Brain Modeling with Brytes

Making Big Brains from a Lot of Little Brains

David Zipser

Institute of Cognitive and Brain Science, University of California Berkeley, 1408 Milvia St, Berkeley, CA 94709, U.S.A.

Keywords: Arm Movements, Brain Models, Distributed Decision Making, Cognitive Control.

Abstract: Brytes are small brains used as subunits to model the cognitive processes of larger, smarter brains. A

previously developed model of scratching behaviour that uses brytes to generate the coordinated movements

of two arms, one with the itch site the other with the scratching hand is described. Then new strategies are

described for using large sets of brytes with virtual locations all over the body to make decisions about

whether scratching is safe in the current context and, if so, which appendage to use. Finally, the biological

plausibility of brytes is examined in the contest of brain evolution and brain functional architecture.

1 INTRODUCTION

It has long been believed that explaining the

extraordinary cognitive abilities of the brain will

require the use of large numbers of high-level sub-

systems. From Minsky’s Society of Mind (Minsky,

1988) to the more recent Mixture of Experts (Jacobs

et al, 1991) many such subsystems have been

proposed. However, none of these are well suited for

building biologically plausible brain models

I have conjectured that to get the right kind of

subunit we should assume that big brains are made

of many smaller brains (Zipser, 2009, 2010).

Computations can then be simplified by distributing

them to many brain-like subunits. This approach has

already been used to build systems that generate

complex motor behaviours. Here I describe how to

extend the paradigm to a higher cognitive level that

is able to make decisions about when and how to

execute a motor behaviour

Thinking about brains made from brains is

confusing so a name is needed to distinguish the

subunit brains from the whole brain. I mixed ‘brain’

with ‘byte’ and got bryte. Brytes are the little brains

of which big brains are made. They have a complete

set of all the usual brain-like abilities, but on a small

scale and adapted to their role as part of a bigger,

smarter brain

Whole brains, even the brains of primitive

animals, are very complex so it is not practical to try

to define such a brain mathematically. Instead of

such an a priori definition, the computational

structure of brytes can be built up incrementally by

using them to implement functioning models of

successively more complex cognitive tasks. This is

the approach taken in this paper

From the bryte point of view brains evolved by

incorporating more and more brytes while each of

these brytes slowly mutates to compute a different

aspect of cognition, Figure.1

From a computational point of view brytes do

computations in a distributed way with each bryte

contributing to the task. This can only work if there

is an efficient way to integrate the contributions of

each byte

In this paper I show how local decisions made by

large groups of brytes can be combined to decide

when an agent can act safely, and how best to

accomplish the action. These higher level cognitive

decision tasks are built on a previously implemented

model that uses distributed sets of brytes to generate

the complex coordinated arm movements needed

when the hand on one arm scratches an itch on the

other arm (Zipser, 2012)

The previously implemented itch-scratching

model is reviewed here in some detail to provide a

concrete example of how the local computations

done by brytes can be integrated to generate the

coordinated movement of whole appendages.

124

Zipser D..

Brain Modeling with Brytes - Making Big Brains from a Lot of Little Brains .

DOI: 10.5220/0005036901240129

In Proceedings of the International Conference on Neural Computation Theory and Applications (NCTA-2014), pages 124-129

ISBN: 978-989-758-054-3

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: Schematic depiction of brain evolution. The

‘command’ symbol is used to represent a basic unit brain

or ‘bryte’. All brytes have access to sensory input and can

contribute the final common path for motor output.

2 USING BRYTES TO

GENERATE COORDINATED

ARM MOVEMENTS

Lets start at the last step in the process of scratching

after the decision to scratch has been made and an

appendage chosen to do the job. As we scratch an

itch on one arm with the opposite hand the

scratching arm continuously changes posture to

bring the hand to the itch while the itch arm moves

to make the itch site more accessible and bring it

closer to the scratcher. These coordinated

movements are made without the arms touching

each other. All the information for this action is

available directly within the nervous system so it can

be done with the eyes closed—vision is not required

Figure 2: Brytes are physically located in the brain, but

can have arbitrary and changing virtual locations in the

outside world. In the scratching model many brytes are

used with virtual locations all over the arm surface. A

sample of this distribution is shown in the blue ovals.

Examples of an avoider bryte, blue, and a seeker bryte,

red, are shown with the movement vectors they output.

The output vectors of all the brytes are summed and

passed to the gradient arm movement algorithm to

generate movements (See text for details).

