Why the Brain is a Prediction Machine
Prediction shapes the way we perceive and understand the world -this view has become more and more influential in the system neuroscience community, and it also provides a framework for us to understand neuropsychiatric disorders -in general, A priori information will affect our perception and perception, and this mechanism is disturbed in people with mental disorders. This has resulted in patients being unable to form an optimal model that characterizes the world.
More and more people conceptualize the impact of predictions as a "top-down" process-from the perspective of information processing, predictions that originate in (cortical) higher-level areas affect areas that are located at a lower level. However, this view does not take into account the prediction information generated in the process of "bottom-up" information processing. As this article will reveal, it is extremely important to distinguish between "bottom-up" and "top-down" prediction processes. This distinction is also the key to constructing a predictive processing framework, which will help us understand how the disorders in the predictive processing process cause the symptoms and experiences of mental disorders.
Life uses sensory input to discover surrounding structures, thereby creating a representation of the environment. Such a characterization is a doer (Agent) * key regulatory interaction with the outside world. However, the sensory input is ambiguous and full of noise. Therefore, we believe that an accurate representation of the environment also requires a priori information. The predictions provided by the a priori information may eliminate the blurring of sensory signals and allow one to speculate on the external stimuli that lead to sensory input. According to this prediction processing framework, making predictions based on a priori information about the outside world is a key function of the brain.
Just as people do in the field of reinforcement learning and artificial intelligence, in the field of cybernetics, we also subdivide the system—that is, a unified whole composed of a set of interactive components—into many actors and their environments. How we should distinguish between the system, the actors and the environment depends on the specific problem. For example, the system may be a fish (activator ) responding to vortices and water currents (environment). In addition, the system can also be a part of the fish's skeletal organ (activator), which is active throughout the body (environment) . Or, we can think of the water flow and all living things in it as a system.
Cyberneticists consider how the agent interacts with the environment to maintain internal stability when facing changes in the environment. Early cyberneticians realized that this hypothesis could also be used to structure brain research-the the best response of the actor to environmental disturbances is proactive rather than reactive; that is, prediction Is necessary. One of the pioneers in this field, William-Ross- Ashby (WR Ashby), and his colleagues proposed two theorems to illustrate how actors can successfully resist environmental disturbances:
Since the environment is ever-changing, so must the actors. In the field of control theory, the variability (Variety) refers to a system (environment or actors) the total number of states in the energy. If an actor controls or regulates the environment's influence on them in order to maintain his internal steady state, then the total number of states that the actor can be in must be greater than the total number of ways the environment affects them. This is called the necessary variability Theorem (Law of Requisite Variety). It is worth noting that this theorem refers to the ability of the agent to maintain the internal steady-state within the ideal range by coping with the impact of the environment. Therefore, the number of necessary states of the agent must be equal to or greater than the total number of ways in which the environment may disturb or affect it, rather than the total number of states that the environment can reach.
William Grey Walter and his famous automated "tortoise" robot, this series of robots can respond to external light, meet obstacles to adjust the direction, follow the navigation lights to return to the charging place, They can also interact with each other and even react to themselves in the mirror. —Cmu.edu
A good regulator can build an excellent model for his environment. Closely related to the necessary variability theorem is the excellent regulator theorem. This is a mathematical concept: if the actor succeeds in mitigating the impact of the state of the environment, in some senses, the actor must have established a model of the environment. This means that the actor must know what kind of internal state to use to deal with what kind of environmental state.
Early cybernetics were mainly concerned with the performance of simple actors. A famous example is a series of automated "tortoise" robots. These robots can simply respond to changes in the environment in real-time, thereby participating in complex interactions with the environment. Cyberneticians soon realized that the same laws also apply to information processing in the brain: in order to successfully respond correctly to changes in the environment, actors need to model and predict relevant aspects of the environment. This means that if we want to understand the design of the brain, we can carefully observe the external environment facing the brain.
This view has become more and more influential in the fields of psychophysics, primate electrophysiology, cognitive computing neuroscience, and clinical neuroscience. In these areas, this perspective provides a whole new perspective on disturbances in perception, perception, and behavior. However, the definition of "prediction" itself in these areas is often unclear, so the term is also used inaccurately. In almost every place, the prediction is considered to be in the level of information processing, and high-level processing acts on low-level processing in a "top-down" manner. In this article, we will provide evidence to challenge this view, which only considers "top-down", and shows that the nervous system also contains many other forms of prediction information, presented in a "bottom-up" processing method.
