The Brain from Inside Out

by György Buzsáki

As an alternative to the empiricist outside-in tabula rasa view, I raise the possibility that the brain already starts out as a nonsensical dictionary. It comes with evolutionarily preserved, preconfigured internal syntactical rules that can generate a huge repertoire of neuronal patterns. These patterns are regarded as initially nonsense neuronal words which can acquire meaning through experience. Under this hypothesis, learning does not create brand new brain activity patterns from scratch but is instead a “fitting process” of the experience onto a preexisting neuronal pattern.

While a plot such as that in Figure 2.1 shows only the values of observations, we often wish to read much more from a static cloud of data points. For example, there are many empty areas without measured values on the graph. We can simply assume that those values exist so that, if we had more time and more funding to generate more data, those new points would fill in the blanks between the already-measured ones. This operation, performed intuitively or with mathematical rigor, is called interpolation (the arrow in Figure 2.1), but we can also call it explanation or deduction. Postdiction would be another possible term because interpolation is done based on data collected in the past. If we feel comfortable filling the graph with imagined data between the dots, perhaps it is not much of a jump to imagine more data that extend the cloud, again following the trend in the observations. This operation is called extrapolation or induction. Alternatively, we can call it prediction or guessing from the currently available data, an act of “logical inference.”5 It is an abstraction from the present, which amounts to predicting the future. How far can we go in our extrapolation of the graph? Just because two sets of variables correlate now, there is no guarantee that this correlation will continue into the future.

In Figure 2.1, we identified some regularity between two variables: one is the thing-to-be-explained (explanandum), called the dependent (y) variable, and the other is the thing-that-explains (explanans), called the independent (x) variable.7 In more general terms, we can also call them the “defined concept” and the “defining concept.” Of course, our strongest desire is to conclude that one variable predicts or causes the other. Ideally, the defining concept is more general than the defined concept, with some asymmetry between them. “The brain is a machine”; however, the other way around—“the machine is a brain”—does not apply. Accordingly, our choice to call one variable independent and the other dependent is often arbitrary and may reflect our preconceived bias.

Yet much of the world has moved forward and onward without accepting such causality rules. The idea of causation is quite different in Asian cultures than in European cultures. For example, in Buddhist philosophy, nothing exists as a singular and independent entity because things and events depend on multiple, co-arising conditions.12 A Buddhist monk or teacher lights three incense sticks leaning against each other, as a demonstration that if any of the sticks falls, the others will fall as well.

A classical demonstration for the critical role of action in perception is the tactile-vision sensory substitution experiment by Paul Bach-y-Rita and colleagues at the University of Wisconsin, which allowed blind people to “see” using an array of electrodes placed on the tongue and stimulating the sensory terminals with weak electrical pulses.

The short message of this long chapter is that perception is an action-based process, an exploration initiated by the brain.

The critical physiological mechanism that grounds the sensory input to make it an experience is “corollary discharge”: a reference copy of a motor command sent to a comparator circuit from the action-initiating brain areas. This comparator mechanism allows the brain to examine the relationship between a true change in the sensory input and a change due to self-initiated movement of the sensors. The same corollary discharge mechanism also serves active sensing, the process by which sensory receptors can be most efficiently utilized to sense the environment.

More complex brains are organized in a “multiple loop” pattern. These parallel loops, each with increasing levels of complexity of wiring and temporal dynamics, are inserted between the output and input (middle panel). For example, in mammals, the most direct circuit between sensors and muscle activation is the monosynaptic spinal cord/brainstem connection. More complex loops connecting these same inputs and outputs include the thalamocortical system and bidirectional communication between the neocortex and several subcortical structures. In addition, a side loop of the hippocampal system is dedicated to generate sequential neuronal activity such that content in the thalamocortical loops is organized into orderly trajectories, allowing simultaneous postdiction (memory) and prediction (planning) operations (as elaborated in Chapter 13). Large and small brains share the same main goal: to predict the future consequences of their actions. However, the multiple interacting loops and the hippocampal postdiction-prediction system allow more sophisticated brains to predict action outcomes more effectively at much longer time scales and in more complex environments.