These scratching arm movements were

previously modelled using brytes as described

elsewhere, (Zipser, 2009, 2010), and will be only

briefly reviewed here. The model uses a gradient-

based optimal control technique for making goal

directed movements with multi-jointed (Todorov,

2006: Torres & Zipser, 2002). This technique moves

a point on an arm toward, or away from, an external

goal point using gradient decent on the function that

relates joint angles to the spatial locations of point

on the arm. Because the gradient is computed

continually and incrementally so the goal can also

move. The arm movements are actually controlled

for both translation and rotation so both the location

and the orientation of the goal can be matched.

Movements are generated de novo in real time. No

pre-computations, motion capture or learning are

used. Most important for the scratching problem is

the fact that the gradient is a linear operator so

gradients for the movement of different points on an

arm toward different goals can be added to get a

single movement. This allows moving toward one

itch point while avoiding all other points on the

contralateral arm.

To understand how brytes are used to do the

scratching task, imagine that a single ‘seeker’ bryte

with a virtual location on the tip of the scratching

finger, red bryte in Figure 2, has access to the spatial

location of itself and the moving itch. This bryte can

compute how to move to bring itself closer to the

BrainModelingwithBrytes-MakingBigBrainsfromaLotofLittleBrains

125

itch site on the contralateral arm. This function is

coded in parietal cortex neurons (Ferraina et al,

2009). A different seeker bryte at the itch site on the

other arm can at the same time move the itch site

toward the finger that is trying to scratch it. This will

eventually bring the scratcher to the itch, but

unfortunately the arms will not only collide, they

will also try to move through each other. To prevent

this, a several hundred ‘avoider’ brytes are scattered

over the whole surface of both arms. These brytes,

blue bryte in Figure 2, have access to the location of

many points distributed on the contralateral arm.

Using this information each avoider bryte, computes

a direction to move away from the other arm so as

not to hit it. This function is coded in neurons in pre-

motor cortex (Graziano, et al, 1994). The

appropriately weighed sum of the movement

directions of all these avoider brytes will keep the

arms from hitting. When the movements specified

by the scratcher brytes are added to those of the

avoiders, the two arms move so that the itch is

scratched and collisions are avoided

Figure 3: Examples of starting and ending postures of both

arms moving to scratch an itch. In A, B and C the

movements are successful while in D a local minimum is

encountered.

Examples of starting and ending arm postures are

show in Figure 3. The important point is that once an

itch and scratch site are specified, brytes can be used

to get the job done. This depends on having a way to

combine together all the computations done by the

brytes from their local viewpoints. In this case a

simple summing operation was all that is needed.

Sometimes the movement vectors for scratching

and avoiding contact cancel and movement is caught

in a local minima before reaching the goal, Figure

1D. This problem arises in many behavioural

contexts. Particular solutions to individual cases are

usually easy to find, but more general cognitive

solutions are of more interest. Some of these general

solutions to local minima are implemented

elsewhere (Zipser, 2009, 2010) in bryte models of

grasping in the presents of obstacles

Movies of the running simulation and a GUI that

allows many factors to be manipulated are available

on line (http://crcns.org/data-sets/movements/zipser-

1/). The code for the simulation of this model is

quite concise and the simulation runs rapidly. Those

interested in the mathematical details of the

simulation and a discussion of how the required

information is represented in the nervous system can

find them here (Zipser, 2012).

3 DECIDING TO ACT

3.1 Selecting an Appendage to do the

Scratching

Suppose you have an itch on your right calf. You

could bend over and scratch it with either hand or

you could scratch it with the side of your foot on

your left leg. How do you know that you have these

options, and how do you decide which to use? This

is an example of a kind of problem you constantly

confront, so finding a fairly general solution is of

some interest. One general solution that is often

proposed is ’simulation’ i.e. imagining what will

happen if we do something and then not doing it if it

leads to bad results. Since we can apparently do

mental simulation, it seems reasonable to use it for

this kind of task. But I have found that attempts to

model mental simulation always involve some

method for evaluating the outcome. In what follows

I show that sometimes these ‘outcome’ evaluations

can be done without actually simulating the action.

Doing this involves using many distributed brytes

that are a bit brighter than the ones used so far

Each bryte in the coordinated arm movement

model made a contribution to the overall movement

based on its own point of view, i.e. virtual location.

These contributions were combined by summation

to get a global movement. In the same spirit,

imagine that brytes are distributed with virtual

locations over the whole body. If each of these

brytes can compute a value that increases with how

appropriate it is for it to be the scratcher bryte, then

the bryte with the highest value can be chosen to do

the task.