In this opinion piece, we think, to distinguish two laws in the world (regularity) or "mode" (pattern) is very important, these two laws form the basis for the prediction of the nervous system. We will introduce when empty global (or constancy) does not distinguish the law level (spatio-temporally global or constant, non-hierarchical regularities) the local temporal regularity (spatio-temporarlly local regularities) between the fundamental difference, which is dependent on The situation is therefore hierarchical.
We will provide evidence that the predictions of prior information based on these two laws are mechanically different: they are related to two different forms of information processing (bottom-up and top-down). Finally, we illustrate the importance of distinguishing between these two forms of prediction. Although we believe that as research in related fields matures, these two forms of forecasting will be further refined, we consider this main distinction-based on form rather than the content of the forecast-as health and disease Field, an important first step towards a more comprehensive and diverse predictive processing mechanism.
Prediction is information theory (information theory) and Bayesian decision theory (Bayesian decision theory) important section. Bayesian models to help us understand the brain function once played an important role, perception and cognition inference (perceptual and cognitive inference) depends not only on the current input but also by the so-called prior knowledge about the environment or structure Background information shapes.
It is worth emphasizing that Bayesian decision theory provides a normative framework: it allows researchers to specify how actors should consider the current input and prior information to maximize specific utility when considering the information available. (Utility) , but the exact mechanism of how this process is implemented is still unknown. Just as a map can detail the best route but not provide the best mode of transportation, the Bayesian decision framework is about the overall objective prediction, not how the prediction is formed in the nervous system.
Predictive coding (predictive coding) concept is largely shaped how we consider neural information processing, and in normal and disease groups, this concept has inspired some of the most detailed neuroscience prediction mechanism. People have proposed a series of predictive coding models. These models are computationally similar, but the assumptions at the neuron level are very different. However, these most predictive coding models allow predictions to be further conceptualized as top-down processing.
In the field of cybernetics, which foreshadows most of the thinking in today's predictive processing frameworks, early research provides a broader understanding of the nature of prediction. It is worth noting that the “predictive processing” we use here is a general term that includes predictive coding; the latter is a specific form of predictive processing.
The term cybernetics comes from the Greek word “ Steerman ”, which means that a successful actor must somehow control the effects of his surroundings: this control is not necessarily by limiting the environment, nor This can be done by adaptively responding to changes in environment-related parameters. The helmsman does not have to control the wind, waves, and ocean currents, but the helmsman can make adjustments to minimize the impact of the environment on the ship's route.
Ashby made the necessary variability Theorem (Law of Requisite Variety) and excellent regulator Theorem (Theorem Good Regulator) , these two complementary theorems closely related to our ideas. These theorems provide a perspective on the nature of prediction. These two theorems mean that structures that take into account relevant environmental impacts can provide us with important principled insights into the fundamental feature design of actors . With the deepening of this idea, the environment can have an influence on the actors, and the actors can also model the environment. We believe that taking environmental rules into account can help us sort out the necessary predictions.
We can classify the space-time dimensions that affect the actors according to different kinds of environmental rules. It is generally believed that some rules are global in space and time (that is, not limited to a specific space or time node), and are related to each encounter between the actor and the environment. These rules are independent of situational factors, and therefore is non-hierarchical: there is rule is immutable, and independent of the other state the current situation or environment. Therefore, we can assume that modeling these rules requires actors to experience and deal with similar situations. The evidence supporting this hypothesis is constantly increasing, and this evidence will be mentioned below.
Starting from this evidence, we infer that the a priori knowledge that allows people to predict global rules of space-time is the basis of the bottom-up processing structure. The impact of these a priori information on us is predictive; in a sense, this impact is an overall prediction of the surrounding environment of the agent (not related to the current specific situation) , and does not depend on the current sensory input. We assume that this prediction occurs automatically and unavoidably every time the actor encounters a stimulus, deciding and shaping our long-standing interaction with the world.
Although the concept of global prediction limits is not entirely new, it is ignored by existing prediction processing interpretations. Existing explanations focus more on another prediction related to the local rules of time and space . These rules will only appear in certain circumstances and affect the actors. Due to their situational dependence, these rules have a hierarchical structure , in which the state of the situation determines whether the rules appear or are relevant .
