Machine learning and pain
This blog by AICoreSpot is with regards to computational and biological reinforcement learning. The role of paining in the context of learning has been a subject that has been researched for several years, and a natural extension of this investigation is the relationship between pain and machine learning.
Most of us have undergone and can provide exhaustive input on the phenomenon of pain, the experience associated with falling off of your motorcycle and smacking your head on the pavement, striking our head on a faucet, or the emotional pain of bereavement.
Pain is common to all humanity. In a cultural context, pain is the main reason individuals seek medical intervention. The treating of chronic pain makes European Healthcare Systems approximately 200 billion Euros in expenditure and within the United States up to 635 billion USD, frequently appearing as the costliest health-related issue to treat. For patients, pain is typically specified as the dominant factor in reduced quality of living and the leading cause behind lost years of productive life.
Provided the consequences of pain, it would appear that pain must have a critical part in biological learning, and this blog explores the concept that pain can serve as a source of inspiration and inform us how we perceive and think about the activity of generating machine learning systems.
Emphasis should be placed on why the research of pain is critical for machine learning. Let’s classify the research of pain as either scientific pain, as it is studied in a medical context, or folk pain, as experienced by individuals in daily experience. Both perspectives are useful for machine learning research. The scientific perspective, which is our primary concern, imparts insights and inspirations into alternative strategies for sensing, generalization, and learning. The folk view provides perspective on how pain is comprehended in a cross-cultural context and how normative perspectives on pain perception are determined and perpetuated. The folk perspective facilitates the cultural and social dimensions of pain to give way to critical questions with regards to ethics, respect, participation, and equity.
There are several themes to look into, with regards to taxonomy and philosophy, of sensing and generalization, of society and science, and cumulatively, these are emergent subjects in the quickly altering landscape of our domain for responsible, fair, ethical, decolonial machine learning research, and to which the research of pain can furnish new insights.
Pain fuels action and behavior that looks to uphold the integrity of the body when exerting its will on the world. Pain can be short-term, or be recurrent throughout an individual’s life span, pain is by its nature subjective, and can also be experienced without a physical stimulus, as is the case with psychological pain. The intricacy of pain makes it tough to give it a simplistic layman’s definition, particularly as even this short informative piece demonstrates that pain requires a way of defining it that is multi-layered, and not one connected to physical stimuli.
The most broadly-used definition from the International Association for the Study of Pain (IASP) defines pain as “an unpleasant sensory and emotional experience connected with actual or potential tissue damage, or described in the context of such damage.
There is, perhaps shockingly, very minimal consensus with regards to what we are speaking in reference to when we bring up the topic of pain. A detailed note follows the medical definition of pain that was just specified, as a supplement.
“Pain is always subjective, it is undoubtedly a feeling in portions of the human body, or brain, but it is always unpleasant, and hence, a deeply emotional experience. Biologists identify that these stimuli which lead to pain are likely to damage tissue, or cause emotional duress in the form of depression, anxiety, and addictions. Several patients indicate pain in the absence of tissue damage or any probable pathophysiological cause and this typically occurs for psychological reasons, and there is typically no way to distinguish their experience from that owing to tissue damage if we consider the subjective report.
These clauses have the outcome of compounding issues.
- For scientists, pain consisting of both a discernible biological reason, however, existing as an essentially subjective state, makes it tough to quantify the phenomenon of pain experiences. The quantification of pain is one of the domain’s holy grails.
- For practitioners, having no way to differentiate between pain of psychological origin and pain of physical origin makes efficient treatment tough.
- Lastly, for patients, ongoing pain conditions and the incapability to sufficiently treat, or describe, their experiences typically results in comorbid mental health conditions that aggravate their existing illnesses.
