

As we trained our neural network using the Wake-Sleep algorithm, we simultaneously trained a separate classifier network based on Wake-phase neural activity in the second network layer on a cross-entropy loss, to identify the stimulus class of the input to the system. For our classifier, we used a multilayer perceptron neural network with a single 256-unit hidden layer and tanh(·) nonlinearities.

We then quantified the accuracy of the classifier on the test set, based on neural activity drawn from the final time step T of hallucination simulations with various values of α. We further measured the average variance of the 10-dimensional output logits of the neural network.

Quantifying correlation matrix similarity before and after psychedelics

To quantify how similar the pairwise correlations between neurons in our model networks were before and after the administration of psychedelics, we recorded hallucinatory network dynamics for an ensemble of 512 test images, and measured pairwise correlations between neurons in the first network layer. To compare these matrices, we then report the correlation coefficient between the flattened N × N matrices. For this metric, a value of 1 indicates that the correlation matrices are perfectly aligned, while a value of -1 indicates that pairwise correlations are fully inverted.

Quantifying interareal causality through inactivations

To quantify changes in interareal functional connectivity induced by psychedelics, we performed two different types of inactivation. In the first, we inactivated the apical dendritic compartments of all neurons in the stimulus layer, and measured how this inactivation affected across-stimulus variability of neurons relative to the fully active state. In the second method, we inactivated all neurons in the deepest layer, and measured the same effect in across-stimulus variability in the stimulus layer. For both inactivation schemes, we report the mean and standard error of the variance ratio:

where we added ϵ_v = 10⁻³ to the denominator to prevent numerical instability and to the numerator ensure that the ratio evaluates to 1 if the two variances are equivalent.

6 Ethics Declarations

Psychedelic drug research has a long history fraught with many instances of unethical research practice [93]. Further, psychedelic drug use itself has long been stigmatized and punished through legal measures [94], often at the expense of indigenous peoples who have incorporated psychoactive substances into their cultural and spiritual practices for millennia [5]. In the interest of avoiding a repetition of past mistakes, we feel compelled to provide explicit guidance on how our work should be interpreted and used. To do so, we will take inspiration from two principal ethical frameworks: the Montreal Declaration on Responsible AI [95], and the EQUIP framework for equity-oriented healthcare [96, 97]. We strongly encourage anyone considering extending our research or using our work in any form of clinical setting to ensure that subsequent research adheres to these frameworks.

Below, drawing from these ethical frameworks, we will provide a set of guidelines for how our work should be interpreted and used. Though these guidelines are by no means exhaustive, our hope is that adherence to them will help promote the potential positive outcomes of our work while limiting potential negative consequences.

Guidelines for the ethical use of this study

Do:

1. Ensure that the elements of our hypothesis have been adequately tested, as outlined in our discussion, before using our framework in any form of clinical or therapeutic setting.

2. Use our ideas to inform further basic neuroscience research on perception, learning, sleep, and replay phenomena.

3. Explore our ideas as an opportunity to inform your own understanding of cognition, learning, and perception, with the understanding that these ideas have not yet been validated experimentally.

4. Feel free to ask us if you are worried that your proposed use of our work may have negative impacts.

Do not:

1. Report our results as scientific fact. We have outlined a hypothesis, which is designed to be tested by the experimental neuroscience community.

2. Cite or interpret our results without an adequate understanding of the mathematics involved. Feel free to ask us if you are worried that you may be misinterpreting our results.

3. Use our results to extract undue or inequitable profit. The ideas developed in this paper are the product of decades of research and public funding, built upon millennia of exploration of psychedelics. Any knowledge or value contained within this paper is the common heritage of all humanity, with particular recognition due to the indigenous and marginalized communities that have historically suffered and are currently suffering from oppressive government and industry policies.

4. Use our results for any application that could violate human rights or harm human beings in any way.

A Supplementary Materials

Recurrent network model

To explore the extent to which our results hold for different neuron models, and to give our generative model more expressive power than the traditional Helmholtz machine [98], we constructed a network model with a single timestep of within-layer recurrent denoising in each layer, which gives our model some similarities to denoising diffusion approaches [99]. For both our ‘inference’ mode and our ‘generative’ mode we specify both a denoised network state and a noise-corrupted network state r^(l) for layer l; specifying a neural network model is then equivalent to specifying, for each layer, a joint probability distribution over denoised and noise-corrupted network states for both the inference and generative modes, i.e. for the ‘inference’ mode we must specify a probability distribution , while for the ‘generative’ mode we must specify a separate distribution . As a notational convention, we will use letters when referring to mathematical objects from the generative, top-down distribution, and their vertical reflection when referring to the inference, bottom-up distribution (e.g. p and b). Notice here that activity in ‘inference’ mode is conditioned on ‘bottom-up’ network states (r^(l−1)), while activity in generative mode is conditioned on ‘top-down’ network states (Figure 1a).