In order to model these rules, the actor's brain must reflect the characteristics of the rules. This can be achieved through a hierarchical, top-down processing system: the high-level information processing mechanism extracts the current situation information, and then returns the prediction results to the lower-level unit to modulate the previous processing.
The correspondence between the two rules to the prediction has the equivalent helmsman analogy in cybernetics : the material and shape of the ship are unchanged—because the core characteristics of the helmsman's dependence on the medium of navigation are unchanged. In contrast, the use of ship structure features must have undergone context-related changes to respond to the challenges posed by changes in environmental characteristics (wind, waves, and ocean currents) . In simple terms, the actor has constant characteristics, which are used to regulate the constant influence of the surrounding environment, and also have the characteristics of situation dependence , which reflect and regulate the situation dependence characteristics of the surrounding world. In the next section, we will give an overview of the evidence supporting this view.
Now on, we will use the term " restriction " (constraint A) to refer to the basic rules and independent situational context thus formed independent predictors. This term comes from the field of computer vision. In this field, restriction refers to a similar point of view. It intuitively points out that the structure of the nervous system forces information processing to follow a predetermined path. The trajectories of these roads are estimates of the environment based on a priori information by the actors, and therefore are predictive.
It is worth noting that we do not support the expansion of the concept of the term "forecast". On the contrary, we believe that the consistent use of the computing concept of this term can lead us to think about how the limitations of bottom-up processing are predictive. In contrast with the limitations, we use the term " desirable " (Expectation) associated with the a priori information to refer to the context-dependent rules. The effect of the top-down processing of attention is different from the prediction processing in function. The prediction processing is the focus of this article, so we will not discuss the details of the top-down processing of attention.
Laws and predictions
Context independent regularity and constraints
Context independent regularity and constraints
The natural world may seem ever-changing, however, from highland meadows to the Mediterranean coast, the natural landscape shows inextricably linked in the image: different images in such as local line direction, contour shape and object placement The distribution is very similar. The sensory system uses these laws to optimize the encoded information when interpreting the image , so as to achieve the best performance while minimizing energy consumption . Today, a growing number of studies have looked at from a neurological point of view, to explore the sensory information (sensory evidence) and based on a priori knowledge of the predicted (knowledge-based predictions) is how to integrate the.
The results of these studies indicate that the laws of globality and situation independence that people have summarized based on prior information are implicitly embedded in the structure of the information processing mechanism. For example, the direction information in the landscape image is not evenly distributed - vertical and horizontal direction in the picture there is often overexpressed (over-Represented) . As the sensory system on the spindle (cardinal axes) direction senses variations exist, as well as neuronal higher degree of sensitivity of the direction, people perceive Bureau toward (local orientation) when there are rules based on a distribution tendency .
Importantly, based on some of the electrical physiology and functional magnetic resonance imaging study results, people think that this tendency / restriction (constraint) is implicit present in the primary visual perception cortex (V1) structure FM in (tuned to) the main to (Cardinal) direction neurons are overexpressed in the V1 region, and the frequency modulation function is narrower than other neurons . In this structure, based on one can perform implicit knowledge based on the direction (a priori information) Bayesian inference (the Bayesian Inference) , without a priori information in a hierarchical structure from top to bottom (top-down hierarchy) of Way to make explicit expression. A study on the basic property law of another environment, the moving speed of an object, shows that (a priori) restrictions are also implicitly characterized on structures with uneven distribution of neuron density and narrow frequency modulation of related neuron groups .
People also group local representations in the environment through overall, situation-independent laws . There are also regularities in the outlines that define the visual boundaries of objects. When extracting the outline information in the picture, people's visual system will use the typical outline shape in the prior information to group the single position information to make it a larger processing unit. This grouping mechanism called the " Association Field (Field, Associate) , an information packet and distribution profile nature of the association field direction is extremely relevant. The contour information is very important for defining the features and object information in the picture.