This hassle of defining pain makes this a good juncture to analyze a critical discourse in the philosophy of pain. Five decades ago, people were querying if we would ever produce a machine that could feel pain as we humans do. In his seminal research, “Why you can’t make a computer that feels pain, Daniel Dennet puts forth a thought experiment regarding developing a machine with a biologically-driven pain system. His work was two-fold, to both generate an improved type of computer system, and to better comprehend the foundational neurobiology of pain states. Provided all the intricacy we just observed – pain states while lacking injuries, and injury without the presence of pain, as well as observing other conditions which were the cause of therapeutic interventions, the researchers ultimately came to the conclusion that this task was impossible. This is due to the fact that a machine couldn’t sense or feel pain, but as there is no singular cogent pain theory to code into it. This begins the tradition known as pain eliminativism which puts forth the argument that as the idea of pain fails to be in reference to anything empirical, we would be better served by eradicating it, deleting it from our usage and rather identifying other vocabularies to function in its place. As we conduct more research with regards to pain in machine learning, we will also get into this debate. Rather, we should regard this an opportunity, pain pinpoints a discrepancy between pain perception states and environmental awareness, and we can ‘rescue’ pain by describing this discrepancy.
Models of pain
As we haven’t mentioned reinforcement learning, you might have conceived in your mind that pain is merely another signal and that is how pain slots into our current medical knowledge. Although natural and relevant, it’s wise to be cautioned against the assumption of this default perspective, as this is one of the many potential perspectives. So, let’s delve a bit deeper into these varying perspectives on pain. There 3 noteworthy models of pain.
The model we’re going to look at first is the sensation model. Within this model, pain is merely the painful feeling, or sensation. Pain is not combined with any representative physical state or stimuli, but is viewed to happen in correlation with them. This is simplistic and intuitive, but not ideal, for the simple reason there appears to be no homogenous pain medically observed. The McGill Pain questionnaire, the most noteworthy subjective pain scale leveraged in healthcare environments, consists of a listing of >50 terms to detect prominent features of an individual’s pain experiences to assist in medical practice, every one of them informing a unique aetiology for pain. Were pain a singular sensory aspect as indicated, such a comprehensive listing of descriptors would not be needed to distinguish both symptoms and underlying cause.
The second model we will be exploring is the representational model, and we are already likely to be aware of this. We can have either sensations, perceptions, or an emotional concept of pain. What is shared between all representational theories is the debate that pain is representative, or an abstraction, or a perceptual facet of one’s surroundings or physiology. Then again, we have proof that deviates from this model. For instance, phantom limb and referred pain, where the physical location the pain is intended to indicate either is not present, due to amputation, for example, or is only indirect, such as with heart problems, indicates that pain is not definitively representative of a physical facet in our physiology or surroundings.
The third model is the motivational model. In it, pain is a request or an instruction to safeguard portions of your body. Again, this is restrictive. Individuals typically feel pain beyond the point where any behavioral intervention impacts the initial reason. There are also instances of pains that are actively sought out, either via treatments, like acupuncture, or in masochistic tendencies, which appear to be contrary to the concept that pain is merely a motivator that deters specific behaviors.
At this juncture, we want to leave you with food for thought, a rather critical point. We’d like to put forth the rather hotly debated claim, emulating the contention by Ann-Sophie Barwich, that “theories of perception suffer from one basic flaw: they are theories of vision.” So, with regards to machine learning, have we got into the same trap, overfitting to our comprehension of vision? It appears that we too have connected a lot of how we contemplate about learning to the identification of perceptual objects that are then connected to states, which then provide feedback on actions. Therefore, our critical proposal is to alter our frame of states, which then provides feedback on behavior. So our important proposal is to alter our frame of perspective to one of perception as situational assessment. Through this, we mean that over concentrating on perception as the identification of perceptual objects, and as operating as independent systems, we can look at perception as unanimously bringing together several sources of contextual data to give shape to final perceptual states. This does not differ from how we view perception algorithmically, but that this perspective is unique in that it enables for painful scenarios to be described and enables us to furnish an object-less account of sensory states. Like several other experiences, pain is a multimodal strategy to perception and can be a part of learning without the need for an association with discrete awareness of your surroundings.
Therefore, what we require is a distributed situational assessment framework of pain to assist us in informing new methods of learning. Let’s take a brief glance at two proposals that will be natural with regards to machine learning, pain as reward, and pain as inference.
The pain as inference model regards pain as Bayesian application of learning, where an agent does not have overt awareness of their inherent states and surroundings, and must therefore make inferences about final states based on vague and incomplete data. Pain is viewed as an active predictor of upcoming bodily states, in addition to being an evaluator of present afferent data, and the Bayesian updating transforms multimodal historical experiences into future evaluations.
