The way canaries remember the last few seconds of their singing lets them to produce much more complex, language-like songs, researchers say.
When a canary sings, it maintains a memory trace of the notes produced in the previous five to 10 seconds, a process that allows the bird to produce songs with long-range rules or structure similar to sentences, according to the new study.
In the project, researchers used tiny, head-mounted microscopes to track the activity of the output neurons that reside in a canary’s high vocal center, a brain area involved in song motor control.
“They create a complex syntax with long-range rules resembling properties of human behaviors like speech, dance, and playing a musical instrument.”
In prior studies, researchers had identified the activity of these neurons in simpler singers, revealing one of the most precise patterns of neural activity observed in any organism.
When researchers applied the same method to the more complex song of canaries, they saw neurons activating in specific sequential contexts, with the rules of activation spanning up to 40 syllables over four seconds.
The research opens a window on theorized “hidden states” of the brain, a form of short-term memory that integrates past information with ongoing motor control, says Tim Gardner, an associate professor and the chair in neuro-engineering at the University of Oregon’s Phil and Penny Knight Campus for Accelerating Scientific Impact.
Studying short-term motor memory in canaries provides an opportunity to examine a high-level motor phenomenon in a controlled model system, one that’s a bit like how studies of the hydrogen atom helped crack the code of quantum mechanics at its inception, Gardner says.
“You want to examine a new phenomenon using the simplest possible model that captures the essence of the problem,” he says.
“We often think of songbirds in a similar way. Birdsong is a very quantifiable behavior. Sensory motor learning is 50% or more of what brains are all about. It’s learning to integrate sensation and action to effectively control movements, in this case, vocalizations.”
Songbirds are known to form detailed sensory memories for their tutor songs, and to use the memories to guide the development of their own song to match the tutor over many months. However, until the new study there was no evidence for short-term memory of song that could form a substrate for more complex song rules.
“Speech algorithms used in Siri and Google Assistant networks use these types of hidden states seen in the canaries.”
“These birds produce songs that contain hundreds of syllables organized in a way that indicates that they are using the short-term memory of preceding song syllables to guide the choice of the next elements in song,” says lead author Yarden Cohen, a neurosurgery research fellow at Massachusetts General Hospital, which is affiliated with the Harvard Medical School.
“They create a complex syntax with long-range rules resembling properties of human behaviors like speech, dance, and playing a musical instrument,” Cohen says. “We discovered that their song circuitry reflects the working memory required for their complex syntax.”
The research, Gardner says, delivers a new way to study the principles of short-term memory.
“If you reflect on the nature of speech, the choice of what to say next is guided by working memory that integrates over many timescales, from the overall aim of the communication episode to the local rules required for proper grammatical form,” Gardner says. “Canary song is much simpler, but it follows long-range syntax rules such as ‘sing syllable D’ only if five seconds ago I sang A rather than B.”
This deep structure, he says, contains simple similarities to speech where the ending of a sentence is dependent on how the sentence began. In both systems, correlations between past and future parts of the vocalization require a form of short-term memory.
“What is clear is that a lot of cellular rules that underlie learning and memory are highly conserved,” Gardner says. “For example, there are cells in the basal ganglia in songbirds that have incredibly similar patterns of activity to what has been seen in rodents. While brain architecture may differ, the fundamental computations expressed at a cellular level are the same.”
Gardner will continue to use the tools from the study. Ideally, he says, it could lead to not just to improved understanding of complex behaviors but also to enhanced machine-learning methods.
“A lot of what we see in the canary resembles computational models that have been used for speech recognition and general artificial intelligence algorithms,” he says. “Speech algorithms used in Siri and Google Assistant networks use these types of hidden states seen in the canaries.”
Eventually, Cohen says, studying the neural basis of canary song production may make it possible to understand how working memory mechanisms adapt to new conditions or fail when brain circuits are damaged. Developing such a model, he adds, may point to new therapies for speech and comprehension deficits that come with aging and in neurodegenerative diseases such as Parkinson’s and Alzheimer’s.
The research appears in Nature. Additional researchers from Boston University contributed to the work. The National Institutes of Health supported the researchers.
Source: University of Oregon
