Hannah Tobias 5-14-13
Final Paper – CogSci100 Livingston
Throughout your life, you will make millions upon millions of choices –
what to wear, what type of omelet to order, whether or not to move to Los
Angeles, who to spend your life with and what to spend your life doing, etc., etc.
Each choice you make is defined by an entirely unique set of circumstances,
events, and mental states, yet each of those millions of decisions, trivial or
pivotal, can be shunted into one of four discreet categories of choice: slow and
high-impact, slow and low-impact, fast and low-impact, or fast and high-impact.
Each of these categories houses a somewhat unique arsenal of cognitive,
decision-making weaponry. Believe it or not, the choice to rescue a dog from the
middle of the road engages a slightly different set of mental functions than does
the decision to attend Vassar College, and those functions are different still from
what’s going on in your head when you order a hamburger – though at the most
basic level, all choices are in fact produced from the same basic prediction-
making, goal-modeling machinery of the brain.
As Hawkins states in his critically acclaimed On Intelligence, “[Prediction]
is the primary function of the neocortex, and the foundation of intelligence“ (2004,
pg. 89). The brain is an algorithmically savvy, computationally intricate, parallel
processing predictor. It is hardwired to receive and record inputs as computable
signals, encoding and grounding the real world into trillions of network patterns.
These patterns become recognizable with repetition and are linked to other
patterns that typically precede or follow them, which ultimately results in a mind-
blowingly complex matrix made of billions of cells able to communally predict
future events by comparing current input patterns to those it has already ‘learned’
– this matrix is our brain. We can make predictions – so what?
It is this seemingly simple capacity that allows us to do all we do: you
reach out a hand to catch a ball, predicting and interrupting its trajectory for the
desired catch. If your prediction is not accurate, you may be surprised or
annoyed as the ball slides past your hand – the real-world inputs entering at the
bottom of your brain’s computational nets don’t match the top-down prediction
signals generated and sent down by higher order, pattern-linking predictors to
meet them. This same prediction-making is described on a higher level as the
brain’s ‘modeling’ capacity by Read Montague in his Your Brain is (Almost)
Perfect. On a neuro-computational level, Hawkins’ networks predict future sets of
incoming sensory inputs, and what Montague describes as ‘modeling’ is an
interesting bigger picture parallel on the level of much more complex pattern
generation: the brain is able to predict, often based on prior experience, the
results of certain choices and creates models of the possible scenarios that might
result. “Predictions allow an organism to evaluate future events before they
actually occur, permit the selection and preparation of behavioral reactions, and
increase the likelihood of approaching or avoiding objects labeled with
motivational values” (Shultz, 1998, p.1). This ability to model possible outcomes
of choices also allows the brain to more clearly make an informed decision based
on what potential future is most suited to its goals.
For example, in choosing whether or not to cross the street with heavy
traffic flow in order to be on time for work, one might first run through the possible
outcomes: make it to the office on time, or be flattened by a 16-wheeled semi.
These models may come from past personal experience, but many don’t – if you
had been run over by a semi before, you likely wouldn’t be in any condition to
caution yourself against being run over again. So where do our models, or
predictions, of the futures that our choices may result in come from when they do
not come from personal experience? For one, they come from others’
experiences: after seeing someone or something else get run over, our brains
transfer the experience to our own arsenal of memories. We often learn even
more from other’s mistakes than we do from our own. For another, our brain is
exceedingly good at drawing parallels between scenarios (this talent is a function
of the Hawkins’ pattern recognizing networks discussed above). After seeing an
object in motion collide with another object, for example a bowling ball and a
bowling pin, our brain catalogues the results and can apply it to the real life
version: you become the bowling pin and the semi becomes the bowling ball. In
addition to these analogies, others’ experiences, and the first person past, other
predictions may come from past fictional models made in the process of making
different decisions. In the end, wherever the predictions come from, this modeling
capability ultimately allows for ‘smart’ choices to be made – choices that favor
goal-fulfillment.
