We attach considerable value to those higher brain functions that allow us to contemplate the world rather than respond reflexively to it. We interpret sensory information, decide upon its meaning and significance, and respond in ways that reflect our biases, memories, and goals. My laboratory studies the neural mechanisms responsible for these mental processes that intervene between sensation and behavior.
To do this, we measure—and sometimes manipulate—the activity of one or a few neurons in the brain during the formation of a decision about a visual stimulus (e.g., What color is it? What direction is it moving? How long was it present? Is it useful?). The detailed analysis of brain function at the cellular level is made possible by combining microelectrode recording and behavioral experiments in monkeys that have been trained to perform complex tasks. Advances in systems neuroscience over the past 40 years allow us to target neurons in the cerebral cortex of the monkey that are thought to play a role in these tasks. Our experiments involve no pain or discomfort. The monkeys seem content to forage for reward in the challenging, unthreatening environment we create for them.
A Simple Perceptual Decision: Direction Discrimination
We study the neural computations that underlie decision making. In general, the decisions we make are based on information received through the senses. We study a simple version of this phenomenon in the lab. We train monkeys to view a video display of randomly positioned, moving dots and to decide whether there is a tendency for some of the dots to move in a particular direction, say up or down. We make this task difficult by allowing only a small fraction of the dots to move and then only briefly. In fact, most of the dots that the monkey sees are just shown briefly at random locations and are not moving at all. After many months of training, the monkey can usually make the correct decision, even when only a small fraction of the dots are moving. This level of performance is comparable to that of a practiced human observer. This task was deployed by William Newsome (HHMI, Stanford University), Anthony Movshon (New York University), and their colleagues to determine which neurons in the brain represent visual motion, and how well they represent it. These experiments led to the idea that the electrical activity from a small group of neurons in the visual cortex—about 100 that are selective for either upward or downward motion—comprises the signal that the monkey uses to make its decision. This idea is the foundation for our experiments.
We have discovered that neurons in the association cortex of the parietal and frontal lobes receive the motion signals from the visual cortex and convert this information to a categorical decision: Was the motion up or down? Consider what this conversion might entail. As the random dots appear and disappear, neurons in the visual cortex increase and decrease their level of activity. When the net motion is upward, there is a weak tendency at any moment for the group of "up" selective neurons to respond more than the group of "down" selective neurons. It is only a weak tendency, but over tenths of seconds, the accumulated activity (i.e., the total number of action potentials) will tend to be greater from one group of up neurons. We say "tend to" because it may turn out that even with time, the down neural signal exceeds the up signal, which explains why the monkey makes errors occasionally. If this idea is correct, the monkey's decisions should reflect the accumulation of information as a function of time, and there should be neurons in the brain that represent this accumulation. Our experiments test these ideas.
Accumulation of the Evidence
By varying the amount of time that we allow the monkey (or human) to view the visual stimulus, we can ascertain how information is accumulated toward a decision. As time passes, the monkey's accuracy improves in a manner that requires perfect accumulation of the neural signals that have been recorded in the visual cortex. Over the course of about a second, the monkey's brain works like an integrator, storing and accumulating all the information that it can. This implies that the neural circuits responsible for decision making compute something like a time integral of the activity that they "read" from the visual cortex. Indeed, the neural activity that we measure in the parietal and frontal lobes seems to have exactly this character. As time passes, the activity ramps up gradually as mounting sensory evidence supports a particular decision, or it declines gradually as evidence mounts for the alternative decision. When the monkey makes an error, the neurons respond as if evidence accumulated to support the wrong direction.
How Long Does It Take to Decide?
If decisions arise from the accumulation of evidence, then it should take longer to reach a decision when the evidence is weaker. Moreover, if the brain behaves like an integrator of the evidence, then there should be a predictable relationship between stimulus strength (affecting ease or difficulty), the amount of time taken to reach a decision, and the accuracy of the decision. This interplay is evident in the activity of neurons that perform this integration. In recent experiments, we have shown that this mechanism applies to more complicated decisions among several options and in circumstances that require a form of "reasoning" from pieces of evidence. The neurons add and subtract quantities that lead to an estimate of probability or "degree of belief." (Sample neural recordings from these experiments can be downloaded from www.shadlen.org/mike/movies/ProbClass/Movies.html.)
What Are the Computations?
Our experiments suggest that neurons in the association cortex help us make decisions by accumulating sensory evidence for and against a particular proposition. We suspect that propositions are represented in the brain in a way that is similar to planning an action—an idea with roots in philosophies that link meaning with intention. Our experiments suggest that the neural circuits that are capable of accumulating evidence are the ones that can sustain activity across time. Viewed in this way, decision making and the interpretation of sensory information have something in common with motor planning and short-term memory. All involve neurons that can sustain a level of activity in response to fleeting information that has come and gone tenths of seconds to seconds in the past. From a computational perspective, the persistent neural activity represents the output of an integrator (or accumulator): it responds to an impulse (of information) with a sustained step of activity.
I believe that understanding the neural time integrator will be among the premier topics to unite systems, cellular, and computational neuroscience over the next decade. What mechanism is responsible for the sustained activity of neurons in the association cortex? Do these neurons possess different types of electrical properties than neurons in the primary sensory and motor areas of the cortex, or is the sustained activity the result of interactions between neurons in different layers of the association cortex? What controls the level of activity? What resets the integration process so that it can begin again?
We are gaining insight into these questions as we attempt to understand how factors such as bias, reward expectation, time pressure, and attention affect decisions and the brain cells that underlie them.