Deep CANALs: An ML-inspired framework for understanding the therapeutic effects of psychedelics

Arthur Juliani
10 min readMay 20, 2024


(image generated using Stable Diffusion)

Psychedelic-assisted therapy is emerging as a compelling treatment option for mental health conditions ranging from depression and anxiety disorders to substance use and obsessive-compulsive disorders. The apparent ability of psychedelics to treat such a wide range of conditions has led Dr. Robin Carhart-Harris and his collaborators to propose that these drugs may be addressing a single unified underlying cause of these diverse disorders, a phenomena which they refer to as canalization.

Specifically, by canalization they mean the tendency to develop overly-precise beliefs about what we expect to perceive in the world, how we expect to act, and what the outcomes of our action will be. According to this framework (CANAL), psychedelics work therapeutically by disrupting these priors, allowing new and more adaptive ones to be learned instead. This theory can account for many of the findings in the psychedelic research literature, and I have previously written an entire blog post describing their framework, which you can find here.

The CANAL theory is limited though in its ability to predict who might benefit from psychedelic therapy, and who might not. This is a critical feature of any usable clinical framework. In a paper published this year in the journal Neuroscience of Consciousness, my collaborators and I outline these limitations and propose an expanded framework based on principles from machine learning theory. We refer to our model as Deep CANALs. This blog post is designed to give an overview of the ideas we present in more depth in the full paper.

The two novel contributions of our Deep CANALs framework are quite straightforward. The first is that we distinguish between two types of canalization that can happen in the nervous system, each at a different temporal scale and with a different neural substrate. The second is that we identify pathologies of both too much or too little canalization. Both of these elaborations to the original canalization theory allow us to better characterize the expected therapeutic effects of psychedelics, and predict who may benefit the most. Let’s dive in.

A Recurrent Neural Network model of the brain

A reasonable first question to ask about the original CANAL theory is: what is being canalized? In our model there are two unique yet interrelated forms of canalization. To understand this, it is helpful to use a simple metaphor from the world of machine learning. Recurrent neural networks (RNNs) are a popular class of models which are widely used in both the worlds of artificial intelligence and neuroscience. An RNN is a function which takes an input (x) and a hidden state (h) and computes some output (y), along with a new hidden state (h’). The function is parameterized by learnable weights (theta). The hidden state allows the model to keep track of information over time and maintain beliefs about the stream of incoming data.

RNNs are trained by alternating between two distinct phases. First is the inference phase, in which the model is used to recurrently compute a sequence of (y) outputs given a sequence of (x) inputs and a continuously updated hidden state (h). Second is the learning phase, in which backpropagation is used to update the parameters of the model (theta). The hidden state (h) changes every step of inference. We can use an energy landscape to describe this change, which corresponds to a transition function. We can also use an energy landscape to describe the evolution of (theta) over the course of learning. This is typically called the loss function. Notably, the point in the learning landscape that (theta) is currently in will determine the inference landscape that (h) uses.

RNNs have frequently been used to model the function of the brain. The many millions of neurons in the brain communicate with one another via synaptic connections. The strength of these connections change over time as the person learns and adapts to its environment. These synaptic connections are the equivalent of the model parameters (theta) in the RNN, and are updated as the brain learns over time. The synaptic connections between neurons determine what the neuron firing patterns will be over time, which is equivalent to the hidden state (h) of an RNN. The neural activity pattern instantiates the actual set of beliefs, internal representations, or behavioral plans which a person is engaged in at any given moment.

Two types of canalization for two energy landscapes

What we have in the brain then are two separate energy landscapes, one corresponding to inference dynamics and the other to learning dynamics. If either of these landscapes are non-convex, they can lead to suboptimal outcomes, either in the space of specific or possible beliefs and behaviors. In machine learning this is often described by the presence of local minima, steep gradients, and other forms of ill-conditioning. The CANAL theory refers to these optimization pathologies as forms of canalization, and equates them with various psychopathologies.

The Deep CANALs model allows us to distinguish between different kinds of canalization. We can first consider what canalization looks like in the inference landscape. Canalization here refers to the dynamics of neural activity getting stuck in or cycling between specific fixed points or attractors. Concretely this might look like rumination or obsessive-compulsive thought patterns, or in the case of substance-use disorder it might manifest as a craving to use a drug which the person is unable to ignore. For depression it may consist of a particular negative evaluation of self-worth which an individual is unable to escape from.

Canalization in the learning landscape looks somewhat different. Here the issue is not that the brain gets stuck in a specific pattern of neural activity. Instead the issue is that the brain gets stuck in a specific configuration of synaptic connectivity. Practically this means that the person is unable to learn and adapt to changes in their environment. Whatever current mental or behavioral world an individual is inhabiting, they are unable to get out of, even with effort. If a person is already perfectly adapted to their environment, then there is no need to adapt further, and this may not be an issue. If on the other hand a person is trapped in unhealthy and maladaptive cognitive, emotional, or behavioral patterns, then it would be beneficial to be able to adapt and change.

We can more precisely describe these two kinds of canalization using machine learning terminology. Canalization in the inference landscape is equivalent to overfitting, where a model is too rigid and unable to generalize well to the domain it is being evaluated in. Notably this maladaptivity is in the absence of additional learning. In contrast, canalization in the learning landscape is equivalent to plasticity loss, where even with learning a model is unable to adapt to changes. Overfitting is undesirable, but potentially resolvable with learning. Overfitting and plasticity loss together however create a situation which is much more problematic.