When deciding whether or not to cross a dangerous intersection in order
to be on time, what are your goals and how are they prioritized? Though it may
seem clear that ‘being late’ is better than ‘being flattened’, how does the brain
itself make any distinction between the two? The answer is value. “The brain
simulates possible future scenarios, values the fictive outcomes of each
scenario, and uses the valuation to help choose a course of action” (Montague
69). It is important to note that value systems are a very “specialized kind of
cognitive mechanism, a summery mechanism” (Livingston, COGS 100 lecture,
May 13, 2013) that assigns a single number to distill a complex thing, like a
grade on a paper. “The brain starts out life richly pre-equipped with lots of value
functions” (Montague, 2007, p.54) such as the values of basic survival elements
like food and water, but the brain also learns the values of many things through
experience – learning what it should want. Because ‘life’ is innately at the top of
the brain’s value scale and is far more highly valued than the alternative, ‘being
late’, the brain chooses survival, favoring the choice that favors the outcome with
the highest value. The next time you are deciding whether to cross the road,
your brain will run a quick value-probability calculation on the results of your
potential choices and will choose the option that best supports the more highly
valued goal.
This outcome (with the highest ‘value’) is also most likely the option with
the most pleasing neurochemical result. What exactly does ‘pleasing
neurochemical result’ mean? Think about ordering food at a café: you are not
sure what to get, but you know what you like and do not like. You think about
chocolate muffins and as you do so, your mouth starts to water. You order the
muffin and enjoy its sumptuousness with satisfaction. How did the mere idea of a
muffin turn into a physical response (salivation), and how did that response affect
your decision? Montague’s reward-prediction error theory places dopamine
neurons near the heart of the answer to that question.
“[Dopamine] neurons encode reward prediction error (critic signals) as
bursts and pauses in their production of electrical impulses” (Montague, 2007,
p.106). When the neurons fire excessively, the presence of an unpredicted
reward is indicated; when the dopamine neurons fire at a regular rate, the reward
is just as predicted; and when the neurons fire at a rate lower than the normal
rate or cease firing altogether, the reward experienced is less than what was
initially expected. A dopamine surge indicating a ‘reward’ can be caused by
something as simple as the taste of chocolate, or something as complex as the
idea of having enough money to buy chocolate. Yes, an idea can cause these
neurons to fire, and it is in this way that dopamine neurons assume their more
powerful predictive, goal-orienting, decision-swaying abilities.
The idea of a reward like a muffin – which might consist of a mental
image, a ghostly gustatory sensation, or “the physical appearance of [a muffin –
is] used for predicting the much slower vegetative effects” (Shultz, 1998, p. 3). In
other words, though the ultimate ‘reward’ effect of a food like a muffin is found in
the muffin’s chemical content’s influence on our biology, our dopamine neurons
have learned to pre-emptively fire when the stimuli that consistently predict the
eventual carbohydrate reward– either the sight, the thought of, the smell, etc., –
are experienced. In fact, “about 75% of dopamine neurons show phasic
activations when animals touch a small morsel of hidden food during exploratory
movements in the absence of other phasic stimuli” (Shultz, 1998, p. 3). In this
way, when you see a muffin, your dopamine neurons may fire in expectation of
the gustatory and digestive implications of eating the muffin, and in so doing
these dopamine neurons entice you to fulfill their prediction by choosing to eat
the muffin. Analogously, ideas about other goals can act as ‘rewards’ in the
brain’s dopamine driven prediction-error system: the idea of getting a good job is
a predictor of a steady income which in turn is a predictor of a steady food source
and comfortable living conditions, two of the body’s evolutionarily programmed
goals. This ‘idea’ of a job hijacks your dopamine system and steers you toward
success, helping you make the choices that will most likely lead to that idea –
and the results of that goal’s realization. However, it is worth noting that
dopamine circuits are no substitute for thinking; after a while dopamine circuits
become acclimated to rewards (after the system learns to ‘predict’ a certain
reward through feedback loops, the dopamine neurons will no longer fire at
heightened levels when that stimuli is encountered – the neurons will adjust and
therefore assume the ‘normal reward’ firing rate), which detracts from potentially
ongoing important goals, nor can dopamine circuits generate alternative choices
to problems (that job is left to your memory and invariant representation recall).
So far we have seen that all decisions are made through prediction and
model generation, whose values (assigned in part by dopamine neurons) with
respect to certain goals are used to select the ‘best choice’. So what happens
when you make a quick, low-impact decision to bike around the left, as opposed
to the right side, of a puddle in the middle of the sidewalk? It turns out that
though we may not know a whole lot in detail about what goes on, we do know
relatively where it goes on thanks to studies on patients with Parkinson’s
disease. Patients given “dopamine washes” (injections of dopamine) restore
much of their lost capacity to function normally. Before the wash, patients have
emotionless expressions, and shaky, almost paralytic movements, but after the
injection they restore much of their mobility and personality (Montague, 2007,
p.155). This suggests that the extra dopamine has in some way temporarily
repaired their damaged valuation system: “The difference in value between two
behavioral options is ‘read out’ by dopamine fluctuations… but in PD these
fluctuations are so small that they are all about the same size” (Montague, 2007,
p.156). This merely emphasizes the subtlety of dopamine’s role. When PD
patients are given the wash, however, the increased amount of dopamine allows
them to detect smaller discrepancies between the values of different
experiences. Clearly the ability to distinguish between subtle changes in
dopamine concentration is essential to decision-making; without it, a PD patient
freezes up to save energy rather than choosing one of two equally dopamine-
valued actions. With the ability to see different values in states other than your
present one, your brain is able to move forward, taking the steps – around the
puddle in the middle of the sidewalk – to reach those more highly valued goals.
