Hi, I'm Aaron, and I like ducks.

This talk is an overview of Donegan's (1995) proposed development of phonological representation and a computational account thereof.

Donegan, P. (1995). The innateness of phonemic perception. In Proceedings of the 24th Western Conference on Linguistics (Vol. 7, pp. 59-69).

Adults perceive speech through the phonemic categories present in their native languages.

Infants are sensitive to phonemic distinctions present in all languages, not just the languages of their environment.

It's really unusual for an infant to outperform an adult. Adults have greater experience, greater cognitive, motor, and physical development, etc. Why are the squishy-brains beating us?

This loss of perception isn't present in music perception or anywhere else in auditory perception...

Decline in sensitivity limited to language (e.g., Werker and Lalonde, 1988; if ISI reduce to 500ms or less, Werker & Tees, 1984; Werker and Logan, 1985)

What makes language special?

Language is special because we're buried in incredibly salient input with tons of regularity. That's not the surprise.

The surprise is how early this happens - before the first year for consonants and around six months for vowels (and some results argue for earlier onsets).

But kids also seem to develop phonologically accurate representations early and quickly too. They don't accept adults mimicry of their inaccurate production, don't confuse lexical items which they produce identically (e.g., a child produces [bi] for baby and bird but does seek feathery creatures when someone uses baby in a sentence), many alternations in child speech seem best accounted by adult-like underlying representations, and new articulatory distinctions are not over-applied.

Let's get down to brass tacks - what do we mean by phonemic perception? Phonemes are not tied to lexical distribution, but are meaningful sounds within a language. We limit allophones after Bazell (1954) to articulatorily motivated alternations clustered around phonemes.

Stampe (1987) claims that hearers won't cluster allophones around a phoneme unless there's an articulatory motivation for the alternation. He also pointed out that there are reasons to target the phoneme as well as an allophonic variation - optimal clarity vs. optimal ease.

Optimizing clarity and optimizing articulatory ease interact to produce the surface forms we encounter, pulling the form closer to the phonemic representation or further from it in turn.

Phoneme is a sound that can be perceived and produced as itself (not as a variant of something else).

But where do phonemes come from?

Babies quickly begin to explore the capabilities of their bodies and their environments. This includes reaching behaviors (learning depth of field, gross motor control, proprioception, etc) as well as prelinguistic behaviors of vocalization and babbling.

Through auditory and proprioceptive feedback children make connections between their articulations and the resulting sounds (Locke & Pearson, 1992; Menn, 1992), creating a map between types of gestures and acoustic effects (i.e., phonetic features).

This connection between articulatory gesture and sound is essential for imitation.

Fortitions occur when the speaker optimizes for clarity - the child learns which gestures give you the best sound match for a phonemic category they've learned.

Lenitions wear away at the optimally clear gestures by making concessions to ease and timing. You must fortify before lenition can erode away.

This vocal exploration allows the child to discover the boundaries of what sounds they can make and the shortcuts their language allows them to take.

What are the implications of these boundaries? All speakers share the articulatory bounds which create lenitions but some languages require you to find gates through these walls to carry your message through. The phoneme distribution of a language describes how often these gates are needed as well as where they're found.

VD - Vowel Denasalization, a fortition
VN - Vowel Nasalization, a lenition

English allows both processes, giving us a single category with two realizations.

Hawaiian allows VD but not VN, giving us a single category with no allophonic variation.

Hindi does allows VN but not VD, giving us two categories, one of which varies.

French speakers must overcome both VD and VN, giving us two categories, neither of which varies.

To summarize: Children explore their abilities through vocalization and babbling while simultaneously learning the categories of the ambient languages. All children discover the limits of the articulatory apparatus (and the resulting natural lenitions), but some of these limits must be overcome to produce clear speech in a given language. The limits of articulation tend to constrain the size of segment inventories while the needs of the language can push to expand them.

I work with a computational model of how people learn and reason called DORA - Development Of Relations by Analogy (Doumas, Hummel, & Sandhofer, 2008). This model is a symbolic-connectionist model - it implements symbols through distributed connectionist representations, supporting the strengths of both symbolism and connectionism without resorting to a hybrid model design.

An enhancement of Hummel & Holyoak's (1997) model of analogical reasoning LISA (Learning and Inference with Schemas and Analogies), incorporating a developmental account for the representations both models depend on.

Another model in the same design lineage is BART (Bayesian Analogy with Relational Transformations; Lu, Chen, & Holyoak, 2012). Within our academic family we name our models after cartoon characters.

Why DORA? It's neurally plausible with strong correlates to key recent neuroimaging findings that fell out of computational constraints of the model (e.g., Knowlton, et al., 2012).

DORA accounts for 35+ findings surrounding the development of relational reasoning and 50+ findings from the adult analogy-making literature.

DORA is built on a small set of assumptions (three or four, depending on how you count). Let's focus on the two ones that allow the model to capture this theory: sensitivity to time and comparison-based learning.

We know that humans are sensitive to time. DORA exploits this sensitivity in two ways - firing things in sequence (to build structure) and firing things simultaneously (to discover and represent symbols).

This is built on simple Hebbian learning; what fires together wires together. This mechanism is exploited to discover the overlaps between things by firing the distributed representations for both simultaneously and learning the differential activation strengths across both. The resulting overlap is fired as the representation of the commonalities. This process supports concept discovery and refinement through the same mechanism!

This intersection extraction engine provides an account for how phonemes might be learned from input, and the resulting representations provide a natural account for categorical perception based on development of knowledge rather than maturational processes.

These are the sensory and motor homunculi. They depict how humans would look if body part size was proportionate to the areas of the brain devoted to the perception and motor control, respectively. Linguists will find it unsurprising that the motor homunculus has a rather large tongue.

Motor programs that correspond to a particular gesture / reach live in the motor cortex (posterior frontal cortex) and are fired to produce the same gesture again.

Back to those babbling babies - they fire the motor program for an articulatory gesture and associate that gesture with the resulting sounds, building connections via simple Hebbian learning.

One of the advantages of this comparison-based learning mechanism operating over distributed representations is that distinctions that don't carry information (e.g., vowel nasalization in English) fall out of the representation over time. This allows language learners (like babies!) to discover what phonemes map to which allophones, informed by their own babbly explorations of the soundscape.

The child's growing knowledge of the sounds within their language help them know when it's okay to get close to the target (e.g., allophonic variation)

And when you need The Real Thing. Sometimes it's hard to find The Real Thing (you haven't figured out where it's sold yet, or how to make the right gestures) but after more time exploring how that tongue works we figure out where they sell Mexicoke (Mercado de la Raza and Fujioka Wine Times).