Juvenile zebra finches learn the underlying structural regularities of their fathers' song

Abstract

Natural behaviors, such as foraging, tool use, social interaction, birdsong, and language, exhibit branching sequential structure. Such structure should be learnable if it can be inferred from the statistics of early experience. We report that juvenile zebra finches learn such sequential structure in song. Song learning in finches has been extensively studied, and it is generally believed that young males acquire song by imitating tutors (Zann, 1996). Variability in the order of elements in an individual's mature song occurs, but the degree to which variation in a zebra finch's song follows statistical regularities has not been quantified, as it has typically been dismissed as production error (Sturdy et al., 1999). Allowing for the possibility that such variation in song is non-random and learnable, we applied a novel analytical approach, based on graph-structured finite-state grammars, to each individual's full corpus of renditions of songs. This method does not assume syllable-level correspondence between individuals. We find that song variation can be described by probabilistic finite-state graph grammars that are individually distinct, and that the graphs of juveniles are more similar to those of their fathers than to those of other adult males. This grammatical learning is a new parallel between birdsong and language. Our method can be applied across species and contexts to analyze complex variable learned behaviors, as distinct as foraging, tool use, and language.

Keywords: birdsong; song structure; statistical learning; vocal development; zebra finch.

FIGURE 1

Three illustrations of simple song corpora, representing the song of three birds (a, b, c), and their description as a SYL type grammar in the form of a graph and in the form of an adjacency matrix. The BEGIN and END symbols are added by the grammar inference procedure. Each element e_i,j in the matrix represents the weight of the edge that links syllable i to syllable j in the grammar.

FIGURE 2

Corpus, lexicon, and grammar of a single individual (#423). Clockwise from left top: the entire corpus (each line represents a song bout); the lexicon (syllables + motifs) for the COL type grammar without introductory notes (the BEGIN and END symbols are added by the grammar inference procedure); and the COL type grammar inferred by our model.

FIGURE 3

The mean probability assigned to a song produced by a bird, after training on the rest of that bird’s corpus. The result presented is a mean, per grammar, of means calculated per bird. The letters above the bars signify groups: bars marked with the same letter do not significantly differ according to the Tukey HSD test, with p < 0.0001. Grammar COL+i (COL, allowing introductory notes in units) assigns a significantly higher mean probability to the withheld test songs than all other grammars, and is accordingly the grammar on which the subsequent analysis focused.

FIGURE 4

Similarity among pairs of song grammars. The distribution of grammar similarity values for the 31 pairs of related individuals (SAME family) and for the 320 pairs of unrelated ones (DIFFerent families), for the COL grammar type (based on syllables + motifs) and the Spectral grammar similarity measure (based on the eigenvalue spectrum of the matrix), showing medians, first and third quartiles (box), limits of 1.5 times the inter-quartile range (whiskers) and outliers (+ symbols), where higher values indicate greater distance between grammars and thus lower similarity. The median of the SAME distribution is significantly lower than that of the DIFF distribution (p < 0.003, Kruskal–Wallis rank sum test), indicating greater similarity in songs of related individuals. Of the 31 pairwise similarity values for SAME birds, 21 were significantly lower than the median similarity value for DIFF birds (Wilcoxon sign rank test with Bonferroni-corrected alpha of 0.0016). A binomial test showed this pattern to be significant (p < 0.036).

FIGURE 5

Examples of grammars of three individuals: these graphs describe the finite-state grammar of type COL without introductory notes, derived for two unrelated juveniles and the father of one of them. From top to bottom: unrelated juvenile, son, father. Transitions with probability <0.1 were omitted for clarity. To avoid confusion, the syllables in each bird’s song were assigned unique characters. Thus, syllables in the repertoire of individual 625 are denoted by digits, syllables in the repertoire of individual 423 by uppercase letters, and syllables in the repertoire of individual 303 by lowercase letters. Importantly, our analysis does not rely on syllable-level correspondence among individuals. The spectral similarity measure, according to which grammars of fathers and sons are more similar to each other than grammars of unrelated individuals, is too complex and spatially distributed to be visually apparent in a casual inspection of the graphs.