domeet webmaster

It is expected to become the predictive coding of the Great Unification Theory and encounter the “Little Black House Challenge”.

Editor’s note: This article comes from the WeChat public account “neural reality” (ID: neureality) .

Author | Sun Zekun

Cover | COCO

Edit || EON

Typography | Little sunflower

Imagine you are alone in a dark, empty room. It is quiet and stable here, everything is in your awareness and control (follow me: unagi ~). There are no unpredictable sources of disturbance here, such as relatives, bear children, the mother who raided the rounds, and a reminder editor who you thought he forgot but he did not.

Maybe you do n’t think this scene is too magical? In the global social isolation, we have more or less tested how much we like to be paralyzed, housed, or paralyzed. We suddenly realized that moving bricks really makes me happy, and it is better to move to the construction site.

Then the question is coming, why is it so difficult to “stay quietly”? This “difficulty” may be related to the nature of the human mind.

Now, let us abstract the human mind as a series of processes from receiving information to making feedback. If you want, you can also call these processes functions or algorithms. So, is there a more basic theory that can explain and guide all the algorithms involved in the mental process? If we can discover this “great unification theory”, we will undoubtedly be closer to the essence and origin of human intelligence.

This article, we want to talk about the players who are expected to win the title of “Great Unity” in the past two decades-predictive coding, and what does this code have to do with your bedroom without lights on relationship.

A brain that is always full of “expectations”

The first time I heard “pre-The person who “tests the code” probably doesn’t think how close it is to our lives. Double-click to open a JPEG image, and you will be exposed to the predictive code.

You can imagine a seascape photo. Now, we want to encode this picture with as little information as possible so that you can download and transfer it quickly. We found that this megapixel photo is too “blue”, and most of the neighboring pixels are so blue! Only a few places have changed color, such as where the sea meets the sky, or the edge of the beach.

In this case, we do n’t need to repeat encoding one pixel after another, we only need to focus on encoding those pixels that do not maintain formation, and use the deviation between the predicted value and the true value to encode the feature part . For example, if the pixels around a pixel are all blue, then the pixel is likely to be blue; but if it is actually yellow, it will seriously deviate from the prediction. The “prediction error” here is what we want to encode. In this way, we only encode the difference between each pixel and its predicted value, and we can compress the amount of information in this picture.

What does the brain have to do with a JPEG image?

Although the term “predictive coding” was first born in the field of information processing, this idea of ​​”coding error” appeared in cognitive / brain related research as early as the 19th century. After all, our brain is also a system with limited capacity and processing power. For such a system, compressing information, improving efficiency, and how to “fidelity” are crucial. Is it possible for the brain to use “predictive coding” to improve the recognition, storage, and output of information?

The physicist Hermann von Helmholtz (Hermann von Helmholtz) described perception in 1860 as a probabilistic inference process. According to this thinking, when the brain recognizes and understands the external world, it does not just accumulate externally input information, but more importantly, the brain continuously predicts the current information based on the existing knowledge.

Of course, brain coding information is much more complicated than computer coding JPEG pictures. It may be a “multi-layer processing structure”: the relatively low-level processing layer transmits signals to the high-level layer, and the “high-level” is based on existing models ( Knowledge system) for “reverse matching”. That is to say, each level is predicting the activity of the processing layer of the next level. If the predicted signal matches the signal from the lower layer (for example, predicted signal = real signal = “blue”), thenNo further coordination is required in the brain; but! If the two do not match, there will be a “prediction error” that we need to focus on encoding. This error signal will tell the brain’s advanced processing layer: “Hey, your speculation and model need to be adjusted! Next time long Remember! “

This is why when the first yellow pixel appears, the prediction error is very large, and when the 100th yellow pixel appears, the brain has adjusted the “probability model” through the previous 99 predictions to 100 At that time, it can be predicted quite accurately.

-Nino Bosikashvili-

How does predictive coding explain cognitive processes? Let’s see an example. Visual perception has a phenomenon called “binocular competition”. If we use a special machine to present a house to your left eye and a human face to your right eye. While keeping our eyes receiving stimulation, we will see the face in a while and the house in a while. That is to say, our visual perception’s interpretation of constant stimuli switches between “face” and “house”.