In contrast to the entorhinal grid patterns, hippocampal neurons display highly flexible spatially tuned patterns, called place fields. Typically, in a given environment, most pyramidal neurons are silent, and the minority forms a single place field: that is, the neuron becomes active only when the animal reaches a particular location in the testing apparatus. This response is why John O’Keefe named them place cells.

The transmembrane potential of neurons can be measured by electrodes, which can report neuronal activity with submillisecond time resolution. Electric current flows from many neurons superimpose at a given location in the extracellular medium to generate a potential with respect to a reference site. We measure the difference between “active” and reference locations as voltage. When an electrode with a small tip is used to monitor the extracellular voltage contributed by hundreds to thousands of neurons, we refer to this signal as the local field potential (LFP). 10 Recordings known as the electrocorticogram (ECoG) are made with larger electrodes placed on the brain surface or its surrounding dura mater to sample many more neurons. Finally, when even larger footprint electrodes are placed on the scalp, we call the recorded signal an electroencephalogram (EEG).

There are numerous brain rhythms, from approximately 0.02 to 600 cycles per second (Hz), covering more than four orders of temporal magnitude (Figure 6.1).12 Many of these discrete brain rhythms have been known for decades, but it was only recently recognized that these oscillation bands form a geometric progression on a linear frequency scale or a linear progression on a natural logarithmic scale, leading to a natural separation of at least ten frequency bands.13 The neighboring bands have a roughly constant ratio of e = 2.718—the base for the natural logarithm.14 Because of this non-integer relationship among the various brain rhythms, the different frequencies can never perfectly entrain each other. Instead, the interference they produce gives rise to metastability, a perpetual fluctuation between unstable and transiently stable states, like waves in the ocean. The constantly interfering network rhythms can never settle to a stable attractor, using the parlance of nonlinear dynamics. This explains the ever-changing landscape of the EEG. Traditionally, the frequency bands are denoted by Greek letters.15

The different oscillations generated in cortical networks show a hierarchical relationship, often expressed by cross-frequency phase modulation between the various rhythms. This term implies that the phase of a slow oscillation modulates the amplitude of a faster rhythm, meaning that its amplitude varies predictably within each cycle. In turn, the phase of the faster rhythm modulates the amplitude of an even faster one and so on. This hierarchical mechanism is not unique to the brain. Spring, summer, fall, and winter are four phases of a year that “modulate” both the amplitude and duration of day length.

Inhibition is the foundation of brain rhythms, and every known neuronal oscillator has an inhibitory component. Balance between opposing forces, such as excitation and inhibition, can be achieved most efficiently through oscillations. The output phase allows the transmission of excitatory messages, which are then gated by the build-up of inhibition.

As discussed in Chapter 3, the fundamental properties of myosin and actin are largely conserved across mammals. Therefore, the motor command computations in the motor cortex, cerebellum, and basal ganglia should be performed in comparable time windows, and the command signals to the spinal cord should be delivered within the same time range in different species. However, the distances of these structures vary by orders of magnitude across species. Thus, all of the timing constraints required for adequate function have to be reconciled with the complexity imposed by the growing size of the brain. This is not a trivial task, given a 17,000-fold increase of brain volume from the small tree shrew to large-brain cetaceans. The constancy of the many brain oscillations and their cross-frequency coupling effects across species suggest a fundamental role for temporal coordination of neuronal activity.22