Animal studies have shown, for early retinal membrane skin (early retinotopic cortices) between neurons connected in parallel (horizontal connection) structure [MOU3] can not be played in the establishment of the Association Field of the important role indispensable. Recently, an animal study explored the " silent " background in orientation tuned V1 neurons. Silent background means that when the stimulus is located in a specific visual area, the stimulus itself is not enough to activate the activated neuron, but it can regulate the activity of the neuron. This study shows that V1 neuron-level connection structure with human psychophysical in their spatial arrangement junction field (psychophysical association field) very similar.
The discovery of the parallel connection arrangement structure in the V1 region provides neurophysiological evidence for people to use a priori information related to the environmental structure to predict the contour recognition. The embedded limitations of these structures ensure that the local orientation information in the outline is consistent with our prior knowledge of the world structure. The context-independent mechanism of contour integration may be affected by the top-down hierarchy.
Even in complete objects there are basic laws of situational independence . One example is an object with the outline of objects together is generally curved (Convex) . Although the mechanism consists of a series of top-down process, including all involved in the visual system to separate objects from the background task, there are still studies have shown that situation to an independent predictor of curvature (convexity) restrictions embedded in a bottom-up process The method within the program plays an important role in this task. In the V1 region, the feedforward neural connection and the feedback neural connection are terminated in different layers . This is for our top-down adjustment and bottom-up adjustment including horizontal influences . Separation provides an opportunity . A study of rhesus monkeys using laminar recordings showed that the parallel connections in V1 played an important role in early image-background separation.
Importantly, the calculation model shows that the specificity of the parallel connection between the promotion and suppression of neural loops in the V1 region guarantees the implementation of the prediction, and the curvature limitation in this prediction comes from the situational independence law of the object structure.
Another example that proves the law of background independence comes from the fact that some object categories often appear in positions that conform to the rule of experience . Since our physiological senses are used to recognize the environment, the regularity of the coordinates of objects in the real world is therefore converted into regularities in retinal imaging . For example, the coordinate positions of lawns and carpets can be converted into our lower half of the field of view, while faces and text can be converted into our upper half of the field of view.
A recent study on population receptive fields in neurons showed that the prediction of the location law was generated in high-level visual cortical neurons related to receptive fields related to specific object categories, that is, receptive fields to specific object categories There is a bias in positions that often appear in the field of vision. For example, among native English speakers, the group of neurons associated with text recognition appears in the receptive field to be smaller in scope, biased toward visual center, and more horizontally (rather than vertically) . This finding has also been verified in face and scene recognition.
All in all, the above research shows that the regularity information of the external environment is implicitly encoded in the information processing system and is a stable structural component of the system . The prediction thus affects / limits the bottom-up information processing process. In this paragraph, we mainly focus on the process of perception, but there are more studies from learning or other cognitive fields that also provide evidence for the independent limitation of context . For example, some of the experiments show that different organisms (organism) level of environmental rules for acquisition are different. Different organisms seem to be restricted to varying degrees in their readiness to learn, and the intensity of the restriction may be determined by the degree of relevance of environmental rules to the animal.
A recent study on prepared learning for fruit flies provides evidence for this finding. The research team applied different experimental treatments to the two groups of fruit flies: one group applied annoying stimuli after applying visual signals, and the other group applied annoying stimuli after applying olfactory signals. After 40 generations of reproduction, the two groups of fruit flies have different preparation states for different stimuli (visual / sniff) . Establishing predictions within the limits of the bottom-up information processing process can enable a particular animal to maximize the use of environmental information, and reduce metabolic energy consumption while ensuring its performance. Therefore, it is not difficult to understand why people embed context-independent rules into artificial intelligence networks when developing artificial intelligence algorithms.
And where does this restriction embedded in the structure come from? Either from theory or is it the methodology, we are difficult to phylogeny (phylogeny) and individual development (ontogeny) split off from the impact on animals. This article will not discuss this issue further, but the existing evidence shows that there is no one theory that can explain all relevant phenomena. Phylogeny and individual development may interact to influence this limitation in the nervous system . In some extreme cases, it has been found that this limitation may be related to genetic coding in a way that is independent of perceived experience. In other examples, the limitation may also be related to adult experience.
Situation-related laws and expectations
Other environmental rules to contextual relevance (context-dependent) of the form. Imagine a scenario where we are more likely to observe woodland birds than wading birds while walking through the forest. Recognition of bird species thus activates the representations in the human brain that are used to predict specific features-such as the representation of a certain form of a beak. The appearance of the beak indicates certain lower-level rules, such as certain contours and certain visual field positions toward the edge.
