"The Experience Machine" by Andy Clark

Along with Karl Friston, Clark has been on the forefront of the effort to understand the human mind via the concept of predictive processing. This is a pop-neuroscience book intended to detail this concept to the layman. He begins with levels of processing, “However complex or high-level the predictions, it is prediction errors that must then carry the news, signaling differences from the expected and thereby keeping us in touch with a changing and sometimes surprising world…. Human brains seem to benefit from intelligent prediction strategies of just that kind, and they do so in an especially powerful way, thanks to the use of multiple “levels of processing.” In these multilevel contexts, simple predictions are nested under less simple, more abstract ones…. In this kind of multilevel arrangement all that flows forward (from the sensory edges ever deeper into the brain) is news—deviations from what is expected. This is efficient. Valuable bandwidth is not used sending well-predicted stuff upward…. Systems like that are wonderfully frugal in their use of the incoming stream of information. Instead of trying to deal with everything from scratch they effectively sift and filter the incoming data by highlighting only what was unexpected. This is the nugget of truth in the notion that human brains hallucinate reality. It means that the world we experience is to some degree the world we predict. Perception itself, far from being a simple window onto the world, is permeated from the get-go by our own predictions and expectations. It is permeated not simply in the sense that our own ideas and biases impact how we later judge things to be, but in some deeper, more primal, sense…. What we might think of as simple or “raw” sensory evidence is itself never experienced. Instead, experience always and everywhere reflects those rich webs of prior knowledge and here-and-now expectation…. Since all human experience is constructed from mixtures of expectation, attention, and sensory stimulation, it will never be possible to experience either the world or your own body “as it really is.” Indeed, it rapidly becomes unclear what this could even mean.”

Action is just as important as perception in the framework of predictive processing. “Ordinary daily actions (according to predictive processing) are caused by predictions of bodily sensation. They are caused, more precisely, by predictions of the flow of bodily sensations that would occur if that very action were to be performed…. Successful action involves a kind of self-fulfilling prophecy…. By making prediction the common root of both perception and action, predictive processing (active inference) reveals a hidden unity in the workings of the mind. Action and perception form a single whole, jointly orchestrated by the drive to eliminate errors in prediction…. Actions come about because we mentally represent the completed effects of the action…. This is sometimes said to reverse a commonsense notion of causality, since instead of the action causing the effect, it is the representation of the effect (the completed action) that causes the action itself to unfold…. This became known as the “ideomotor theory of action,” since the idea (or mental image) of the completed motor action is what brings the actual movements about…. By launching a cascade of sensory predictions, and then rendering them true by means of action (thus eliminating the resulting prediction errors), the brain creates the desired movements…. The deep unity (under predictive processing) of perception and action should now be apparent. There are two different, but equally effective, ways to minimize prediction errors during our encounters with the world. The first is by using prediction errors to help us discover the best guess about how things are out there in the world. But the second is to act so as to make the world fit some of our predictions.”

Clark references Lisa Feldman Barrett’s work on the role of perception and interpreting the signals of the interoceptive system in creating emotions. “The central idea is that a single kind of process combines inner and outer sources of information, generating a context-reflecting amalgam that is experienced as emotion…. The very same bodily information can thus feel very different according to how we represent the larger context in which the bodily signals arise…. According to interoceptive predictive processing, feelings and emotions are what result when we integrate basic information about bodily states and general arousal with higher-level predictions of their most probable causes…. Emotions and feelings reflect a process that combines interoceptive (inward-looking), proprioceptive (action-guiding), and exteroceptive (outward-looking) information with model-based predictions of all those signals as they are occurring…. The winning predictions will be the ones that best “make sense” of that large and varied body of information.”

Precision-weighting is integral to the system of predictive processing. “The most successful predictive model always sculpts the way the brain deals with the incoming signals. It alters the response of neurons at multiple cortical levels, amplifying and dampening them (courtesy of all that variable precision-weighting) in ways that reflect the brain’s best guess at the structure of the objects out there in the world…. Precision-weighted prediction (usually) serves our purposes by highlighting some things at the expense of others.”

Our personal histories shape our predictive models. Clark argues that perception is neither passive reception nor free fantasy, but a controlled hallucination continuously corrected by prediction error. “The most basic way that we actively construct our world is by selective sampling. We move our body and aim our gaze in ways that reflect what we expect to encounter…. Humans with different individual histories, will harvest different sets of stimulations from the very same world. But as we selectively harvest those simulations, our brains impose structure a second time, processing the sensory information in ways that amplify and dampen, extracting meaningful structure that itself reflects our own prior experience…. We cannot help but base our current waves of prediction on our own native tendencies and particular life histories. Where those predictions vary, so does human experience.”

