I am very pleased with the recent work on psychedelic replications by communities such as the wonderful Psychonaut Wiki and r/replications. There is a lot of great work in the area, a little too much to discuss at length in one post. Keep up the good work!
A recent source of marvelous psychedelic replication techniques has just come into the scene, and from an unlikely source. Of course, we are talking about inceptionism applied to deep belief networks.
First of all, who says these pictures are actually trippy? Is there evidence of that? I intend to fully operationalize the concept of trippiness for the classification of pictures; I believe the question is empirically approachable. In the meantime I will simply point out that a lot of people are talking about the peculiar trippiness of these pictures. To give an example, look at some of the comments on the Google blogpost:
Help! We’ve created AIs more powerful than us, and now we need to feed them hallucinogenic drugs to subdue them…. – Urs
Either somebody has been feeding hallucinogens to Google’s image-recognition neural networks, or computer comprehension is alien! Well, actually, I wonder how this compares to visualizations of how the human brain stores images for pattern-matching purposes. – Stephen
Computers are all on drugs. – Matt
And from the Vice article:
“Its incredible how close it looks to an LSD trip, that is normally so hard to describe.” – corners
There are ongoing discussions in a lot of forums about this right now. Somehow, it seems that these new pictures are hitting a particular component of the psychedelic experience that previous replications have missed or at least not fully captured. What is that?
For the purpose of this post I will use a particular classification of phenomenal effects caused by psychedelics. Specifically, the one proposed by Psychedelic Information Theory. In order to fully grasp the motivation for this classification I highly recommend reading the control interrupt model of psychedelic action. In summary, it seems that there are natural inhibitory processes that prevent features of our current experience to build up over time. Psychedelics are thought to chemically interrupt inhibitory control signals from the cortex, which in turn results in a non-linear interaction between the unmitigated characteristics of your conscious experience. I will explain in a bit how this model provides a good framework for explaining the way recent Google Inceptionist (GI) pictures fit into the broader world of visual psychedelic replication.
But now let’s start with the three classes of hallucinations discussed:
- Entropic hallucinations describe the visual effects of gently pushing one’s eyes as well as the amazing interaction between LSD and strobes
- Eidetic hallucinations are the result of interpreting ambiguous stimuli using high-level concepts
- Erratic hallucinations result from the chaotic binding and over-saturation of sensory modalities, which affect the stability of the global perceptual frame (and probably disrupts the continuity field too)
Zooming into the phenomenology of eidetic hallucinations:
The most commonly reported eidetic hallucinations seen on psychedelics are of people, faces, animals, plants, flowers, spirits, aliens, insects, and other similar archetypes. Eidetic hallucinations can sometimes take the form of entire virtual worlds, spirit dimensions, invisible landscapes, and so on. Eidetics often emerge within a pre-existing entoptic interference pattern that grows in intensity over time to produce more photographic or 3D rendered objects. Eidetics under the influence of psychedelics are most often reported with eyes closed or while sitting motionless in meditative trance. On high doses of psychedelics eidetic hallucinations may materialize with eyes open on any surface, pattern, or texture that’s gazed at for more than a few seconds.*
If you surf the internet looking for replications of psychedelic experiences, you will notice that there are great examples of a wide range of effects, but compelling software-generated images of eidetic hallucinations are rare. The challenge here is the complexity of creating actionable tools that highlight high-level features in pre-existing pictures. Amazingly, people can make successful and stunning pictures with eidetic tones, but this requires a lot of dedication and artistic experience. The mighty human artistic effort is unstoppable, though:
Thanks to this 3-fold classification of psychedelic effects we can isolate the quality of experience that both Dali and the recent GI pics specifically enhance. Of course, the phenomenology of most psychedelic experiences incorporate elements of each of these classes, and the interaction between them is certainly non-trivial. In addition, specific substances may have a larger loading of each type, and signature proportions with peculiar results.
It is also worth mentioning the existence of other classification systems, within and beyond visual phenomenology. For example the subjective effect index of Psychonaut Wiki and even the various circuits proposed by ancient Leary and Dass writings have very worthwhile observations that may come useful in one context or another. For the level of resolution here discussed giving eidetic hallucinations their own class is particularly useful.
How the Inceptionist method and psychedelic experiences work similarly
Here is the core of the explanation for how the Google trippy pictures were made:
In this case we simply feed the network an arbitrary image or photo and let the network analyze the picture. We then pick a layer and ask the network to enhance whatever it detected. Each layer of the network deals with features at a different level of abstraction, so the complexity of features we generate depends on which layer we choose to enhance. For example, lower layers tend to produce strokes or simple ornament-like patterns, because those layers are sensitive to basic features such as edges and their orientations.