In this example, the dependence on context depends on whether the low-level local feature law of the input stimulus is determined by its higher-level feature or by the information that is not related to the stimulus. Therefore, in addition to the situational independence feature, there is also a hierarchical nesting feature in the environment structure : the high-level regularity in the environment determines the low-level regularity. We predict that there must be similar hierarchical nested information processing mechanisms in the brain that deal with context-related environmental laws. We believe that there is a top-down structure in the processing mechanism, which enables advanced programs to extract contextual information , make inference predictions, and feed back the predictions to the model, thereby affecting the early perception process .
There is a lot of evidence in the cerebral cortex that can support top-down processing . In the ornithological example given above, environmental context information determines which birds we are more likely to encounter. Studies have shown that the brain can sense objects with the aid of scene-object dependency . For example, people recognize objects placed in regular situations faster. It is believed that the prediction of high-level context-specific cortices from the situation is fed back to the characterization of low-level objects, resulting in this promotion [73].
When the rough edges of the object are extracted and separated from the background, the human brain derives regular predictions, which produces a dynamic interaction between local feature processing and prediction. Many electrophysiological studies on primates show that the top-down structure can greatly filter all levels of information processing, and this filtering even affects the early information processing stage of subcortical structures.
In the above example, the prediction information mainly comes from sensory input. Recently, however, more and more studies have pointed out that the top-down structure does not only use sensory input information. For example, ornithologists can quickly discover and identify the type of bird. Mental and physical functional brain imaging studies such evidence is given that the professional knowledge of certain objects (knowledge) can be in the form of a top-down play a significant role in promoting the object recognition. In addition, psychophysical research shows that the expectation of specific object components and even semantic meaning can affect the recognition of early visual features in a top-down manner. It is expected to be generated before sensory input and stored in the advanced memory system. Brain functional imaging studies have also given similar evidence—a priori knowledge of objects stored in distributed networks such as parietal/frontal regions has dynamically affected visual imaging in the early retinal cortex. Interestingly, psychophysical and brain functional imaging studies have found that only the expectation of stimulus components can activate the characteristic templates in the early visual cortex.
Social interaction and social human animal (social interaction) there is a strong correlation situation. Some studies have shown that expectations based on prior social knowledge have an extremely important influence on information processing . Psychophysics research shows that people can use the context-related laws in social interaction knowledge to understand the information in other people ’s action patterns. This guiding relationship is also top-down. Even the anticipation of the psychological state of others, such as the intention of others to produce a certain action, may affect the early perception of social information processing .
Continuous exposure to the same or like that may affect people's perception processing the sensory stimulation, a phenomenon commonly referred to adaptation (Adaption) . The phenomenon of adaptation is complex and complicated, and people do not fully understand the mechanism behind it. Existing models are mostly opposed to it as passive "nervous fatigue" (Neural Fatigue) and more likely to be understood as an active process , but we still do not know whether the phenomenon has been adapted to predict the impact of processing. It is believed that phenomena such as decreasing the intensity of neural response to repeated or predictable stimuli and increasing the intensity of response to unpredictable stimuli are related to the independent prediction of top-down processing scenarios. But another part of the research shows that bottom-up processing also affects adaptation . Researches that consider the phenomenon of adaptation to bottom-up processing generally use similar experimental treatments, while experiments that consider the phenomenon of adaptation to top-down processing use various experimental treatments. These findings may reveal two mechanisms: one is bottom-up processing affected by context-independent constraints, and the other is top-down processing affected by context-related predictions. The latter can function in a changing environment.
All in all, many studies have pointed out that environmental predictability may be contextually relevant . In this scenario, the bottom-up processing of the prediction information that we discussed in the previous section is less helpful. Individuals that produce the best performance can usually use predictions to flexibly adjust information processing in a top-down manner, and most of the research on prediction processing models today usually focus on this form of prediction.
Meaning and application
Some people believe that there is no need to distinguish between different forms of prediction because they ultimately serve the same purpose : to promote the inference of the state of the world, thereby optimizing the interaction between the organism and the world. This kind of thinking makes sense. When we try to make a high-level and functional description of behavior, it is like many models related to optimization. These models provide a valuable benchmark for evaluating the performance of an actor from a functional perspective.
