Songbirds

We use songbirds to investigate neural mechanisms of motor learning, performance evaluation, courtship and parenting

The spiking of a single dopamine neuron during singing (Gadagkar et al, Science 2016).

Songbirds and humans have similar brain circuits

Basal ganglia in birds and mammals

The basal ganglia thalamocortical circuit shown at left is an evolutionarily conserved brain pathway required for trial and error learning in humans and song learning in birds. This pathway is associated with movement disorders such as Parkinson’s, Huntington’s and dystonia, and psychiatric illnesses such as schizophrenia and addiction. In past work we discovered that the behavior-locked activity patterns of BG cell types is similar in singing birds and behaving primates.

The 'trial' part of learning: baby birds babble - and get better with practice

Trial and error song learning

Adult zebra finches sing a stereotyped song with a fixed sequence of syllables, e.g. 'a-b-c.' Juvenile finches hear their tutor song, and begin a month-long process to try to imitate it. Like babbling infants juveniles don't know how to move their vocal muscles to make the right sounds, so they 'babble.' Birds practice - singing thousands of songs per day - and gradually learn to imitate their tutors.

The 'error' part of learning: Birds get a spritz of dopamine when a they hit the right note

We discovered that when a singing bird unexpectedly hears itself sing the right note, its dopamine neurons are activated in the same way as when a thirsty monkey unexpectedly receives juice. And following song mistakes, its DA neurons are suppressed as when a primate experiences disappointing reward omission.

An unanswered question was how performance is evaluated as ‘good’ or ‘bad’ during practice. It was well established that dopamine activity contributes to reinforcement learning by encoding reward prediction error in tasks where animals learn for primary rewards such as food or juice. But it was unknown if dopamine could also encode performance error in tasks that did not involve reward. To test if dopamine encodes performance error, we recorded dopamine neurons in singing birds while controlling perceived error with distorted auditory feedback. Remarkably, dopamine activity encoded performance just like reward: phasic bursts following better-than-predicted outcomes, and phasic pauses following worse-than-predicted ones. We next identified origins of dopaminergic error signals in a ventral pallidal region outside the classic song system, revealing a previously unidentified ‘actor/critic’ circuit motif inside the songbird brain. This discovery was high impact because it demonstrated that circuit motif associated with drug addiction and foraging in mammals is ancestral and can be repurposed for learning a motor sequence like birdsong. Finally, dopamine is, of course, not the only neuromodulator likely to be important for a complex behavior like birdsong learning. The basal forebrain also has strong cholinergic projections to motor cortical areas, yet the role of acetylcholine for motor learning remained unclear. We found that manipulation of cholinergic signaling in vocal motor cortex of juvenile birds did not affect vocal babbling, yet chronic blockade over weeks impaired learning, resulting in an impoverished song with excess variability, abnormal acoustic features and reduced similarity to tutor song.

We discovered brain signals suggesting that songbirds know exactly when they are coming into a tricky part of their song

We discovered an actor-critic circuit motif, similar to one used in common deep reinforcement learning networks, inside the songbird basal ganglia. This motif computes dopaminergic performance error signal by comparing the actual (heard) to the predicted quality of individual song syllables.

An animal foraging a familiar environment in search of food will have a memory of where it got rewards, resulting in a place-dependent reward prediction. Now imagine a bird practicing a song with many syllables. Will it similarly have a memory of when in the song it made mistakes, resulting in a syllable-dependent error prediction? We discovered that the songbird ventral pallidum (VP) signals the predicted quality of syllables during singing. We view VP’s role in computing the predicted quality of syllables as conceptually similar to its long-established role in computing the predicted reward value of cues or places. And as in mammals, VP sends these prediction signals to VTA dopamine neurons, enabling them to compute prediction error by comparing the actual reward (or song error) received to the predicted reward (or error).

Males turn off their dopaminergic self-evaluation signals when they perform for females

We discovered that dopamine signals are gated by social context. When a male courts a female, its DA neurons no longer respond to his errors. They are instead activated by her vocal calls.

It was long assumed that a main function of dopamine in the songbird was to facilitate the reduction in variability that occurs when a lone male bird transitions to female-directed song performance. We recorded from dopamine neurons during female-directed singing and made three surprising discoveries. First, tonic discharge of DA neurons did not change with the appearance of the female, countering decades of untested assumptions about activation of dopamine neurons by mating contexts. Second, DA performance error signals that were present when males sang alone mostly disappeared when males sang to females, demonstrating that DA error signals can be dynamically gated. Third, this gating suggested that during courtship the male attended less to evaluating his own song and more to the behavior of the female. Consistent with this idea, we found that DA neurons were activated by female vocalizations, providing the first demonstration that DA can by activated by the vocalization of a conspecific. In ongoing work, we are testing how DA signal gating is implemented in upstream circuits.