There appear to be at least two mechanisms that allow scaling of neuronal networks while conserving timing mechanisms. The first mechanism compensates for the increase in neuronal numbers and the enormous numbers of possible connections by shortening the synaptic path length between neurons, defined as the average number of monosynaptic connections in the shortest path between two neurons. This problem is akin to connecting cities with highways and airlines to obtain an efficient compromise between the length of the connections and the number of intermediate cities required to get from city A to city B. Directly connecting everything with everything else is not an option in most real-world networks as the number of connections increases much more rapidly than the number of nodes to be connected. An efficient compromise can be achieved by inserting a smallish number of long-range short-cut connections into local connections. The resulting “small-world-like” architectural solution allows for scaling while keeping the average synaptic path length similar as brains increase in size. While such scaling prevents excessive volume growth, longer axons increase the travel time of action potentials. This would pose a serious problem for neuronal communication. Thus, a second mechanism is needed to compensate for these time delays. This requirement is solved by using larger caliber, better insulated, and more rapidly conducting axons in more complex brains.

Similarly, babbling in human babies may reflect a self-organized intrinsic dynamic. When the uttered sounds resemble a particular word, the happy parents regard them as a real word. They reinforce such spontaneous utterances with a corresponding object, action, or phenomenon, until it acquires a meaning for the baby. Exploitation of the default self-organized patterns of the brain prior to language is a more effective mechanism of pattern formation than a de novo, blank-slate solution. (Chapter 13).

Birdsong and grooming are examples of neuronal trajectories that can be interpreted by relatively simple reader mechanisms due to their limited variability. In contrast, large recurrent circuits can generate numerous trajectories, each of which can become a meaningful pattern when matched to experiences. The richness of the information depends not on the sequence pattern generator but on the reader mechanisms. This may explain why the neocortex (reader) is so much larger than the sender (hippocampus).

The brain areas in charge of generating plans and thoughts share many similarities with the motor cortex in terms of cellular architecture and input–output connectivity. The main difference is that prefrontal cortex does not directly innervate motor circuits. Instead prefrontal cortical areas can be designated collectively as an internalized action system, and thus plans and thoughts can be conceived of as internalized neuronal patterns that serve as a buffer for delayed overt action. Thoughts are only useful if they lead to action, even if that action is delayed by days or years. These same brain areas and mechanisms are also responsible for externalizing thought in the form of artifacts, language, art, and literature.

Discussions of space and time often cite the philosopher Immanuel Kant, who argued that these concepts are a priori categories that cannot be studied directly. “Space is nothing but the form of all appearances of outer sense . . . [that] can be given prior to all actual perceptions, and so exist in the mind a priori, and . . . can contain, prior to all experience, principles which determine the relations of these objects.”4 Space and time are thus the axioms of the universe, orthogonal to each other and independent from everything else.

Turning the volume knob on your radio is a gain control mechanism for sound production.6 Gain modulation requires two sources, an amplifier and a modulator (Figure 11.1). When the sources are multiplied (or divided), such interaction generates an output whose magnitude is larger or smaller (in which case the gain is negative) than expected from adding (or subtracting) the sources. Such nonlinear gain modulation is ubiquitous in the brain, with far-reaching consequences, and is achieved by different mechanisms in individual neurons or synapses or at a larger scale in neuronal networks, indicating that it is a critical operation.

Still, visual neurons have to deal with a wide dynamic range. The solution is to respond not to the absolute magnitude of the input but to a relative change in intensity, known as contrast. By receiving information about both the mean background intensity and about the deviation from this mean, neurons can calculate the ratio or contrast against any light intensity. This process is a type of normalization, to borrow a term from statistics—essentially a division.7

A third family of gain control mechanisms acts on the inputs to the neurons. When many nearby inputs on a dendrite are activated, they amplify each other’s effects so that their impact is larger than their algebraic sum.12 Even more selective gain control, acting at single synapses, can be achieved via the short-term plasticity of synaptic strength. The synaptic efficacy of the action potential from a presynaptic neuron is not fixed over time but shows large fluctuations. Some synapses increase (“facilitatory synapses”) whereas others decrease (“depressing synapses”) their response to the later input spikes in a series.13

In short, selective attention multiplies or “gain-modulates” the neuron’s response by a constant factor without changing its tuning characteristics. Attention-induced gain modulation works well at all stages of the visual-posterior parietal cortical system28 and likely in many other brain areas.