In some sense this is basically the same eidetic effect we find in psychedelic experiences. For one reason or another, there are moments during a psychedelic experience in which strong eidetic effects manifest. As if a specific layer (or hierarchy level) of one’s model of reality is chosen for being enhanced and fractally iterated in a scale-free manner. Referencing back the control interrupt model of psychedelic action, we can reason that what is going on involves a reduction in the amount of inhibition that highlighted high-level features receive. Again, this resembles the Inceptionist algorithm:
If we choose higher-level layers, which identify more sophisticated features in images, complex features or even whole objects tend to emerge. Again, we just start with an existing image and give it to our neural net. We ask the network: “Whatever you see there, I want more of it!” This creates a feedback loop: if a cloud looks a little bit like a bird, the network will make it look more like a bird. This in turn will make the network recognize the bird even more strongly on the next pass and so forth, until a highly detailed bird appears, seemingly out of nowhere.
Now, this only really shows a snapshot of a psychedelic experience with a heavy eidetic bent. In actual psychedelic experiences there are other common factors that come into play that influence the experience. First, not only are specific features highlighted, but, on the whole, we could say that there is an increase in the overall amount of sensations experienced together. The overall amplitude of your experience goes up, if that makes sense. In other words, although this is hard to imagine, the overall amount of experience increases relative to baseline. That is not evoked using external stimuli, of course, since the actual change in the intensity of your experience requires direct control interruption. The overall information content globally available in the field of awareness of a person tripping increases in a dose-dependent way.
The second hallmark characteristic of psychedelic experiences, which gives them a powerful edge over current digital techniques, is that the state highlights already salient stimuli. High-level psychedelic pattern recognition seems to be based on attention-modulated saliency enhancement. Let me explain:
Our visual system automatically recognizes salient features in our experience. This is not an exclusive property of visual consciousness, by the way. Here we must notice that awareness and attention are distinct but related aspects of our mind. Awareness happens effortlessly, and its visual variety arises as soon as we open our eyes (within 200 milliseconds therefrom). Even at the level of awareness we see a fast sorting of perceived features by their overall saliency, which is a function both of their intrinsic properties and those relative to every other feature in the awareness field. Attention, which is slower and builds on top of the awareness field, enables a variety of high level cognitive activities to interplay with the features highlighted by awareness. In turn, the overall state of consciousness of a person changes as attention moves the reference point for awareness to bring forth new salient features. Iteratively, these processes allow a mind to surf through states of consciousness.
In summary, awareness creates the marketplace of salient features that compete for attention. As attention is recentered on a new cluster of features, the field of awareness is modified and the new salient features again have a chance to change the focus of attention.
With psychedelic-induced control interruption, the intensity by which saliency of features in the field of awareness is highlighted goes up significantly. In turn, the attention-modulated perception of the intensely salient features highlights specific high-level features suggested by the field of awareness. And finally, this conceptual mental state highlighted via attention, results in an even higher saliency for conceptually-related features. And hence come the eye reality, fish realty, tree reality, abstract concept reality, divine reality, fractal reality, etc. people discover on LSD.
Although a difficult challenge, I predict that a well-trained, dedicated and mentally healthy psychonaut would be able to paint psychedelic experiences of her own that highlight similar high-level features as those highlighted on specific Inceptionist works of art. Probably a long meditative practice would help in the process, since the specific saliency of various features is attention-modulated, and thus requires inhibiting unrelated salient directions (e.g. deep philosophical questions, personal issues, etc.) and focus exclusively on, say, dogs.
Who chooses what is salient?
If you already know what class of features you want to highlight, then the inceptionist method will help you. But what about choosing what to highlight to begin with? This, I believe, is the crux of what makes psychedelic experiences (and minds in general) still unbeatable by neural networks. Once you know what to look for, your cortex and inceptionist methods (and their future incarnations) might be on the same playing field. But what enables you to decide what is worth looking for?
The key unresolved problem standing for a fully-digital psychedelic experience replication algorithm is what I call the saliency-attention mapping. This is: Given a particular conscious experience that is highlighting a set of features, how does attention ultimately find what to focus on? How are the subsequent relevant features to be highlighted? In many cases we choose to ignore all of the immediately salient features in a scene precisely to see more subtle patterns. And during a psychedelic experience, directing your attention to entirely unsuspecting places has the effect of switching off previously salient features and activating a new class of them (for example, choosing to focus on the music rather than the visual scene).
