The torrent of information that constantly bombards our senses poses a problem for our meager mortal brains, which must encode it all using the spikes of a limited set of neurons. One possible solution is to represent every item with its own personal neuron (not to be confused with the misunderstood ‘Jennifer Aniston neuron’ of 2005 ); at the other end of the spectrum, information may be encoded by global activity patterns across entire populations.

These theories offer trade-offs in their ability to faithfully represent our surroundings—dense, global coding offers highly selective information storage and is resilient to damage, while independently meaningful neurons allow quick learning and preclude interference between codes. But recently, both theory and experiment have pointed to an optimal compromise between the two—sparse coding. In sparse coding, information is represented by a relatively small number of simultaneously active neurons, with the majority of neurons inactive.

The happy middle-ground of sparse coding allows efficient learning—there is low interference between codes, and changes can be governed by local, e.g., Hebbian, rules—high memory capacity, simple transformations, and a substantial amount of information storage. Sparse coding is now known to be widespread in the nervous systems of many animals—including the rodent visual, motor, barrel, and olfactory systems, the zebra finch auditory system, and the cat lateral geniculate nucleus—but the sparseness of representations in the rodent auditory cortex has not been explicitly addressed.

Which brings us to this week’s paper—Sparse Representation of Sounds in the Unanaesthetized Auditory Cortex —by Tomáš Hromádka from Tony Zador’s lab at Cold Spring Harbor, which came out in PLoS Biology. To investigate the sparseness of auditory coding, Hromádka surveyed how the neuronal population of the primary auditory cortex responded to particular stimuli.

Unable to analyze every neuron in the auditory cortex, Hromádka optimized his technique to get as unbiased a survey as possible. First, to isolate a representative sample of cells, he used cell-attached recordings with glass electrodes—allowing him to find cells by detecting resistance, rather than poking around looking for flurries of big spikes (which tends to bias other recording techniques toward highly-active neurons).

Second, to determine the population’s response as a whole, they presented each neuron with identical sound repertoires (tones, sweeps, white-noise bursts, and natural sounds), rather than optimizing the sounds for each neuron’s responsiveness.

And third, the animals were awake. Bartiturates and ketamine, which are generally used in global anesthesia, alter neuronal response properties, leading to transient, homogeneous responses. But by using awake animals (with their heads in a stereotaxic rig and their bodies in a plastic tube), they could get the full, natural diversity of neuronal responses.

So they presented their tones, listened for the auditory code, and heard—radio silence.

Only half of the neurons tested showed any change in firing rate for any stimulus presented, and less than 5% of the entire population showed a “well-driven response” (which they defined as a change of at least 20 spikes per second). Further, both the evoked and spontaneous firing rates were low—and in both cases, most of the spikes were actually generated by small subsets of neurons.

In the evoked condition, most of the neurons ignored the stimuli, while a few neurons fired vigorously in response to certain frequencies. In the spontaneous dataset, the authors suspect that the small subset of active neurons may have been broadly-tuned interneurons (although they didn’t actually identify cell-types)—which likely contributed to the overall silence of the population.

Of course, information about how individual neurons behave is somewhat useless on its own. The interesting issue is the code that represents the sounds to higher areas and ultimately guides behavior. This code must take into account not only the highly active neurons, but also any neurons that respond at all. When the authors analyzed such population-level responses to the tones and the sweeps, they found that, at any one time, only about 10% of neurons showed any significant increase in firing rate, and even fewer showed a significant decrease.

These results suggest, for the first time (according to the authors), that sounds are represented in the auditory cortex as sparse codes, strengthening the argument that sparse coding may be the guiding principle for cortical representations of sensory input.

Discussion Questions

1. How would attention/emotion affect the sparseness or other properties of the code?
2. How might sparse coding contribute to the cocktail party effect ?
3. Sparse coding is often proposed as a means to conserve energy—we might not have enough resources for any other type of coding. If this is the primary driver of the evolution of sparse coding, it’s possible that our coding system is sub-optimal, constrained by our anatomy and energy requirements. Which—metabolic advantage or representational power—would you argue is the greater evolutionary force?
4. The more selective a code, the larger the number of neurons required. It has been suggested that invariance allows highly sparse codes to be selective. What mechanisms might underlie this invariance?