Modern neuroscience, like early physics, is striving to describe and condense observations with mathematical equations. Compared to physics, neuroscience has only a modest number of laws. 14 Yet, one of these, the Weber law or also called the Weber-Fechner law has a breathtaking simplicity and generality. This law is named after two German scientists who laid the groundwork for human psychophysics, a quantitative investigation of the relationship between physical stimuli and the mental states they induce. Ernst Heinrich Weber was a physician who was interested in how we can perceive differences in touch. After extensive experimentation, he concluded that “in observing the disparity between things that are compared, we perceive not the difference between the things, but the ratio of this difference to the magnitude of things compared.” For example, if you are holding a 100-gram object, the second object needs to weigh at least 110 grams in order for you to notice a difference. If the object weighs 200 grams, you can detect a weight difference only if the other object weighs more than 220 grams or less than 180 grams. This threshold change, called the Weber fraction, is also called the just noticeable difference.15

Gustav Theodor Fechner did not conduct experiments. Instead, we can call him an early computational scientist. He trusted Weber’s observations and calculated mathematically that sensation is a logarithmic function of physical intensity. Therefore, when stimulus strength multiplies, the strength of perception adds.

Perhaps the most common skewed distribution in biology is the logarithmic-normal or log-normal distribution.23 This distribution is right-skewed on a linear scale but looks bell-shaped when the logarithms of the observed values are plotted. In other words, a log-normal distribution is a probability distribution of a random variable whose logarithm is normally distributed. Examples from biology include the number of species per family, survival times of species, sizes of fruits, pharmacological effects, times to first symptoms of infectious diseases, and blood pressure distribution within an age group, just to list a few. Basic physiologic processes, such as resting heart rate, visual acuity, and metabolic rate, also show a skewed distribution. In economics, incomes are often well fitted to a log-normal function.24

The log-normal distribution also describes the firing rates of cortical pyramidal neurons.

However, good enough is far from perfect. We would not want to drive a car with 60–80% accuracy or submit a scientific paper with such precision. To perform better, we also need to deploy the second virtual brain: a large fraction of slow-firing neurons with plastic properties that occupy a large brain volume connected by weaker synapses into a more loosely formed giant network. Their work is absolutely critical for increasing the accuracy of brain performance. Of course, I am not thinking about two discrete brains in one skull, but instead a continuum of a broad distribution of mixed networks that performs apparently different qualitative computations at the left and right tails. This distribution allows the brain to implement anything from preexisting rigid patterns to highly flexible solutions. Thus, there is no definable boundary between the “fast decision, low precision” and “slow decision, high precision” networks.6 Brain performance is a tradeoff between speed and accuracy. After one of my talks on this topic, somebody commented: “The dichotomy you were talking about reminded me of Daniel Kahneman’s book Thinking, Fast and Slow.

Mainstream psychiatry, like cognitive neuroscience, is also under the historical influence of the empiricist representational framework. The Diagnostic and Statistical Manual of Mental Disorders (DSM-5) in the United States is another prominent example of how human-conceived terms are used in attempts to draw boundaries among mental disorders. It is the “James’s list” of psychiatry. Each new edition is an attempt to fix problems with the previous ones. According to the American Psychiatric Association, the “goal in developing DSM-5 is an evidence-based manual that is useful to clinicians in helping them accurately diagnose mental disorders.” This aim reflects a collective subconscious desire to reduce a high-dimensionality problem, such as a complex psychiatric condition, to a single word or phrase: the diagnosis. Dimensionality reduction in statistics and information theory refers to a mathematical process of reducing the large number of variables under consideration to extract a set of principal variables or features using methods such as principal component analysis or nonlinear dimensionality reduction.3 Of course, these methods can only be as good as the data we supply to them. A further problem is that “humans are incorrigibly inconsistent in making summary judgment of complex information,” as expressed succinctly by the economist Daniel Kahnemann.