How do the brytes compute their own

appropriateness? There are innumerable factors that

can potentially go into the calculation of

appropriateness. For now we will consider only

two—a default appropriateness, and distance to the

itch. The default value is based roughly on how

likely a bryte is to be chosen as the scratcher, i.e. a

prior. Brytes on the fingertips would have high

NCTA2014-InternationalConferenceonNeuralComputationTheoryandApplications

126

defaults while those on the back would have low

values. The distance contribution to the

appropriateness value simply assumes that being

close to the itch makes a bryte more appropriate to

be the scratcher. A hypothetical example is given in

Figure 4 with only the three most appropriate brytes

shown.

Figure 4: Hypothetical example of the method proposed to

select the most appropriate appendage to use for

scratching. Only the three brytes with the highest priors

are shown. The local estimates of appropriateness are sent

to a central location where the maximum value is

determined and broadcast to all the brytes. The bryte that

produced this value decides that it is the one to do the

scratching.

To pick the scratcher bryte, the appropriateness

values have to be compared and then the bryte with

the largest one informed of its selection. In the arm

movement model the gradients were summed to get

a value that was then sent to a final path for

movement. Here the max rather than the sum is

centrally determined and sent to all the brytes. The

bryte that has this max value acts as scratcher. Note

that all the computations and decisions are made

locally except the max operation, which is context,

independent.

3.2 Deciding Whether to Make a

Movement

How is the decision made that scratching will be

safe in the current context? Here also there are many

factors, but to illustrate how the bryte paradigm can

be used, let us consider just one factor — static

stability, i.e. when there is no movement and no net

force.

Maintaining static stability requires that the ever-

present force of gravity be balanced by a support.

The weight of the body on the support produces

pressure that is measured by sensors on the body

surface. Since body weight is constant, summing all

these pressure measurements that there is sufficient

support to prevent falling. When we are sitting

down, for example the pressure between the chair

and our behind is enough to guarantee stability and

we can sit safely.

The safety of using a selected appendage for

scratching can be determined if brytes on chosen

scratcher set the value of pressure they send to a

central summing device to zero before any

movement is made. This reduces the summed value

of support pressure, indicating a possible loss of

stability. Broadcasting this value to all the other

brytes allows them to detect the possible instability

and output a value of pressure they can handle. This

is a local computation that depends on factors such

as the presence of a support and the strength of the

body part the bryte is located on. If the sum of all the

increased values the brytes can handle is large

enough the scratch will be safe. If, however, the sum

of pressure that can be compensated is too low, the

scratch is unsafe and can lead to a fall.

If stability is being maintained by standing on the

ground or sitting in a chair, there will be no

significant change in net force from moving a hand,

so using it to scratch will not produce instability.

However, if the chosen bryte is at a place that is

contributing a lot of force to stability—as in the case

of hanging by your finger tips— setting its force

vector to zero will have a large affect on the

predicted stability. This predicted instability cannot

be compensated since no other brytes are in contact

with a support. This projected instability can be used

to cancel the scratch.

Our brytes are getting brighter. They can direct

movement, estimate is effects and help decide what

body part should do it. None of these tasks seemed

to require a great amount of intelligence on the part

of the brytes, but the combined contributions of

many local brytes facilitate global cognition that at

first appeared quite difficult.

4 BRYTE WAY TO BRAIN

EVOLUTION

Are brytes biologically plausible? In the 600 million

odd years between the immediate chordate precursor

of vertebrates and the advent of humans the brain

increased in size about a million fold and got a lot

BrainModelingwithBrytes-MakingBigBrainsfromaLotofLittleBrains

127

Figure 5: Is it safe to Scratch? The top hand shows the

situation as it currently exists for the person in the

photograph. The hands grasping the top of the wall

provide contact with a support that cancels gravity. The

bottom sketch shows what happens if the hand is selected

to do the scratching, the brytes at the fingertips determine

that movement will remove support. Setting their pressure

outputs to zero before movement will lower the global

pressure sum by half, allowing the movement to be

cancelled.

smarter. The brains of the vertebrate precursor had

some distinctly vertebrate features as can be seen in

their surviving descendants, i.e. 20 or so species of a

minnow like sea creature called Amphioxus.

The rate of brain evolution has continually sped

up. Only 3% of the time from Amphioxus to humans

separates us from our common ancestor with the

chimpanzee. During these 16 million years the brain

nearly quadrupled in size adding about three times

as many neurons as had been added in the preceding

600,000,000 years. To account for this rapid

increase in size and intelligence it is reasonable to

assume that evolution has used some efficient

strategies for improving things fast.

One simple, but very powerful way evolution has

been accelerated is genome duplication. Animals are

generally diploid--they have two copies of each

chromosome. Each time a cell divides all its

chromosomes are replicated so there is a time before

cell separation when there are 4 copies of each

chromosome in a single cell. Occasionally all four

copies go to the same daughter cell and it becomes

tetraploid. If this occurs during the formation of a

gamete the whole organism becomes tetraploid.