So, what “algorithm” does the brain use to make the stable “input” finally show the unstable “output”? The description of the prediction code is quite reasonable: within a certain period of time, our brain forms an optimal prediction for the information input by the outside world (binocular stimulation), such as “I see a house on the screen”. The “house” stimulus passed in from your left eye is consistent and predicts success;

But at the same time, this prediction cannot match the “face” stimulus from your right eye, which produces a prediction error signal. This signal prompts the brain to find another explanation for the stimulus, which is “I see There is a face on the screen “. At this time, the visual perception switches to the” face “mode, but at this time, the original” house “stimulus input has caused a new error signal. So back and forth, you will see the face in a while and the house in a while. In general, this unstable alternating phenomenon of perception is actually the result of the “prediction error” caused by the rotation of house signals and face signals.

At this point, we have learned the core function of predictive coding: “prediction error mi”nimization). Supporters of predictive coding theory believe that the correction and minimization of prediction errors are the core goals of human mental activity. In other words, “reducing prediction errors” can explain all human behavior and mental processes, including attention, learning, memory, actions, emotions, and motivations, which are essentially processes of receiving information, forming assumptions, and correcting errors.

It’s quite like “Great Unity Theory”.

“crack” predictive coding

The popularity of predictive coding is undoubtedly related to its explanatory power. In his book The Predictive Mind, Jokob Hohwy said: “The brain is a complex hypothesis verification system, all it does is receive information from the outside world, And constantly reduce the prediction bias for this information. This mechanism aims to explain all mental activities from perception to action. “

Is it a little too much to bear the excitement? Want to immediately join the vast ocean of fancy prediction models? But wait, what seems wrong? What did we say at the beginning?

Smart, you found a strange contradiction: If the ultimate task of the brain is to eliminate all prediction errors, then, as the “error terminator”, why don’t we simply find a dark room and curl up? Here, the external information is constant, there is no new stimulus, and the brain is always in a comfortable state of “in line with expectations”.

This is the “The Dark Room Problem” for predictive coding theory. The principle of “minimization of errors” leads to a paradoxical assumption: we should strive to find the most boring experience while avoiding all possible “surprises”. Can predictive coding theories overcome this problem and tell us why people cannot tolerate a constant environment?

-Francesco Bongiorni-

Is the “black room question” a serious question? Maybe you, like many people, don’t think so. There are usually several ways to answer “Why not stay in a dark room”:

• “You must be hungry? You have to leave the room to find food!” That’s right, but at the same time it didn’t answer the original question at all. According to predictive coding, if the goal of the organism is only to reduce the prediction error, then only if one state increases the prediction error (for example, in the previous example, the signal error between the “face” stimulus and the “house” perception leads (Conscious interpretation is converted to human face), then it will prompt the organism to change state to achieve “expected accuracy”. However, for someone who is doing nothing in the dark room, the predictability of the hunger signal is quite high. Since the signal “hunger” introduced by the body matches the signal “hunger” predicted by the brain, the organism lacks a mechanism to leave the room.
  

  • “We are curious, we want to explore the world!”

    Long-term, leaving the dark room may allow us to better predict the world, and exploring the unknown may improve our ability to predict. This response still does not capture the essence of the black room problem. But even Andy Clark, one of the representatives of predictive coding, admits that even in the long run, the motives that drive us out of the dark room cannot be collectively referred to as “the instrumental value of exploration.” We dance, ride roller coasters, aid charities, and read poetry; we are even deliberately seeking unexpected “surprise” and “stimulation”. Even with resorting to longer time and space scales, predictive coding does not interpret all behavior as an effort to “reduce errors” as it expected.
    
    

    • “Did evolution not include the inherent motivation in our genes? Maybe our nature is to seek fresh experiences!”

      Yes, but this interpretation sacrifices most of the explanatory power of predictive coding itself. The most exciting aspect of predictive coding theory is that it is an independent and powerfully interpreted mental “theory of all things” that combines all aspects of extremely diverse mental activities under a single principle. However, if this principle needs to resort to external supplements and explanations, then it is no longer a complete theory of unity.
      