Would there an appreciable difference in the choice you made concerning
the puddle if you had more time to deliberate? Not really: the fact of the matter is,
that no matter how much time you had to decide which way to walk around the
puddle, you wouldn’t take more than a few seconds to decide before following
through on that decision. Why? Because time efficiency is also a large
component of decision-making. The many goals your brain keeps track of all
have individual values, and if wasting time deciding on ‘how to reach the other
side of the puddle’ goal keeps you from fulfilling your goal of ‘reaching the office
on time,’ your brain will recognize this and override your indecision by stimulating
action. A relatively unimportant action with virtually no time constraints will be
carried out in virtually the same way as an unimportant action with time
constraints due to the fact that time efficiency and the importance of other goals
supersedes the need to ‘get that choice right’.
Because ‘getting it right’ takes time! When faced with a heavily loaded
decision, we often take days to weigh our options, making pros and cons lists
and thinking through possible outcomes. Long-term, high-impact decisions are
also usually less easily facilitated by prior experiences than are low-impact
decisions. With less immediate feedback (unlike short term decisions where the
result is more closely linked to its cause by temporal proximity), it is likely that
each long-term, high-impact decision will require a new, innovative solution. The
making of such a decision, for example choosing one of two very different jobs,
requires all the mental power we can muster. It takes time to dredge up
memories, to examine invariant representations, or to put emotions into words
and put words into lists. The computational speed of our brain is limited, as are
the pattern searching equations we run to look for solutions amidst the billions of
stored network connections. For this reason, we must deploy all the cognitive
tools at our disposal to tackle far-reaching decisions that must be made, and
each one of these tools is fully loaded with complexities about the functioning of
the choice-making brain. One of these tools, language, plays an extremely
important and interesting role in our decision-making processes.
A research team from the University of Chicago set out to test the
boundaries of this question by asking one hundred and twenty-one American
students a question in English concerning a hypothetical epidemic. Half were
asked, “Should we develop a medicine that saves 200,000 lives, or a medicine
with a 33.3 percent chance of saving 600,000 lives?” while the other half were
asked, “Should we develop a medicine that saves 200,000 lives, or a medicine
with a 66.6 percent chance of saving no lives at all?” (Keim, 2012). Since “people
are, in a nutshell, instinctively risk-averse when considering gain and risk-taking
when faced with loss, even when the essential decision is the same,” (Keim,
2012) the results of the study were logically unsound: those who answered the
question framed in the negative with emphasis on the loss of life chose the less
secure option while those who answered the positively framed question were far
more willing to take a risk. This experiment, based on the work of Nobel Peace
Prize winner Daniel Kahneman, shows that choices made from language-based
questions are dependent on the phraseology of the problem. This indicates that
language’s power to highlight and emphasize certain sides of a scenario has an
incredibly large impact over our decision-making abilities.
The Chicago team probed a little farther into the effects of language on
choice, testing the results of the same question in the second-language of a set
of new subjects. Astonishingly, or perhaps not so astonishingly, the percentage
of people who chose the safe option stayed consistent, no matter in what light
(positive or negative with respect to loss of life) the question was phrased. This
is, again, indicative of language’s innate power over us that we escape when we
are forced to think about language as an interface between our brains and real
world information. Answering a question posed in anything other than one’s
native tongue requires a good deal of dry dissection – grammatical re-
arrangement, word translation, etc. – which, as it forces more intimate contact
with the material, no doubt helps the subject to better grasp the nature of the
question being asked. “The researchers believe a second language provides a
useful cognitive distance from automatic processes, promoting analytical thought
and reducing unthinking, emotional reaction” (Keim, 2012).
