Summary

Researchers from the Imperial College in London and the Université de Genève in Geneva have created a model using two generative deep convolutional networks to reproduce psychedelic hallucinations related to DMT, aiming to better understand the effects of DMT on the brain and aid physicians and therapists in planning better treatments.

Abstract

In recent years, psychedelics such as DMT have gained attention for their therapeutic purposes and mental health effects, including highly visual hallucinations. Researchers from the Imperial College in London and the Université de Genève in Geneva have developed a model using generative deep convolutional networks to simulate DMT-induced hallucinations, building upon previous work by Google DeepDream. The study aims to understand the effects of DMT on the brain, particularly its impact on visual processes and the induction of synesthesia. The researchers use StyleGAN, a type of generative adversarial network (GAN), to perturb images and reproduce the generated synesthesia caused by DMT. Although this work is a promising computational approach to understanding DMT's effects on the brain, it is not yet sufficient for a complete representation of DMT's impact on serotonin receptors and conscious/perception expansion.

Bullet points

Researchers from the Imperial College in London and the Université de Genève in Geneva have created a model using two generative deep convolutional networks to reproduce psychedelic hallucinations related to DMT.
The study aims to understand the effects of DMT on the brain, particularly its impact on visual processes and the induction of synesthesia.
The researchers use StyleGAN, a type of generative adversarial network (GAN), to perturb images and reproduce the generated synesthesia caused by DMT.
The study builds upon previous work by Google DeepDream, which also attempted to simulate psychedelic hallucinations using artificial neural networks.
Although this work is a promising computational approach to understanding DMT's effects on the brain, it is not yet sufficient for a complete representation of DMT's impact on serotonin receptors and conscious/perception expansion.
The study could aid physicians and therapists in planning better treatments, especially for major depression, where the scientific evidence for the use of psychedelics is significant.

Computational Ayahuasca: Simulating DMT on Artificial Neural Networks

What can we learn from psychedelics and AI?

In our data and machine-obsessed society, psychedelics together with cannabis are starting to be decriminalized and used for pharmaceutical purposes. There is now even an ETF on pharmaceutical companies related to psilocybin and other psychedelics considered a promising market.

DMT (acronym of N,N-dimethyltryptamine) is the main component of the Ayahuasca brew traditionally used by shamans in South America for therapeutic purposes. Together with the therapeutic purposes and mental health effects, it introduces highly visual hallucinations. Those are generally geometric patterns and saturation of visual spaces. A great lecture about the different types of hallucinations can be found in the video below from the Harvard Psychedelic Society.

Recently, researchers of the Imperial College in London and the Université de Genève in Geneva have created a model using two generative deep convolutional networks to reproduce psychedelic hallucinations related to DMT. The idea is not new, and it has been already attempted in 2015 by Google DeepDream. Although, the authors here elaborate a bit more on the serotonin perturbance and effect of conscious experience.

DMT has a chemical structure very similar to serotonin and for this, it leads to several relevant benefits, especially against depression. However, at the same time, due to its “alien” presence in the brain, visual processes are not working properly. Moreover, additional inputs from other sensory systems are being interpreted as visual (it is a form of induced synesthesia).

Practically, those geometric patterns, which are seen, are the results of normal visual processes being disrupted. This includes persistence of images overlaying, 3D pattern filler being over-applied (as in the movie Inception or Dr.Strange), and difficulty distinguishing edges so objects blend and melt together. Moreover, saccades become asynchronous, and synesthesia occurs, which causes other sensory input to be interpreted as visual.

Why trying to reproduce those disruptions with artificial intelligence?

Deep learning is literally everywhere. We are now quite experienced in using generative adversarial networks (GANs), which is the type of artificial networks that generate images not only classify. There is a growing trend of neurotech companies which are exploring the use of psychedelic for medication use capitalizing on centuries of shaman experiences. Especially for major depression, the scientific evidence is quite big. Having a better idea of what happens to the brain during those powerful treatments will aid physicians and therapists, even if we consider hallucinations as a side effect of the treatment.

Therefore, despite AI and psychedelics come from completely separate domains, they can be used in symbiosis to plan better treatments. Doing an oversimplification, GANs work in the following way:

We have a discriminator, which is a neural network firstly trained by exposing it to thousands of images to learn a specific feature (e.g. trained to find cats in a picture). Then we have a generator which is creating fake images from random values until the discriminator is not able to distinguish the true from the generated images. This model has become very popular, allowing deepfake to go viral, creating pictures of humans and animals which do not exist, and even in the medical field for data augmentation:

Neural Networks and Malaria

What Generative Adversial Networks have to do with mosquitos ?

towardsdatascience.com

A particular case-use of GANs is the “style transfer”, where the content of an image is preserved but the style coming from another image is integrated. E.g. see your picture as painted by Van Gogh. In this case, we perturb an image until the discriminator is not convinced that the perturbed image belongs to a specific style. This is at the core of this study. The researchers aim at answering the question of what happens if given a “traditional” picture seen by the brain, we perturb it, transferring the DMT style? In this way, the GANs are practically reproducing the generated synesthesia given by the DMT perturbing the brain.

Original image, and example output of a style-transfer network with different weights (image from the discussed paper: “Neural network models for DMT-induced visual hallucinations”)

The used architecture is the StyleGAN of Kerras et al. as shown in the following figure.

example architecture StyleGan and the perturbed images (image from the discussed paper: “Neural network models for DMT-induced visual hallucinations”)

The human visual stream is generally considered the key for visual experiences, and this work is promising, especially as a computational approach to reproduce perturbation in the brain due to DMT.

This is certainly not sufficient for a whole DMT and other psychedelics representation, as it does not fully represent all effects of DMT on the serotonin receptors need to be understood. We are also far from a complete grasp of conscious/perception expansion and understanding, but it is a start. If we want to be critical, it is just a revisitation of the GoogleDeepDream using StyleGAN. Nevertheless, now the times are different. With the current open mind-set on this topic, this can be the start of a more serious computational representation of DMT.

References

Schartnert Timmerman “Neural network models for DMT-induced visual hallucinations” Neuroscience of Consciounces 2021

Karras, Laine, and Aila “A style-based generator architecture for generative adversarial networks”. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019, 4401–10.

Davis, Alan K., et al. “5-methoxy-N, N-dimethyltryptamine (5-MeO-DMT) used in a naturalistic group setting is associated with unintended improvements in depression and anxiety.” The American journal of drug and alcohol abuse 45.2 (2019): 161–169.

Connect if you want