      

      In addition to the above unsuccessful explanations. Can predictive coding theory itself overcome the black room problem? The leading scientist in predictive coding, neuroscientist Karl Friston (Karl Friston) also answered this question.

      Friston believes that the black room problem deviated from the right direction from the beginning. This is because it incorrectly assumes that the prediction bias caused by the black room is very low! “The black room itself is actually very” unexpected “because our prediction is that we will not stayBlack room ’. “So,” staying in a dark room “actually caused a high prediction error. At this time, the mechanism of” minimizing error “will drive us out of the room. Unlike the previous intuitive responses, Fries Dayton ’s answer seems to solve the problem. “Minimizing prediction error” itself explains why we do n’t want to live in a black house.

      However, this answer puts the “Great Unity Theory” in a more dangerous situation.

      Self-prediction vs. self-reinforcing

      Actually, predictive coding is not the first player to refer to the “great unification theory”. It has a very famous predecessor: reinforcement theory.

      Half a century ago, two scientists who had profoundly transformed the field of cognitive research-Skinner and Chomsky-confronted the question of whether “reinforcement theory can explain all human behavior.” In a series of rebuttals by Chomsky against behaviorism, there is a point that is not particularly eye-catching: Chomsky enumerates a series of daily activities, and none of them seem to stem from strengthening principles, or “rewards.” . Chomsky said that regardless of whether it is an adult or a child, we will talk to ourselves, we will hum when there is no one, we will imitate the sound of cars and airplanes, and there will be no reward for doing so. So, how does behaviorism explain such behavior?

      Skinner ’s answer is “self-reinforcing”: we speak to ourselves because doing this makes us feel satisfied, that is to say, the so-called “self-reward” can strengthen our behavior.

      In this regard, Chomsky pointedly pointed out that resorting to “self-reinforcement” actually dispelled the explanatory power of behaviorism. This type of explanation is either wrong (is it true that self-reporting is self-reward?), Or it is meaningless superfluous (all behaviors can be said to be “self-reinforcing”). A “mechanism” that can explain anything can ultimately explain nothing, because it is empty and unfalsifiable. Why would a person read a book, play a piece of music, say something, etc …? Oh! Because he reinforced these behaviors. The term “strengthening” has lost its explanatory power.

      Will Friston’s defense of the “self-prediction” of the black room problem put predictive coding theory into the same position? possible. Why do we dance? Because we predict that we will not stand still; why do we donate to charities? Because we predict that we will do good; why do we need to socialize (and hate social isolation)? Because there is such an expectation in our brain, “Ah, our brain doesn’t like staying alone.”

      On the one hand, these explanations may imply important principles of the mind, but on the other hand, these answers may not be called explanations or mechanisms at all. If we only resort to the self-expectation of “we will not stay in a black room” to overcome the “black room problem”, then we may neverNo behavior can be explained by predictive coding, because all behavior can be explained by a reverse self-prediction.

      It is true that predictive coding promotes our understanding of human mental processes in many ways. The “black room problem” is not intended to deny these important contributions. However, when a theory tries to explain everything about humans as a “universal theory” of mental science, we have to be cautious.

      Postscript

      Finally, at the end, please allow me to speak for my laboratory and recommend our opinion article The Dark Room Problem published in Trends in Cognitive Sciences. A little spoiler, the article has an interesting argument not mentioned here. A big spoiler, we have received the letter from Friston to the magazine, and we are talking about science issues friendly. If you are interested in the subsequent plot, please stay tuned!

      References

      Sun, Z., & Firestone, C. (2020). The dark room problem. Trends in Cognitive Sciences, 24, 346–348.

      Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and brain sciences, 36 (3), 181-204.

      Clark, A. (2016) Surfing Uncertainty: Prediction, Action, and the Embodied Mind, Oxford University Press, New York

      Chomsky, N. (1959) Review of Verbal Behavior by B.F. Skinner. Language 35, 26–58

      Friston, K. (2013) Active inference and free energy. Behav. Brain Sci. 36, 212–213

      Friston, K. et al. (2012) Free-energy minimization and the dark-room problem. Front. Psychol. 3, 130

      Klein, C. (2018) What do predictive coders want? Syn- these 195, 2541–2557

      Hohwy, J. (2013) The Predictive Mind, Oxford University Press, Oxford