The hierarchical organization of the brain finds many parallels in the hierarchical organization of silicon. Hierarchies The hierarchical structure of the brain can be related to the principle of hierarchical organization of computation.
Introducing the Brain
Evolution discovered a general way to regulate the state of the body with chemical signals. The core calculations for all these tasks are performed by the forebrain, which is the focus of the next chapter.
Computational Abstraction
But among all these jobs, the operating system runs the program you may have written—the user program. The reason is that the lower level instructions are too detailed to suit the details of the machine.
Different than Silicon
From these analyses, the consensus is that the way the brain should do this is to precompute most of the responses in a tabular format so that they are simply looked up. As shown in Figure 1.5, the size of the gates in silicon is comparable to the processes of a living cell.
The Brain ’ s Tricks for Fast Computation
Limited input and output sizes When analyzing sorting algorithms on silicon computers, it is assumed that the dominant factor is the size of the input. For now we will look at some of the pessimistic views.
More Powerful than a Computer?
A program can be represented by a table showing what to do for each state and input. One important message from this point of view is that as much as the universe can be described by calculus, so probably can our little brains.
Do Humans Have Non-Turing Abilities?
Therefore, the referents of logical statements are regularized and no special machinery is required. 19 His main theme was that ever-increasing abstractions of physical systems are necessary to tame the chaos of atomic levels.
Summary
If you want to show that computer science is a good model, then you need to show how the things people do can be described by recipes or programs that exploit brain neurons in a workable way. The introduction to the brain in the previous chapter highlighted the bigger picture by focusing on the parts of the brain that have essential regulatory or life-sustaining functions.
Spinal Cord and Brainstem
As we'll get to in a moment—and we'll spend a lot of time on this in Chapter 4—the cortex has the job of encoding sensory-motor states in a vast table in a way that makes every response lookup. 24 Basic drives in the hypothalamus set a plan that the rest of the brain tries to satisfy.
The Forebrain: An Overview
The state of the cortex at any given time can be thought of as analogous to the real web page for the user. This new state again sets up another cycle of the cortex - basal ganglia - thalamus loop. The cortex calculates the state of affairs and a list of what needs to be done, and the basal ganglia perform one of the actions.
Cortex: Long-Term Memory
By now, the reader should have a good idea of the overall functioning of the brain's subsystems. This means that their responses are sensitive to the position of the gaze points of both eyes. We can see that the gaze vector rapidly scans the image in very different patterns motivated by different questions posed to the viewer.
Basal Ganglia: The Program Sequencer
Tonically active neurons make up only 10% of the neurons in the basal ganglia, but they play an important role in signaling task segments. These data make the following important point: The basal ganglia do not define the details of movement. The view of the basal ganglia as a general program sequencer is further supported by observations of an important brain function called working memory.
Thalamus: Input and Output
For the next step, you need information on how you will measure when the water boils and the location of the cup. Of course, after discussing the role of the cortex, you know that the main locus of information in working memory is the cortex, but the basal ganglia must refer to this information in order to do its job. The time for the signal to reach the thermal sensors of the palm to the cerebral cortex is on the order of hundreds of milliseconds.
Hippocampus: Program Modifications
The hippocampus (figure 2.16) plays the central role in the permanent analysis and recording of the series of momentary experiences. In addition to this account of the expansion of permanent storage, however, another extremely important issue is the use of the hippocampus in interaction with stimuli in the here and now. The mainstream research view is that novelty detection in the hippocampus contains the information that makes this adaptation possible.
Amygdala: Rating What ’ s Important
Such an ability is likely to be scalable and applicable to a general cortex–thalamus–basal ganglia loop that includes the hippocampus. This loop can be run in real time, during behavior guidance and action selection guidance, as well as in “planning time” where it can use the intended outcomes of actions to explore their ancillary consequences. In this case, the evidence suggests that amygdala circuitry is used to maintain a more direct encoding that elaborates on the details of the imminent error.
How the Brain Programs Itself
It seems likely that in this case the amygdala handles the encoding in a simple way that bypasses or complements the cortico-hippocampal pathway. The hippocampus encodes the difference between the current situation and remembered programs, and the amygdala scores this difference in terms of importance. This hippocampus continuously cuts the interpretations of the current context into a format that corresponds to existing stored programs.
Summary
Recent results relate to the role of the basal ganglia in learning and suggest roles for specific neurotransmitters. To understand the scope of the problem, one can start with the issues of levels of abstraction. We will describe one of the more speculative ones to illustrate some of the crucial computational issues.