Since no bad genes have been added, tetraploids are

often completely viable. This new copy of the

genome facilitates evolution because it can be

‘experimented’ with without damaging the basic

genetic kit of the original organism. Recent genetic

analysis has found that the genome of Amphioxus

doubled two times on the way to vertebrates.

Amphioxus has six chromosomes we have 23 plus

the runt Y chromosome.

If the genome can get more versatile by

doubling, perhaps the brain can get bigger and

smatter by a similar process. This actually happened

in going from Amphioxus to the earliest vertebrates.

Amphioxus is bilaterally symmetrical, but does not

have paired structures. It has one brain, one frontal

eye, etc. Vertebrates on the other hand have a right

and left brain, two eyes and other paired structures.

The doubling part seems straight forward, perhaps a

single developmental mutation. However, the

modifications required to make this work are not

trivial since they include a requirement that the two

new brains don't give conflicting behavioural

signals, Figure 2. Note that doubling the brain did

not require doubling the genome. It could have been

accomplished with just a few small changes.

This original doubling took place when the brain

was a tiny fraction of its current size. We know from

Amphioxus that this brain already had many of the

basic features of modern vertebrate brains. The bryte

POV suggests that our current brain is made from

NCTA2014-InternationalConferenceonNeuralComputationTheoryandApplications

128

Figure 6: The top of the figure illustrates the two

doublings of the genome in the 100,000,000 or so years

that separated pre-vertebrate chordates from vertebrates.

The doubling of the brain that occurred in the same time

interval is shown below it. Below that one step in the

hypothesised increase in be number of basic brain units or

brytes is shown.

subunits corresponding to a huge number of these

early brains. As with chromosomes, making an

exact copy of an existing structure will not lead to

disruption, but provides an extra copy of a

functioning subunit to experiment with. Note that

unlike the genome, the bryte hypothesis is that the

brain evolves by adding additional copies of a small

basic unit brain, not by doubling the whole brain—

except in the first step. Of course, the subunit brains,

i.e. brytes, get modified in the course of time in

ways to increase their computational power and

facilitate their interaction.

The brain architecture that results from bryte

duplication is not inconstant with existing brains.

Figure 2 shows that in situ multiplication of subunit

brains maps onto the brain structure if we slice the

brain ‘vertically’ so that each slice has a bit of all the

principle brain parts such as cerebral cortex,

cerebellum, basal ganglia, etc..

5 CONCLUSION

My goal here has been to show that the odd point of

view that big smart brains can be constructed from

large numbers of smaller simpler brains is

computationally useful and biologically plausible.

Even from the few things that have been presented

here -- scratching an itch, and hints at possible

biological plausibility— it can reasonably be

concluded that the bryte point of view can be

extended to account for more challenging aspects of

cognition. Consider that most of the computations

used here for scratching are also components of

other tasks. And there are many aspects of brain

function that have not been used at all—most

notably vision. There are many avenues to extend

the bryte paradigm and only time will tell if they

lead to some really new insights.

ACKNOWLEDGEMENTS

I want to thank thr Redwood Center for Theoretical

Neuroscienc at UC Berkeley for providing me with a

venue to air my ideas and for their usefull feedback.

REFERENCES

Ferraina, S., E. Brunamonti, M. A. Giusti, S. Costa, A.

Genovesio, and R. Caminiti, 2009. Reaching in Depth:

Hand Position Dominates over Binocular Eye Position

in the Rostral Superio Parietal Lobule. The Journal of

Neuroscience 29(37):11461.

Graziano, et al., 1994. Coding of visual space by premotor

neurons. Science 11: 1054-1057.

Minsky, M, 1998. The Society of Mind. Simon and

Schuster, New York.

Todorov, E ., 2006. Optimality principles in sensorimotor

control. Nature Neuroscience 7:907-915.

Torres, E. B., Zipser, D., 2002. Reaching to Grasp with a

Multi-jointed Arm (I): A Computational Model.

Journal of Neurophysiology 88: 1-13.

Jacobs, R., Jordan, M. I., Nowlan. S. J. and Hinton, G. E.,

1991. Adaptive mixtures of local experts. Neural

Computation, 3, 79-87.

Zipser, D., 2009, 2010 Brytes or How to make big brains

out of lots of small brains, parts 1, 2, 3 - Redwood

Center for Theoretical Neuroscience.

archive.org/search.php?query=david+zipser.

Zipser, D., 2012. Distributed Control of Complex Arm

Movements in Living Machines Springer-Verlag

Berlin Heidelberg, 309-320.

BrainModelingwithBrytes-MakingBigBrainsfromaLotofLittleBrains

129