Is there any way of modeling the saliency-attention mapping without taking into account all of the information present in the field of awareness at the time? Indeed, an ongoing hypothesis here in Qualia Computing is that consciousness itself is required for this step. The very computational advantage of being conscious seems to be related to the unitary nature of experiences: Your choices are not only the result of parallel processing or implicit information integration. They stem from what you choose to pay attention to considering the entirety of your field of awareness. You do this at every point in time. Thus, a sort of instantaneous and ontological unity is required to account for a significant step of the information processing pipeline of the mind. And this may lead to a saliency-attention function whose runtime complexity is impossible to match with digital computers.
The conceivability horizon
Now, this unitary field of awareness step also has large down-stream effects. In particular, subsets of the phenomenology of experiences can be reinterpreted in very novel ways. Psychedelics are likewise famous for unlocking entirely new conceptual ontologies and points of view that remain with the person long after acute effects subside. We could call this, an extension of the horizon of conceivability. This comes about from considering many of the features of the particular conscious experience at once and identifying a new private referent (such as a concept) whose meaning is derived from the unique combination of those elements.
Without a unitary conscious experience this step would be impossible, and it remains to be seen for an artificial neural network to accomplish this on its own. For completeness, it is worth mentioning that phenomenal binding also has strong implications for memory. Every time we experience a new situation a new ‘situational snapshot’ is added to the collection of — and network of relationships between — memories that can be triggered with temporal lobe stimulation. Thus, incorporating a human (or whatever implements phenomenal binding) into the loop may be unavoidable.
The future of psychedelic replications and consciousness engineering
Eidetic art is marvelous, and for a long time we didn’t have any idea about how to systematize it in software. Now we have some wonderful examples of a fully scalable approach. Inevitably, we will soon have visual editing software that incorporates neural networks.
Deep belief networks applied to replications will allow us to drastically increase the level of realism of simulated trips. This will thus draw a lot more attention to this fascinating field, and bring engineers, artists and mathematicians onboard. They will have a wonderful synergy in this sphere.
But how practical are these techniques? If you want to find fundamentally new patterns in an image, what should you use… neural networks or LSD? The answer is: why do you have to choose only one? Here is where I casually mention that if you were planning on taking a psychedelic sometime in the future, why not tell us how the trippy images of Google look like during a trippy experience? I bet a lot of people would appreciate your input.
Presumably, incorporating a human in the loop could actually empower these networks to recreate remarkably psychedelic progressions of scenes and features (and high level ideas!). To do so you need to somehow identify what the human finds salient in the picture/video being explored, and how her attention is directed as a consequence of that. Obvious candidates here are eye tracking devices and the general class of bio/neurometrics. More speculatively, endocrine measurements of the chemical markers of saliency and attention may be of tremendous value too. What would this look like? A person hooked to a series of tubes that provide fast feedback using a lab-on-a-chip, and a deep belief network with flexible Inceptionist dynamics guided by the person’s measured center of attention. In case you haven’t noticed, I think that this area of exploration is extremely promising. Go ahead and do it!
Now, if you want to figure out a hard technical problem, currently mild psychedelic experiences are more promising than deep belief networks. This, again, is because the attention-modulated saliency enhancement of psychedelics can allow you to discover, explore and reinterpret the features that matter for a particular problem. Assisted digital exploration, however, may someday surpass the effectiveness of psychedelics, or better yet: A smart combination of techniques –chemical, biological and digital– will incite in the field of consciousness research what the Galilean revolution was to physics. The hands-on collective exploration science needs in order to fully thrive is about to arrive for consciousness. Finally!
Honestly, if the deep dream algorithm could be rendered continuously and recursively in a virtual reality setting based on the locus (center of field of view for simplicity’s sake) of the user’s attention, a more meta unitary conscious structure might be learned or taught from the exercise. Much like the present algorithm teaches recognition through repetition of salient images, this could (hopefully) teach consciousness through repetition of attentive behaviors.
The exercise might look like:
Start with a stock virtual environment populated with a variety of objects>User directs attention by turning VR headset to specific feature of environment>algorithm enhances whole field of view based on feature>User redirects attention>algorithm enhances FOV based on new feature>…
That would hopefully isolate the awareness genesis (here forced by the manipulations to FOV)-attention direction loop to the point that a machine learning algorithm could derive a consciousness algorithm from the whole set of behaviors that resulted from the exercise of a bizarrely focused human consciousness.
Just a thought, anyhow–great read, OP made my day
Its nice to see an in-depth analysis of this. I’m writing up one too on my blog, but it won’t be as extensive as your write up. Thank you for posting this! Best wishes, Don