Signaling Strategies
In the hippocampus, spikes are organized according to the beginning and end of the current task of interest. Cells encoding information at the beginning of the task send their information at 0 o (red), while cells at the end send their spikes at 360 ° (green). C) A very different code is used by tonically active cells (TAN) in the basal ganglia. What happens when the movement is in the opposite direction to the one it is most sensitive to.
Receptive Fields
Only two—the LGN and the first stage of visual cortex V1—are discussed, but later stages embody the same kind of thinking. One can have collections of edge cell neurons that respond to distinct edge orientations in the image. Because there are now many neurons, denote the receptive field of k by w k.
Spike Codes for Cortical Neurons
In earlier development of the formation of receptive fields, synapse strength and firing codes were numbers in models. This neuron is taken from an area in the cortex that is measuring motion in a small area of visual space. This would be similar to that used by the hippocampus in the θ frequency band, but tuned to the much higher frequency γ band.
Reflexive Behaviors
If neurons in the visual field signal a moving object of the right size and depth, the frog sticks out its tongue to get it. In this kind of reflex, the response is hardwired in the spinal cord and there is no sophisticated differentiation of the stimulus. But in any case, they can still be used in the role of function detectors in decision-making.
Summary
Amazingly, when Newsome recorded the responses of neurons in a motion-sensitive part of the cortex, the firing rates of individual neurons increased in a way that paralleled the monkey's success at patterning the task. Although on the surface it appears that the motion neurons in area MT perform the same kind of function as those in the frog retina, there is a very important difference: the cortical neurons are trained to do this work. Given the vast amount of different neurons in the cortex, this experiment indicates the enormous flexibility the cortex has in creating different states that can be used by programs.
Appendix: Neuron Basics
When the axon (blue) arrives at a target cell (red), its terminal bud encounters one of the cell's dendrites where it makes a connection. What you see is a beautiful cross-sectional photo of a dendrite (de) that has a raised spine (sp) to contact an incoming axon terminal (bo). Finally, the overall modulation of the circuits of which the cell is a part is handled in another Figure 3.21.
Table Lookup Strategies
As you know, we can do this by describing the state of the network in terms of a state vector x. All we have done is to designate a single symbol to cover the entire state of the network. As time evolves, the panels in figure 4.1 have a different state vector for each of the time steps.
The Cortical Map Concept
A more detailed inspection of the map reveals another important feature of the vision, and that is that the central area of the pattern is exaggerated in the cortical area. This naturally reflects the fact that the central region of the visual field is sampled in the retinas by their high-resolution foveas. Oriented cells can be sensitive to movement perpendicular to their orientation in one of two possible directions.
Hierarchies of Maps
To appreciate these findings, it helps to recall Figure 3.6 in Chapter 3, which represents the layers of the cortex. Van Essen, “Connections of the middle temporal visual area (MT) and their relationship to cortical hierarchy in the macaque monkey,” Journal of Neuroscience , vol. Maunsell's Figure 4.8 summarizes some visual maps, again for the monkey, emphasizing this dichotomy.
What Does the Cortex Represent?
The result is that such areas can be detected and displayed on top of the conventional optical image, as is done in Figure 4.11. Colored highlighted areas show areas of cortex that responded to the different stimuli shown in the immediate right panels. This result is also important as the analysis shows that the results can account for approximately 22% of the BOLD signal, a huge improvement over other methods.
Computational Models
You can see how this might work in the context of the border ownership case in Figure 4.10. A discussion of Bayesian networks assumes that the networks have been trained and their connectivity is established. When trained exclusively on face images, higher-level networks form features that capture the essential features of a face.
Summary
We already introduced the former in the previous chapter as one of the most important tasks of the cortex. How the brain organizes the details is still a mystery, but in broad terms the program of states and actions would of course be encoded in the cortex and the execution of the program would be handled in the basal ganglia. Especially if the effort required to achieve this exceeds the value of the apple, one should not venture out in the first place.
Evaluating a Program
A program may be socially destructive, but from the perspective of one's internal accounting system, it seems like the best thing to do. At the end of the program you can easily assign a value to it based on the result. But with the status and action chart in hand, one can score the program's states that are next to last with respect to the ending.
Reinforcement Learning Algorithms
Formally, we have entered the realm of Markov decision processes (MDPs), where the state and action description is augmented by (1) a probabilistic transition function that determines action outcomes and (2) a reward function that scores the expected value of action taken . The policy improvement theorem requires a specification of a policy and an additional bookkeeping mechanism to keep track of the value of states and actions. In other words, the value of a state is the value of the best action available from that state.
Learning in the Basal Ganglia
The coding of actions in the basal ganglia is one part of the reinforcement learning story; the other component is the encoding of reward itself. Now the reward signal moves to the time of this signal, or the start of the more extensive program. However, in the absence of the predicted reward, a correction signal occurs at the time of expected reward.