Distributed representations are the inevitable consequence of devoting large neuronal circuits to a detailed and adaptive analysis of complex information. The neocortical sheet with its extensive cortico-cortical connectivity is characterized by ubiquitous massive divergence and convergence, sparseness and reciprocity of the vast majority of connections. It therefore appears as an optimized dynamical structure for detailed and adaptive analysis on the one hand and the operation of multiple parallel neuronal processes required for optimizing speed and accuracy of information processing on the other hand. The established concepts of information coding in the cortex are based on tuning functions of many individual neurons thought to express their stimulus specificity in an independent way. A second level of organization is usually attributed to the spatial relations of neurons in topographically organized representations like those in sensory areas and the convergence of neuronal signals carrying information from different modalities into “higher” areas which are more involved in executive functions or the formation of complex memory representations. It has been argued that the collection of neuronal signals from consecutive recording sessions can be used to reconstruct population codes as it has been done with the “population vector” analysis in motor, sensory, and memory areas of the cortex. It is clear that the success of this method relies on fixed neuronal response properties which have been consolidated in cortical circuits over long periods of time such that the spatial pattern and the mixture of neurons contributing to the population response is reasonably stable.

Introduction

Sensory information progresses centrally from the primary sensors in the periphery to the central neural structures that derive relevant environmental information from these sensory data and determine appropriate physiological and behavioral responses. In this chapter, I present a general theory of early olfactory sensory processing in the primary olfactory epithelium and olfactory bulb (OB). The theory depicts olfactory sensory processing as a cascade of representations, each of which exhibits characteristic physical properties and is sampled by appropriate neural mechanisms in order to construct the subsequent representation. The primary olfactory representation is mediated by the activation pattern across the population of primary olfactory sensory neurons (OSNs) in the sensory epithelium. The secondary olfactory representation is similarly mediated by the activation pattern across the population of principal neurons immediately postsynaptic to the OSNs, known as mitral cells. (Mitral cell axons diverge dramatically, projecting to roughly ten different central structures within the brain; the resulting tertiary and subsequent olfactory representations are constructed outside the olfactory bulb and are not discussed at length herein.) The transformation between the primary and secondary representations is a robust, intricate, two-stage process that corrects for artefacts that can hinder the recognition of odor qualities, regulates stimulus selectivity, and transduces the underlying mechanics from a robust but costly rate-coding scheme on a slow respiratory (theta-band) timescale to a sparse dynamical representation operating on the beta- and gamma-band timescales and suitable for integration with other central neural processes.

Information representation in neuronal populations: what is the “machine language” of the brain?

Research in the area of neuroscience and brain functions has made extraordinary progress in the last 50 years, in particular with the advent of novel methods that enables us to look at the properties of neuroanatomy and neurophysiology in much finer detail, and even at the activity of living brains during the performance of tasks. However, the question of how information is actually represented and encoded by neurons is still one of the “final frontiers” of neuroscience, and surprisingly little progress has been made here. How information is encoded in the brain has captivated medics, scientists, and philosophers for centuries. Scholars such as Leonardo da Vinci or René Descartes had already an astonishingly detailed knowledge of the anatomy of the brain, and had made suggestions that it is the brain that processes information and even harbors the seat of the personality or of the soul. However, whenever suggestions are brought forward how information might be processed and represented in the brain, these often turn out to be simplistic and idealistic. These rarely add up to more than a kind of “homunculus” that somehow receives information that is received via the eyes or the ears. This model only transfers the problem of information representation from the brain to the homunculus.

One problem with the research of information encoding is that it is completely counter-intuitive.

Introduction

Auditory cortex function beyond bottom–up feature detection

Until the 1980s the auditory cortex was mainly conceptualized as the neuronal structure implementing the top hierarchy level of bottom–up processing of physical characteristics (features) of auditory stimuli. In that respect, plastic changes in anatomical and functional principles were only considered relevant for developmental processes towards an otherwise stable adult brain. Presently, this view has been replaced by a conceptualization of auditory cortex as a structure holding a strategic position in the interaction between bottom–up and top–down processing (for review see Irvine, 2007; Scheich et al., 2007), in particular auditory learning (for review see Weinberger, 2004; Irvine and Wright, 2005; Ohl and Scheich, 2005).

In this chapter we review experimental evidence from gerbil and macaque auditory cortex that has led to this change of view about auditory cortex function. It will be argued that a fundamental understanding of the role of auditory cortex in learning has required to move beyond the study of simple classical conditioning and feature detection learning, for which auditory cortex does not seem to be a generally necessary structure (see below). Specifically, it will be elaborated that the abstraction from trained particular stimuli, as it is epitomized in the phenomenon of category learning (concept formation), is a complex but fundamental learning phenomenon for which auditory cortex is a relevant structure harboring the necessary functional organization.

Decoding simultaneously recorded spike trains

Pioneering studies of motor cortex by Georgopoulos and colleagues (e.g. Georgopoulos et al., 1982) established that “population vectors,” constructed from weighted averages of the responses of single neurons, can accurately predict behavioral variables, such as movement direction. This approach has been used to study population coding in a number of cortical systems and has led to the view that cortical neurons act as independent processors of information (e.g. Gochin et al., 1994). However, some recent work has challenged this interpretation of neural population activity. For example, Schneidman et al. (2003) proposed interpreting neural ensemble activity by comparing ensemble information with information represented by the single neurons that comprise the ensemble. In a synergistic coding scheme, ensembles encode more than the sum of the component neurons. The advantage of synergy is that there can be a massive gain in information from the activity of multiple neurons. In a redundant coding scheme, the removal of individual neurons has little effect on encoding and thus the ensembles can be less noisy and less prone to errors. In Narayanan et al. (2005), we adapted the information-theoretical framework proposed by Schneidman et al. (2003) to measures of decoding of the performance of a delayed response task with activity from the rodent motor cortex. The predictive relationship between neural firing rates and a categorical measure of behavior, e.g. correct vs. error performance of a reaction time task, was quantified using statistical classifiers.