Part of Advances in Neural Information Processing Systems 10 (NIPS 1997)
Hagai Attias, Christoph Schreiner
It is known that humans can make finer discriminations between familiar sounds (e.g. syllables) than between unfamiliar ones (e.g. different noise segments). Here we show that a corresponding en(cid:173) hancement is present in early auditory processing stages. Based on previous work which demonstrated that natural sounds had robust statistical properties that could be quantified, we hypothesize that the auditory system exploits those properties to construct efficient neural codes. To test this hypothesis, we measure the informa(cid:173) tion rate carried by auditory spike trains on narrow-band stimuli whose amplitude modulation has naturalistic characteristics, and compare it to the information rate on stimuli with non-naturalistic modulation. We find that naturalistic inputs significantly enhance the rate of transmitted information, indicating that auditiory neu(cid:173) ral responses are matched to characteristics of natural auditory scenes.
1 Natural Scene Statistics and the Neural Code
A primary goal of hearing research is to understand how complex sounds that occur in natural scenes are processed by the auditory system. However, natural sounds are difficult to describe quantitatively and the complexity of auditory responses they evoke makes it hard to gain insight into their processing. Hence, most studies of auditory physiology are restricted to pure tones and noise stimuli, resulting in a limited understanding of auditory encoding. In this paper we pursue a novel approach to the study of natural sound encoding in auditory spike trains. Our
• Corresponding author. E-mail: hagai@phy.ucsf.edu. t E-mail: chris@phy.ucsf.edu .
104
H. Attias and C. E. Schreiner
~ 11111111 I II
I II I I 111111
~ I 111111 III 11111111
III II III U II
Figure 1: Left: amplitude modulation stimulus drawn from a naturalistic stimulus set, and the evoked spike train of an inferior colliculus neuron. Right: amplitude modulation from a non-naturalistic set and the evoked spike train of the same neuron.
method consists of measuring statistical characteristics of natural auditory scenes, and incorporating them into simple stimuli in a systematic manner, thus creating 'naturalistic' stimuli which enable us to study the encoding of natural sounds in a controlled fashion. The first stage of this program has been described in (Attias and Schreiner 1997); the second is reported below. Fig. 1 shows two segments of long stimuli and the corresponding spike trains of the same neuron, elicited by pure tones that were amplitude-modulated by these stimuli. While both stimuli appear to be random and to have the same mean and both spike trains have the same firing rate, one may observe that high and low amplitudes are more likely to occur in the stimulus on the left; indeed, these stimuli are drawn from two stimulus sets with different statistical properties. Our present study of auditory coding focuses on assessing the efficiency of this neural code: for a given stimulus set, how well can the animal reconstruct the input sound and discriminate between similar sound segments, based on the evoked spike train, and how those abilities are affected by changing the stimulus statistics. We quantify the discrimination capability of auditory neurons in the inferior colliculus of the cat using concepts from information theory (Bialek et al. 1991; Rieke et al. 1997).
This leads to the issue of optimal coding (Atick 1992). Theoretically, given an auditory scene with particular statistical properties, it is possible to design an en(cid:173) coding scheme that would exploit those properties, resulting in a neural code that is optimal for that scene but is consequently less efficient for other scenes. Here we investigate the hypothesis that the auditory system uses a code that is adapted to natural auditory scenes. This question is addressed by comparing the discrimination capability of auditory neurons between sound segments drawn from a naturalistic stimulus set, to the one for a non-naturalistic set.
2 Statistics of Natural Sounds
As a first step in investigating the relation between neural responses and auditory inputs, we studied and quantified temporal statistics of natural auditory scenes {At(cid:173) tias and Schreiner 1997}. It is well-known that different locations on the basal mem(cid:173) brane respond selectively to different frequency components of the incoming sound x{t) (e.g., Pickles 1988), hence the frequency v corresponds to a spatial coordinate, in analogy with retinal location in vision. We therefore analyzed a large database of sounds, including speech, music, animal vocalizations, and background sounds, using various filter banks comprising 0 -10kHz. In each frequency band v, the am(cid:173) plitude a{t) ~ 0 and phase r/>{t) ofthe band-limited signal xv(t) = a{t) cos{vt+r/>{t)) were extracted, and the amplitude probability distribution p(a) and auto-correlation
Coding of Naturalistic Stimuli by Auditory Midbrain Neurons
105