However, prediction processing frameworks often make an additional premise: that the prediction process is mediated by top-down mechanisms. As we said above, this view is incomplete: predictive information can be implemented in the brain in at least two forms. If our goal is to seek an understanding of neural mechanisms, then it is necessary to revise the current single point of view. In many ways, we have seen the value of enhancing and expanding a single predictive processing mechanism.
If a computational model only traces the top-down prediction model, then it is likely to face the risk that the computational model is disconnected from the neural mechanism before, which will, in turn, hinder the feedback loop between the model and empirical research, so as to hinder the subsequent Research progress. Examining the prediction process involving illusions clarifies this problem. The "illusion" is deliberately created, and it runs counter to the normal operation of the sensory system. In the current forecasting process, the illusion is conceptualized as the result of top-down processing. For example, the classic Cornsweet effect -two light spots are separated by a gradient in the center, causing an illusion of different brightness-has been regarded as an example of a top-down effect. It is related to beliefs about the spatial gradient of brightness and reflection and has been simulated using a top-down prediction network.
However, although it may be regulated by a high level of nerves, most of the impact is derived from the early subcortical and even retina prediction information: the Constellation effect is theoretically related to the receptive field structure of retinal ganglion cells, and to the lateral knee The signal in the lateral geniculate nucleus is strongly correlated (projection of retinal ganglion cells) and has been proven to originate from monocular neurons, indicating that it originated under the cortex. To put it more succinctly, one of the main challenges of explaining illusions in a top-down framework is that the neural circuits that cause illusions can be independent of any previous experience. This has been shown by clinical evidence: people with congenital or early-onset blindness will immediately develop certain optical illusions, such as the Müller-Lyer illusion, immediately after vision restoration surgery.
Cornsweet effect
These illusions may arise from the prediction process. However, we believe these phenomena in large part (but not necessarily all) can be bottom-up process, the scenario has nothing to do with the constraints to better explain . For example, a lot of illusions are related to the light from above. This, we will discuss in detail in the next section.
In this example, the disconnect between empirical evidence and computational models suggests that an explanation may have descriptive validity at the computational level, but not at the neural mechanism level. Interestingly, studies have shown that even Rao (Rao) and Ballard (Ballard) ground-breaking predictive coding framework, but also can be rephrased in such a way: Predicting perhaps through lateral inhibition (lateral inhibition) instead of connecting feedback (feedback connections) to achieve.
Obviously, we need to more tightly integrate neural mechanisms and computational models, and distinguish between different forms of prediction at different scales. For example, a recent study temporarily supported the idea that scenario-related predictions may share a common source, while scenario-independent predictions may rely on built-in constraints that share another source.
In the process of hierarchical prediction processing, predictions are generated at several different levels of the information processing level. The predictions at different levels are considered to interact through a top-down mechanism to ensure that ultimately all predictions are consistent with each other. However, bottom-up constraints and top-down expectations may interact in a completely different way . Since constraints will not change to a large extent due to expected short-term changes, constraints and expectations may affect the same process, but will not directly affect each other and align with each other.
"A priori of light from top to bottom" provides a good example . The direction of the light shining on the visual scene and the resulting shadows provide information used by the human visual system to infer the shape of the object. In the absence of clear information about the location of the light source, human observers judging the properties of objects indicate that the visual system will conservatively predict that light comes from above. Interestingly, this prediction can be corrected through experience; after training, the feedback received by the observer indicates that the position of the light source has changed, and the prediction of the visual system will move to a new position. The traditional prediction processing method for this phenomenon shows that the top-down prediction information has been updated.
However, this is inconsistent with the research results of the new prior and laboratory environment-the prior is caused by other acquired clues. Another explanation is consistent with this scenario-specific and electrophysiological evidence that the a priori from the above light is implemented as a constraint of bottom-up processing and is not affected by the short-term experience. Experimental training did not change the original constraints but produced a new, environment-related expectation that the light has changed.
We make an interesting assumption about the interaction between different forms of prediction: when information from constraints and expectations interact, the former may never be completely covered by the latter. This is in contrast to the pure top-down forecasting process, whose ultimate goal is to ensure that all forecasts are consistent with each other.