But what kind of role does language play in choices made on problems not
originally framed by language, like our dual-job-offer conundrum? In the case of a
pros and cons list, language provides structure. By taking the ‘unthinking,
emotional reactions’ and reducing them to a series of symbols, one can often
better see the issues at hand. In this way, even a first language can provide a
much-needed degree of separation from the visceral, emotional connotations or
values of a meaty decision. Just as the second language helped the subjects of
U.C.’s study see the logic behind the epidemic problem, so may a first language
provide a ‘useful cognitive distance from automatic thought, promoting analytical
thought.’ Language is a useful system of organization and value assignment;
unfortunately, we do not always have the time to use it to help us to solve our
problems.
Nor do we always have time to deploy any of our best decision-tackling
tactics. In an emergency, a fast and high-impact situation, the instinctive
reactions with which we are all evolutionarily programmed take over. Sometimes
this reaction is the quintessential fear or shock-induced paralysis that occurs
when the brain cannot process data quickly enough. “The more complex the
cognitive task, the more expansive the neural circuitry needed, and the more
likely that processing time will exceed the minimum. Non‐optimal circumstances,
such as danger, may further slow information processing” (Leach, 2004). If an
individual has never encountered anything like the emergency situation – for
example an individual in a sinking ship – before, there will be minimal patterns
from which to pull appropriate action from; “This will take at least 8‐10 s under
optimal circumstances and much longer under threat… This produces a
cognitively induced paralysis or 'freezing' behavior’ (Leach, 2004).
For all its wonders, your cognitive toolbox – ready at any point to be
whipped out to make incoming decisions with its memory recall, its language
structures, its dopamine value scales, etc., etc. – is a little odd. Is it just a toolbox
floating in mid-brain? No, of course not. You are there, aren’t you? Isn’t it your
toolbox? But what is, ‘you,’ that inescapable invariant representation that orients
your every thought? It’s ‘consciousness,’ that bizarre quality that, along with
qualia, remains the hard problem of Cognitive Science (Crick and Koch, 2003).
Many of the decisions you make each day are indeed ‘conscious,’ decisions in
which your fictive models are ones you can remember, in which your final choice
can be articulated using language, in which ‘you’ are ‘aware’ of the action you
carry out and why you do so.
Consciousness is not a necessary component of decision-making if
decision-making is defined as, ‘taking one action rather than another to reach a
goal’. Every machine and every living thing does as much. However, making
human-like decisions, conscious choices that balance on the precipice of
emotion and instinct is a different story. To make those types of decisions, an
artificial unit must be embodied, must categorize and process patterns with a
predictive software like ours. An artificial unit like that must operate on
computational, Hawkins-esque networks that learn, grow, and develop. It
depends on your definition of ‘decision-making’, but it is clear that if your
definition is any less than the above, artificial intelligence can most certainly be
made with the capacities to make those kinds of choices. Equipped with
toolboxes stuffed with prediction-action, invariant memory, embodiment, and
feedback loops, these AI’s will develop ‘consciousness’ or something very close
to what many of us understand ‘consciousness’ to be (the definition of
consciousness has yet to be concretely defined) simply as an emergent quality of
an organized information network (Smith, 2011).
Will there ever be an artificially intelligent being make slow and high-
impact, slow and low-impact, fast and low-impact, and fast and high-impact
choices? Perhaps one day far in the future after several miracle discoveries by
yet unknown names in the CogSci world. Our knowledge and therefore
technology on this front is woefully thin, but for good reason. The Cognitive
Science big bang is just beginning – who knows! In the next century we may see
some decisions being made by units that are not just out of our control, but under
their own!
Works Cited
Hawkins, Jeff. (2004). On intelligence. New York: Owl Books.
K.R. Livingston, Vassar COGS 100 lecture, May 14th, 2013.
Read Montague, Your Brain is Almost Perfect (New York: Penguin Group, 2007.
Wolfram Schultz, “Predictive Reward Signal of Dopamine Neurons,” J Neurophysiol,
1998, http://jn.physiology.org/content/80/1/1.full.pdf.
Brandon Keim, “Thinking in a Foreign Language Makes Decisions More Rational,”
Wired, April 24th 2012, http://www.wired.com/wiredscience/2012/04/language-
and-bias/.
John Leach, “Why people ‘freeze’ in an emergency: temporal and cognitive constraints
on survival responses,” Aerospace Medical Association, 2004.
Francis Crick & Christof Koch, “A framework for consciousness,” Nature Neuroscience,
2003, http://www.nature.com/neuro/journal/v6/n2/full/nn0203-119.html.
Kerri Smith, “Neurosceience vs philosophy: Taking aim at free will,” Nature, 31 August
2011, http://www.nature.com/news/2011/110831/full/477023a.html.