Over-canalization and under-canalization

It may be tempting to assume that even if there are two different energy landscapes, canalization should still be looked on as an undesirable property to be eliminated in either cases. Although it is true that there are a number of psychopathologies of too much canalization, there are also pathologies which can arise from too little canalization. In machine learning terminology, these would correspond to underfitting in the case of inference landscapes, and catastrophic forgetting in the case of learning landscapes.

Let’s consider the case of inference first. In order to be well adapted to an environment and act successfully in that environment, we need to be able to form and maintain stable representations or beliefs over time. In this way, canalization can be an aid, because it can enable our well-adapted beliefs and behaviors to be consistent and durable. An example of a failure to maintain stable representations during inference is in attention-deficit disorder, which is characterized by an individual being unable to consistently focus on a given task for the time necessary to complete it.

The same is true of the dynamics of the learning landscape. While it is desirable to be able to change and adapt over time as our world changes around us, we also want to be able to maintain a repertoire of currently successful and adaptive strategies for engaging with the world. If we were to dramatically change our synaptic connectivity in response to only minor changes in the environment, then we risk losing valuable learning which may have taken place in the past. An example of this from the mental health literature would be bipolar disorder, where an individual cycles between very different cognitive and behavior patterns over the course of days or weeks, often in response to only relatively minor changes in their environments.

We can see that the ideal from a long term mental health perspective is to follow the Goldilocks principle. An individual wants neither too much nor too little canalization in either of the landscape types, but rather a balance which enables the maximization of both stability and plasticity over time. Put another way, we want to be both well-adapted to the current environment and adaptable to changes if and when they occur.

These constructs of stability and plasticity have been extensively studied in the psychological literature. The most popular and well studied personality measure is the Big Five. The five traits in this measure can be further grouped into two meta-traits: stability and plasticity. Stability is a grouping over agreeableness, conscientiousness, and inverse neuroticism. Plasticity is a grouping over openness and extraversion. Within our Deep CANALs model, the construct of stability refers to the balance between over-canalization and under-canalization in the inference landscape. This trait is sometimes referred to as meta-stability, as it avoids excess rigidity. Plasticity on the other hand refers to the degree of under-canalization in the learning landscape. A well-tuned nervous system is one that can maximize stability while also maintaining the necessary plasticity to remain adaptive to changes in the world.

Effects of psychedelics on canalization

We can use the Deep CANALs model to start to make sense of the therapeutic effects of psychedelics. In the original CANAL model, it was proposed that psychedelics act to reduce canalization, and thus mitigate a variety of psychopathologies. While this on the whole tends to be true, individuals also occasionally have intense experiences on these substances which can exacerbate certain symptoms, or even trigger psychosis in rare cases. Our expanded model can provide a more nuanced account which takes into consideration these diverse experiences and outcomes.

According to Deep CANALs, psychedelics act to acutely and transiently disrupt the inference landscape. This disruption can manifest in many different ways as a function of the mindset of the individual and the environmental setting they are in. On one extreme they may experience a state of oceanic boundlessness, where boundaries seem to have dissolved and constructs fade away. This would correspond to a clear reduction in the canalization of the inference landscape. On the other extreme, an individual may get stuck in thought-loops which they have no control over, see objects or entities which cannot be accounted for by the visual stimuli their eyes are taking in, or may have intense experiences of insight which are accompanied by a strong sense of confidence. Here we have cases of transient over-canalization manifesting in the inference landscape. Whether this is therapeutic depends heavily on the intentions of the individual and the safety and support of the context in which they have taken the drug.

The effects of psychedelics are more straightforward when it comes to the learning landscape. There is extensive research demonstrating that psychedelics are able to increase neural plasticity across a variety of time scales. Furthermore, psychedelics are able to regrow neurons and connections in a number of different regions of the brain. These effects may be responsible for the experience of “afterglow” which can follow days or even weeks after a psychedelic experience. Collectively, these effects correspond to a decrease in canalization of the learning landscape and an increase in plasticity. While this can be desirable in many (if not most) instances, it may present issues in rare cases in which individuals are already particularly high in plasticity. Negative outcomes in this space might look like triggering a manic or dissociative episode which may persist well after the acute drug effects end. Although this is unlikely to be the outcome from a single session of psychedelic use, it becomes a possible concern for individuals who are frequent users of these substances, particularly in non-therapeutic contexts.


Psychedelics are breaking into the mainstream. We can see this not only in all the clinical research into psychedelic-assisted therapy, but also in the legalization and decriminalization efforts which are gaining support across the United States. As more and more people begin to try psychedelics both therapeutically and recreationally, it becomes increasingly important to understand what these substances are doing to the nervous system. Building on the CANAL model, we developed Deep CANALs as a means of providing a nuanced framework for better understanding both the acute and long-term clinical effects of these substances.

It is clear from the past decade of research that psychedelics have a huge potential to help address many of the societal challenges we are facing around depression, anxiety, and addiction. At the same time, they can be significantly less therapeutic when used in the wrong circumstances and for the wrong reasons. We hope that Deep CANALs can help guide more responsible and beneficial psychedelic use by providing a framework for better understanding how and why psychedelics have the therapeutic effects that they do.

If you are interested in reading our full paper, you can find it here. If you have thoughts, comments, or feedback, please feel free to reach out.



Arthur Juliani

Interested in artificial intelligence, neuroscience, philosophy, psychedelics, and meditation.