A second perspective adopts our first guess with regards to pain, viewing pain as an internal reward indicator or a generic variant of cumulant that is leveraged within a risk-averse intrinsic motivation framework. This model features integration of sensory and nociceptive information. Assessment is not with regards to the particulars of a physical object, within the surroundings, but regarding the scenario via which the afferent data was obtained. The outcome pain signal produced in the primary teaching signal that leads to the formation of value and behaviors.
These perspectives are not complete in a few ways and don’t yet have all the specifics worked out. We can look deeper by exploring three spheres of pain learning, single exposure pain learning, generalization of pain experiences to novel stimulus, and the capacity to socially transfer obtained pain knowledge. All three of them are learning capabilities we have with other perceptual frameworks, but when taken in conjunction, leveraging a perspective of situational assessment, pain demonstrates that we can furnish a view on these strategies towards learning that is not connected to object recognition in the immediate surroundings.
There is another second critical point of consideration. It is that ML can be a utility with which to furnish some of the information that is presently absent. Leveraging Marr’s levels of analysis as a strategy for contemplating about these issues. To save up our time, let’s talk about the case regarding single-exposure learning.
At the computational level, the capability to combine pain perception to injurious stimuli in a swift and efficient fashion is one of the main facets of pain systems. There is new proof demonstrating that drosophila, leveraging odor and food related conditioning, illustrates single trial learning for both punishment and reward. With regards to rewards, we observe learning when an odor is combined with a fructose reward. With regards to punishments, we also observe learning when an odor is combined with a highly concentrated salt solution. And fascinatingly, single-trial learning was not illustrated with regards to adversive quinine solutions.
The capacity to learn from a singular exposure of salt solution over quinine is seen as an outcome of the activation of a multimodal pain framework via micro-wounds, placing emphasis on modality particular pain learning through punishment, over any and all negative or adverse stimuli.
As you would have probably had the experience yourself, the swift onset of pain learning after a singular exposure also happens in human beings, and can give lifelong adaptive behaviors. When swift encoding of the stimuli in your surroundings happens with pain, it’s usually also brings up a fear and emotional reaction, therefore, enlisting structures of the memory system. This swift-onset learning is a critical survival mechanism and is a critical facet of pain behavior throughout the animal kingdom.
At the implementation level, one critical element to swift encoding of learned pain experiences is achieved via the recruiting of the hippocampus. Stimulus that present suddenly with pain outcome is a larger activation within the hippocampus, amongst other areas. The activation of this system in the hippocampus is viewed to underscore its part as a situational comparator, where real states are contrasted to forecasted states, with errors having the outcome of encoded memory creation. Also connected to pain learning is an averse or fear circuitry of the amygdala-striatal system and also the lateralized consummatory system. This all indicates to a comprehensive knowledge base of pain neurobiology we have at our disposal to function with.
However, while we have made advancements describing the computational and the implementing levels, we’ve not yet detailed the algorithmic level, as what constitutes the algorithmic level with regards to these pain issues is still absent. Assisting to complete this absent algorithmic level is where there are several contributions from machine learning to make.
This brings us to the conclusion of our blog post. Our goal is to provide you with two primary takeaways. Starting with you obtaining a brief perspective into the domain of pain research and the several dimensions it takes, be it in social, technical, or sociotechnical fields. Then, you will observe the opportunity for new investigation into both biologic and computational learning by keying in the absent algorithmic level, leveraging the concept of situational assessments as a guideline.
The human pain framework is very refined, skilled learning apparatus critical to our existence as a species, in addition to being a mediator of so many complicated social and cultural conventions. People without pain systems typically have short lives, without the capacity to detect critical protective data about their surroundings. And this framework not working correctly is one of the main healthcare burdens felt globally. Regardless of the criticality of pain to human existence, ML has currently fallen short of leveraging pain as a source of algorithmic inspiration, regardless of the protracted convention of mutual exchange amongst neuroscience and machine learning. This has been a concept that researchers have been curious to explore for a long time.