Clark returns to the role of actions in shaping and creating our world models. Predictive processing thus becomes, for Clark, not merely a neuroscience theory but a general framework dissolving sharp distinctions between perception and action, mind and world, organism and environment. “Practical actions and epistemic actions are determined in exactly the same way, as the predictive brain makes counterfactual predictions about what kinds of futures will result if certain actions are launched. Actions are then chosen that deliver preferred outcomes directly (when possible) or else that probe and sample the environment to bring forth more information, reducing key uncertainties, and making the desired outcome more likely in the future…. Any predictive processing agent able to minimize error relative to future goals will discover both epistemic and practical actions, and how to mix them together. In all such scenarios, all the brain does is select the actions that best minimize future prediction error relative to goals.”

David Chalmers and Clark came up with the idea of the extended mind to describe the scaffolding outside the body that becomes enmeshed with the human mind, creating one seamless entity, “Extended minds arise because predictive brains are naturally expert at exploiting opportunities to use information-gathering action loops to help them achieve their goals…. Our constructed worlds can sometimes take over, transform, and augment functions once carried out by our brains…. I continue to believe that as the resulting weave between brain, body, and external resources tightens, it becomes less and less productive to think of mind as something locked neatly behind the barriers of skin and skull…. Brains are prediction machines that invoke external resources as easily (and for the same reasons) as they engage practical actions and activate different aspects of they dinner circuitry…. This creates a circular causal web in which mind is—at the very least—constantly porous to body and world.”

Finally, Clark returns to the importance of weighting in building world models, “Variable precision-weighting is the single most powerful tool in the predictive processing toolkit…. In the brain, precision-weighting alters patterns of post-synaptic influence (the strength of the signals passed on after the synapse “fires”). This means that specific signals can be selected for enhanced impact. The signals selected for this special treatment will be ones that are expected to be both reliable and important for the task at hand…. Precision variations of this kind underpin both conscious and automatic deployments of attention…. The brain learns how and when to vary its precision estimations as part and parcel of the process by which it acquires the generative models themselves. Those models and associated estimations of precision are learned by repeated exposure to flows of sensory information in the context of trying to act in the world.”

“Surfing Uncertainty” by Andy Clark

Clark is a philosopher who specializes in logic and metaphysics. However, this book combines neuroscience and embodied cognition to give a theory of the predictive brain. Clark has coined this process predictive processing and it relies heavily on Bayesian logic. “It is the kind of automatically deployed, deeply probabilistic, non-conscious guessing that occurs as part of the complex neural processing routines that underpin and unify perception and action…. Brains like ours…. are predictive engines, constantly trying to guess at the structure of the incoming sensory array. Such brains are incessantly pro-active, restlessly seeking to generate the sensory data for themselves using the incoming signal (in a surprising inversion of much traditional wisdom) mostly as a means of checking and correcting their best top-down guessing. Crucially, however, the shape and flow of all that inner guessing is flexibly modulated by changing estimations of the relative uncertainty of (hence our confidence in) different aspects of the incoming signal. The upshot is a dynamic, self-organizing system in which the inner (and outer) flow of information is constantly reconfigured according to the demands of the task and the changing details of the internal (interoceptively sensed) and external context.” Our brains are constantly in action, making top-down guesses about the sensory data to learn about our external world. The brain is creating models and then slowly modulating them, based on past experience, to improve future predictions. Crucially, the brain is not a passive system. Action plays a critical role in predictive processing. “Our massed recurrent neuronal ensembles are not just buzzing away constantly trying to predict the sensory stream. They are constantly bringing about the sensory stream by causing bodily movements that selectively harvest new sensory stimulations. Perception and action are thus locked in a kind of endless circular embrace…. [The brain is able to] use action upon the world to reduce the complexity of its own inner processing, selecting frugal, efficient routines that trade movement and environmental structure against costly computation.”

The brain is constantly learning as it is constantly predicting everything that will happen to the body in the next moment, in an ever-rolling cascade. Perception is very near-term top-down prediction modified by the senses. “Prediction error [is] a kind of proxy for any as-yet-unexplained sensory information. Prediction error here reports the ‘surprise’ induced by mismatch between the sensory signals encountered and those predicted…. Perception is indeed a process in which we (or rather, various parts of our brains) try to guess what is out there, using the incoming signal more as a means of tuning and nuancing the guessing rather than as a rich (and bandwidth-costly) encoding of the state of the world.” It is actually our expectations, to a large extent, that determine what we see, smell, and hear. A prediction does not create our sensory world, but it does focus our attention. “Brains like ours are constantly trying to use what they already know so as to predict the current sensory signal, using the incoming signal to select and constrain those predictions, and sometimes using prior knowledge to ‘trump’ certain aspects of the incoming sensory signal itself. Such trumping makes good adaptive sense, as the capacity to use what you know to outweigh some of what the incoming signal seems to be saying can be hugely beneficial when the sensory data is noisy, ambiguous, or incomplete.” Again, the non-passive nature of the brain is crucial. “Action is not so much a ‘response to an input’ as a neat and efficient way of selecting the next input, driving a rolling cycle. These hyperactive systems are constantly predicting their own upcoming states and actively moving about as to bring some of them into being.”

