Working memory and neuronal calculus

The world offers an awesome, indescribably magnificent profusion of sensory riches. For our meager mortal brains, however, trying to process this deluge of information is akin to taking a drink from Iguaçu Falls: it's tremendously inefficient, and you will likely be violently ripped from your precipice and vanish in a ferocious torrent of natural wonder.

Because the world is too rich for our brains to process at once (or even in a lifetime), we are equipped with mechanisms that restrict the avalanche of information to a manageable trickle. At the level of the brain, this restrictive bottleneck is referred to as attention; when we attend to a certain stimulus, we select it for more comprehensive processing, while relegating the rest to a relatively superficial survey. Importantly, attention endows a capacity limitation onto our brains, not our sensory organs, which latter detect a remarkable embarrassment of sensory details. For example, the sensory neurons on the bottom of your feet are well aware of the pressure exerted by the floor, but you were probably not actively thinking about it until this sentence directed your attention to the sensation.

If our processing ended with attention, we would conduct our lives strictly from information received at the present instant, without any internal state of the mind or abstract thought. But instead of flitting whimsically in and out of our brain, information selected from the world by mechanisms of attention gain access to our working memory, which temporarily holds onto this information for detailed evaluation. For example, when ordering a pizza for delivery, you read from the menu "4-1-5, 6-9-5, 1-6-1-5," hold the sequence in your head, and punch it into your phone. In the interim between reading and dialing, the digits were stored in your working memory, and likely quickly forgotten once you heard the first ring and the number was no longer relevant. In more complex situations, the information in our working memory is the basis for decisions and planning of elaborate behavior, and is thus a critical component in many cognitive processes associated with human "intelligence," such as language.

So what is the neural manifestation of working memory? What happens in your brain between reading and dialing the pizza delivery number? Working memory is dependent on the prefrontal cortex (PFC), which is the region at the very front of the brain, directly behind the forehead. In the monkey PFC (and presumably in that of humans), there are neurons that seem to exhibit many properties of working memory; that is, they are activated by a specific stimulus, and if the stimulus will soon be relevant, they temporarily remain activated even after the stimulus disappears. For example, if a monkey must remember the location of a flash of light for a period of 4 seconds, a certain population of neurons will experience a surge in action potentials in response to the light, and proceed to fire at this elevated rate through the 4-second delay period. When the animal reports the stimulus location, the latter information is no longer relevant, and the population of neurons shuts down accordingly. Such neurons are said to exhibit persistent activity (also called "delay period" activity). Persistent neural activity is thought to represent information about a stimulus even after it is gone, thus reflecting the temporary storage of information, i.e. our working memory.

Persistent neural activity presents an interesting computational complication: action potentials are brief electrical pulses, so how does the system interpret a tonic, persisting pattern of neural activity? In a process called temporal integration (theoretically similar to mathematical integral calculations), the system accumulates information over a certain window of time and "remembers" the sum as a pattern of neural activity. That is, the network dynamics of the circuit can integrate a flurry of brief electrical pulses, and translate the sum into a persistent change in activity. One fundamental question is how the circuitry of neural integrators accomplishes this computational feat, which appears to be so fundamental to working memory.

Emre Aksay of Weill Cornell Medical College, in collaboration with David Tank at Princeton University, recently published a study in Nature Neuroscience that investigated this issue in the neural integrator that controls eye movements in goldfish (the goldfish "oculomotor integrator"). Although goldfish eye movement is not the most intuitive place to study working memory, the goldfish oculomotor integrator is a particularly tractable neural integrator, and may thus provide a framework for understanding similar mechanisms in, for example, our PFC.

Like most animals, goldfish spontaneously move their eyes around, fixating on items of interest (e.g. my finger on the glass of their tank). In order to keep the eyes in that stable, fixed position, the animal must have a sustained neural representation (i.e. a memory) of its eye position, which guides and maintains the activation of the appropriate eye muscles (even if my finger is briefly removed). The oculomotor integrator generates this internal representation by integrating the action potentials of neurons which signal changes in eye position.

When a goldfish is looking to the right, the neurons on the right side of the integrator increase their firing rates (behavior characteristic of a positive feedback system), while those on the left decrease their firing rates. Presumably, the positive feedback occurring on the right is critical for generating persistent firing, thereby enabling integration. However, the connective logic of the circuit that mediates this positive feedback is unknown.

It is known that the oculomotor integrator is a bilateral circuit, with two populations of excitatory neurons (one on each side of the brain); these populations are connected primarily by inhibitory neurons. In light of this neuroanatomy, there are two feasible mechanisms that may mediate positive feedback: a) disinhibition from (in this example) the left side of the integrator or b) excitatory connections between cells on the right side of the integrator.

By using drugs that targeted either excitatory or inhibitory neurons, Aksay and Tank sought to dissect the circuitry of the integrator and solve this dilemma. They found that the persistent activity of the integrator that underlies eye fixation did not require inhibitory neurons, but did require the excitatory connections. However, the inhibitory connections between the right and left sides of the integrator appeared to be important for coordinating the two sides, ensuring that only one has persistent activity at any one time (and thus that the eye only moves in one direction at a time).

And now for the tantalizing extrapolations to which neuroscience lends itself so wonderfully: although the persistent neural activity of discrete neural integrators, (holding a specific set of information in your working memory), does not require inhibitory pathways, the coordination between different integrator circuits (i.e. representations of different sets of information) does. The cornucopia of information presented by our surroundings may require these inhibitory connections to help liaise our working memory at a local level, lest the mayhem of our welter world prevail.