Reduction of Information Redundancy in the Ascending Auditory PathwayIntroductionResultsInformation about Stimulus IdentityInformational RedundancyControls: Stimulus Bandwidth, Anatomical Location, and Frequency SelectivityDiscussionExperimental ProceduresInformation about Stimulus IdentityRedundancy QuantificationElectrophysiological RecordingsANF SimulationsSupplemental DataAcknowledgmentsReferencesNeuron 51, 359–368, August 3, 2006 ª2006 Elsevier Inc. DOI 10.1016/j.neuron.2006.06.030Reduction of Information Redundancyin the Ascending Auditory PathwayGal Chechik,1,5,*Michael J. Anderson,4Omer Bar-Yosef,2Eric D. Young,4Naftali Tishby,1,3and Israel Nelken1,21Interdisciplinary Center for Neural Computation2Department of Neurobiology3School of Computer Science and EngineeringHebrew University of JerusalemJerusalem 91904Israel4Department of Biomedical EngineeringJohns Hopkins UniversityBaltimore, Maryland 21205SummaryInformation processing by a sensory system is re-flected in the changes in stimulus representation alongits successive processing stages. We measured infor-mation content and stimulus-induced redundancy inthe neural responses to a set of natural sounds in threesuccessive stations of the auditory pathway—inferiorcolliculus (IC), auditory thalamus (MGB), and primaryauditory cortex (A1). Information about stimulus iden-tity was somewhat reduced in single A1 and MGB neu-rons relative to single IC neurons, when informationis measured using spike counts, latency, or temporalspiking patterns. However, most of this differencewas due to differences in firing rates. On the otherhand, IC neurons were substantially more redundantthan A1 and MGB neurons. IC redundancy was largelyrelated to frequency selectivity. Redundancy reductionmay be a generic organization principle of neuralsystems, allowing for easier readout of the identity ofcomplex stimuli in A1 relative to IC.IntroductionOver the last 40 years, various general principles of in-formation processing in sensory systems have beensuggested based on theoretical considerations. Theseinclude effective information transmission (Becker andHinton, 1992; Linsker, 1988), efficient use of storage(Barlow, 1961; Miller, 1956) or energy resources (Levyand Baxter, 1996, 2002), achieving sparse codes (Ol-shausen and Field, 1996), and extraction of behaviorallyrelevant stimulus properties (Escabi et al., 2003; Fritzet al., 2003; Rieke et al., 1995). Each of these proposedprinciples predicts specific transformations of stimulusrepresentations along the processing hierarchy, butthe experimental evidence required to assess any ofthem is still very limited.Among the potential changes in stimulus representa-tions, of special interest is the way groups of neurons in-teract to code information about the stimuli. These inter-actions can be synergistic, in which the interactionsincrease the amount of information carried by the groupcompared with the same neurons considered indepen-dently of each other. The interactions can also be redun-dant, in which they reduce the amount of informationcarried by isolated neurons independently because dif-ferent neurons convey overlapping information. At thereceptor level, neurons are often highly redundant sinceeach point in the sensory epithelium is represented bya large number of neurons with overlapping receptivefields. Barlow (1961) advocated the idea that redun-dancies in stimulus representation are reduced as thestimuli are successively processed at different stations.As a result, neurons at higher processing stations maybecome largely independent to allow for easier readoutand more efficient use of coding resources. This idea,together with other theoretical principles, can be inves-tigated experimentally by comparing stimulus represen-tations along a hierarchy of processing stations.To investigate how the stimulus representationchanges along processing stations, it is necessary touse stimuli that potentially engage nontrivial processingmechanisms at all levels of the auditory pathway. Thisrequirement poses opposing constraints on the stimuli:on the one hand, the stimuli have to be rich enough toactivate interesting central processing mechanisms,and on the other hand, their peripheral representationsmust be similar enough to make the task of distinguish-ing between them nontrivial. To satisfy these two re-quirements, we designed a set of stimuli that was basedon natural bird vocalizations that contain rich and com-plex acoustic structures. To these we added systemati-cally modified variants that shared similar spectro-tem-poral structures (Figure 1). These are expected to elicithigh redundancies in the auditory periphery, althoughthey are clearly different perceptually. Furthermore, wehave previously demonstrated that these stimuli evokerich and complex responses in auditory cortex(Bar-Yosef et al., 2002). These stimuli are therefore suit-able to test the fate of stimulus-induced redundancy inthe ascending auditory system.To quantify changes in stimulus representations, weused measures of information content (Borst and Theu-nissen, 1999; Rieke et al., 1997) and stimulus-inducedinformational redundancy of neural responses in threesubsequent stations in the core auditory pathway: theinferior colliculus (IC), medial geniculate body of thethalamus (MGB), and primary auditory cortex (A1).ResultsAll recordings were performed in halothane-anesthe-tized cats using a single set of stimuli consisting of nat-ural and modified bird vocalizations (Bar-Yosef et al.,2002). Figure 2 shows examples of three representativestimuli, together with the neuronal responses they eli-cited in cells from different brain areas. The A1 neurons(Figures 2F and 2G) often responded differently to thefull sound (left column) and to the main chirp componentof the sound (center column), in which the echoes and*Correspondence: [email protected] address: Computer Science Department, 353 Serra Mall,Stanford University, Stanford, California 94305.background noise were removed. Responses to the fullnatural sound and to the background noise and echoes(right column) were often similar (Figures 2F and 2G),even though the echoes were 15–20 dB weaker thanthe main chirp and had different temporal envelopes.In contrast, IC neurons (Figures 2B and 2C) respondedsimilarly to the full sound and to the main chirp, but re-sponded weakly to the noise. MGB neurons were inter-mediate (Figures 2D and 2E). In this study we quantifythese complex