In addition, our explanation indicates that the top-down processing of experimental operations should have a different effect on the newly obtained predictions, but may keep the original priors unchanged. This differential effect should be further observed at the neural level. For example, we can predict that neuroimaging experiments will show the effect of short-term learning on light source position shifts in advanced processing areas, while the neural characteristics of bottom-up processes will remain unchanged.
This form of interaction between two different forms of prediction helps to achieve the best balance between durability and flexibility. Interestingly, this interaction does not seem to be ubiquitous in all living organisms: even after a long period of experience, the way chickens perceive the shape of objects also indicates that their visual system assumes that light comes from above. From birth, chickens seem to have an immutable bottom-up constraint to predict light from above, and they cannot obtain an environment-related expectation to adjust these built-in predictions, which strongly proves that prediction information can Completely be decoupled from personal experience of the world.
In the clinical practice of mental illness, heterogeneous symptom groups are often misdiagnosed as symptoms of the same category, because different diagnoses may be interfered by overlapping neurophysiology. Therefore, research often evades standard diagnostic categories, focuses on single symptoms, and seeks narrower but deeper mechanisms. The predictive processing framework has always been an important part of this business, but its focus on a single form of prediction often limits the framework ’s interpretation of the diversity of symptoms. Take, for example, neurological diseases and schizophrenia mediated by autoantibodies. More and more attention is drawn to people with mental illnesses, and IgA antibodies against NMDA receptors have been found in their serum or cerebrospinal fluid. These antibodies may be, but not necessarily, the underlying cause of mental illness. Since antibody-mediated immunotherapy for mental disorders is only performed when clinically needed, clinicians face an in-depth explanation at the neuromechanical level rather than at the computational level. In other words, from a clinical point of view, predictive processing models require an explanation at the neural mechanism level.
Link the actor ’s neural toolbox with environmental statistics
We have proposed that features in the environment that support prediction are effectively classified as situation-independent or situation-related regularities, and these features are mapped to constraints and expectations, respectively. Recognizing these fundamentally different forms of forecasting may stimulate a broader framework for forecasting. In particular, the goal of linking the information processing of actors with environmental statistics has always been a guiding principle in the statistical work of natural scenes, and has provided a similar principle framework for the study of prediction processing .
For example, an analysis of environmental laws may enable us to predict and explain to what extent this is regulated by top-down processes or bottom-up processing constraints, or why the prediction process has short-term changes related to the scenario The sensitivity varies. More generally, the analysis of the relevant laws in the environment of organisms may help to develop a unified framework, which can explain why the predictions of organisms in different fields may appear in different forms. This perspective provides a way to link the predictive processing framework to sensory ecology and evolutionary theory.
In conclusion, The principle behind our argument is simple: by observing the structure of the environment, one can understand the key design features of the information processing mechanism of the actor . This basic point of view is not new. It not only promoted the thinking of early cognitive biologists, but also promoted cybernetic experts and current work on natural scene statistics. Here, we believe that applying this principle to predict the role of brain function will provide us with a new and useful perspective . In order to achieve a deeper understanding of the brain as a "prediction machine", we believe that it is necessary to recognize that there are different forms of prediction by the nervous system . The immutable regularity in the environment is reflected in the corresponding scenario-independent forecasting mechanisms. These mechanisms are used for bottom-up processing. We refer to these mechanisms as constraints. Similarly, the fluctuations of the external world that depend on scenarios are directed to a flexible scenario-dependent forecasting mechanism implemented by a top-down process, which we call expectations. With the development of prediction processing, it is expected that it can be applied to more and more specific problems based on neural mechanisms.
Other differences may be useful, but we believe that the concepts presented here are primary and basic. Although we mainly discuss these ideas from the perspective of neural information processing, a comprehensive understanding of the role of prediction in information processing must ultimately include an understanding of the interaction between the entire organism and its environment. Predictive information exists not only in neural mechanisms, but also in the morphology of entire organisms, which has been proven in many examples of sensory ecology and behavioral ecology. Finally, the goal of the prediction processing framework should be to combine predictions as part of biological information processing in many different ways to provide a more comprehensive view of how we interact with our health and disease world.
Reference
Compilation source: Teufel, C., Fletcher, PC Forms of prediction in the nervous system. Nat Rev Neurosci 21, 231–242 (2020). Https://doi.org/10.1038/s41583-020-0275-5
Comments
Post a Comment
Please let me if you have any question