The ‘inference mode’ specifies a probability distribution over neural activity, conditioned on the next-lower layer (where the lowest layer is the stimulus layer, i.e. r⁽⁰⁾ = s)—mechanistically it corresponds to activity generated by feedforward projections. For l > 0, layer activity is sampled from the distribution , where h(r^(l−1)) is given by:

Subsequently, we add additional noise to get a noise-corrupted network state ; while noise corruption is a natural feature of network dynamics in the brain [100], we include it here in our model because it has been shown that denoising is a critical aspect of many powerful generative modeling approaches [101, 102, 103], and we have likewise found that it improves the quality of generated images in our learned networks (Supplemental Figure S2).

The ‘generative’ mode specifies a probability distribution over neural activity, conditioned on the next-higher layer—it corresponds mechanistically to activity generated by feedback projections. The highest layer, r^(L) is sampled from an N ^(L)-dimensional independent standard normal distribution, r^(L) ∼ 𝒩 (0, I), and all subsequent layers are sampled from the distribution , where is given by:

where m^(l) is a N^l ×N ^(l+1) weight matrix, and a is a bias term. Subsequently, the network goes through a single timestep of recurrent denoising, so that , where is given by:

where σ(·) is a sigmoid nonlinearity that acts as a gating function similar those used in the LSTM [104] and GRU [105], C₁ and C₂ are N ^(l) × N ^(l) recurrent weight matrices, and c₁ and c₂ are biases. While this is a more complicated nonlinearity than is normally used for single-unit neuron models, functions of this kind could feasibly be implemented by nonlinear dendritic computations [89]; we further found that using this nonlinearity qualitatively improved generative performance. Given these parameterized probability distributions, we then determined the neural activity for each layer l according to Eq. (1). As with our multicompartmental neuron model, inference and generative parameters were updated according to Eqs. (4) and (7), respectively. Recurrent network hyperparameters are available in Table S3.

Simplified neuron model

As a control, we also tested our results using a simplified multilayer perceptron neuron model, which used neither batch normalization nor multiple dendritic branches. For the ‘inference’ mode within the simplified model, for l > 0, layer activity is sampled from the distribution , where for neuron i in layer l, h_i(r^(l−1)) is given by:

where is a 1 × N ^(l−1) matrix of basal synaptic weights onto neuron i, and b_i is the corresponding bias.

The simplified ‘generative’ mode likewise replaces the branched neuron model used in the main text with a multilayer perceptron model. The highest layer, r^(L) is sampled from an N ^(L)-dimensional independent standard normal distribution, r^(L) ∼ 𝒩 (0, I), and all subsequent layers are sampled from the distribution , where for the ith neuron, μ_i(r^(l+1)) is given by:

where is a 1 ×N ^(l+1) matrix of apical synaptic weights onto neuron i, and is the corresponding bias. As with the branched neuron model, inference and generative parameters were updated according to Equations (4) and (7), respectively. For optimization, we used the identical hyperparameters to the multicompartment neuron model (Table S1).

Figure S1:

Figure S2:

Figure S3:

Figure S4:

Figure S5:

Figure S6:

Table S1:

Table S2:

Table S3:

Algorithm 1 Wake-Sleep Pseudocode Algorithm 1 Wake-Sleep Pseudocode

Acknowledgements

We would like to thank members of both G.L. and B.R.’s labs, as well as James M. Shine, Brandon Munn, Christopher Whyte, Veronica Chelu, Jiameng Wu, Matthew Larkum, Santiago Jaramillo, Michael Wehr, Neil Savalia, Alexandra Klein, Sarah Cook, Conor Lane, Anousheh Bakhti-Suroosh, Runchong Wang, Michael Okun, and Jordan O’Byrne for insightful discussions and feedback. This work was supported by: [GL] NSERC Discovery Grant (RGPIN-2018-04821), Canada CIFAR AI Chair Program, Canada Research Chair in Neural Computations and Interfacing (CIHR, tier 2). [BR] NSERC (Discovery Grant: RGPIN-2020-05105; Discovery Accelerator Supplement: RGPAS-2020-00031; Arthur B. McDonald Fellowship: 566355-2022) and CIFAR (Canada AI Chair; Learning in Machine and Brains Fellowship). [CB] is supported in part by the FRQNT Strategic Clusters Program (Centre UNIQUE - Quebec Neuro-AI Research Center). The authors acknowledge the material support of NVIDIA in the form of computational resources.

Additional files

Figure 2a multicompartmental mnist hallucination video

Figure 2c multicompartmental cifar10 hallucination video