Each percept is constructed with the help of the brain’s priors. These priors influence future expectations in a probabilistic manner. The brain then combines the likelihood of these priors with raw sensory data. “Attention, thus construed, is a means of variably balancing the potent interactions between top-down and bottom-up influences by factoring in their so-called ‘precision’, where this is a measure of their estimated certainty or reliability.” When evaluating sensations, the brain is constantly separating the signal from the noise, using prior knowledge. This process happens through hierarchical Bayesian inference based on precision-weighted guesses at every level. These predictions then influence future action in a proactive fashion. Memory is also crucial in predicting our future. Fernyhough suggests, “if memory is fallible and prone to reconstructive errors, that may be because it is oriented towards the future at least as much as towards the past…. similar neural systems are involved in both autobiographical memory and future thinking, and both rely on a form of imagination.” In the end, perception is the brain’s best guess as to reality. “Perception (rich, world-revealing perception) occurs when the probabilistic residue of past experience meets the incoming sensory signal with matching prediction.”

Prediction allows our bodies to live in the present. As Franklin and Wolpert assert, “delays are present in all stages of sensorimotor system, from the delay in receiving afferent sensory information, to the delay in our muscles responding to efferent motor commands…. we effectively live in the past, with the control systems having access to out-of-date information about the world and our own bodies.” This is overcome by predictive processing. “Forward models provide a powerful and elegant solution to such problems, enabling us to live in the present and to control our bodies…. you treat the desired (goal) state as observed and perform Bayesian inference to find the actions that get you there…. Motor control is, in a certain sense, subjunctive. It involves predicting the non-actual proprioceptive trajectories that would ensue were we performing some desired action. Reducing prediction errors calculated against these non-actual states then serves…. to make them actual. We predict the proprioceptive consequences of our own action and this brings the action about….. ‘Active Inference’ then names the combined mechanism by which perceptual and motor systems conspire to reduce prediction error using the twin strategies of altering predictions to fit the world, and altering the world to fit the predictions.”

Frith makes the case that “our perceptions are fantasies that coincide with reality.” Hohwy suggests, “what we perceive is the brain’s best hypothesis, as embodied in a high-level generative model, about the causes in the outer world…. Conscious experience is like a fantasy or virtual reality constructed to keep the sensory input at bay.” Thus, the human brain is bounded by an external reality. “Prediction-driven learning delivers a grip upon affordances: the possibilities for action and intervention that the environment makes available to a given agent.” The brain is then constantly making its best guess as to what course will reduce prediction error. This might be by modifying its predictions or it might be through acting. This results in “‘affordance competition’ in which…. possible motor responses are being simultaneously prepared, and in which ‘the human brain does not wait for a decision to be completed before recruiting the motor system but instead passes partial information to prepare in a graded fashion for a probable action outcome’…. Such pro-active readiness, to be genuinely useful, must necessarily be multiple and graded. It must allow many possible responses to be simultaneously partially prepared, to degrees dependent upon the current balance of evidence.”

The world that humans perceive is “our world” in that it is the world best understood by humans. “What we perceive is (when all is going well) the structured external world itself. But this is not the world ‘as it is’, where that implies the problematic notion of a world represented independent of human concerns and human action repertoires. Rather, it is a world parsed according to our organism-specific needs and action repertoire.” Humans create and actively modify our own world. Through language “our own thoughts and ideas now become available, to ourselves and others, as potential objects for deliberate processes of attention…. Courtesy of all that material public vehicling in spoken words, written texts, diagrams, and pictures, our best predictive models of the world (unlike those of other creatures) have thus become stable, reinspectable objects apt for public critique and systemic, multi-agent, multi-generational test and refinement. Our best models of the world are thus able to serve as the basis for cumulative, communally distributed reasoning…. Our human-built worlds are not merely the arenas in which we live, work, and play. They also structure the life-long statistical immersions that build and rebuild the generative models that inform each agent’s repertoire for perception, action, and reason.”