The second coolest things about birds (the coolest thing is that they can fly, fer goodness sake!) is that they sing. Furthermore, song birds learn their songs from the examples of their local group, just like humans learn to speak.
Learning a new song, just like learning to speak a new language, is hard, especially with only examples to work from. In fact, it would seem to be logically impossible—but everybody does it!
This month Dina Lipkind and colleagues in New York and Zurich report a neat study of how Zebra Finches learn a new song .
Their basic observation is that learning a song (or utterance) requires learning to make the sounds and learning to make them in the right order and timing. In any but the shortest song, the number of possibilities is far too large for trial an error, or even for progressive approximation of the model. It’s computationally infeasible, AKA, impossible. (The imperfection of the model, and the presence of multiple exemplars only makes things more complicated.)
Lipkind and colleagues explore a possible simple, if non-optimal, strategy: decompose the learning into pieces. Specifically, the student might learn to make all the sounds individually, and then fix up the sequence later.
The experiments use Zebra Finches, a favorite and well studied species. The investigators presented recorded and synthetic songs, and recorded the partially learned songs. Specifically, they analyzed the errors in detail. (See the paper for details.)
Their conclusion is that these birds have neural machinery for learning the pitch of tones (i.e., the pieces of the song), which learns the tone relative to neighboring tones, not the whole song. Essentially, this learns each piece, without trying to keep things in order.
A second phase learns the correct sequence, which involves learning the transitions between the pieces. The researchers conclude that this is a separate neural mechanism.
“Thus, zebra finches break down the computationally difficult task of exploring the entire space of possible motor permutations, into two simpler tasks, yielding a search for solutions that is non-optimal, but manageable.” (p. 8)
This is an interesting finding, because it parallels aspects of human learning. Furthermore, it is analogous to the most successful strategy for computing document similarity. This isn’t so much bioinspired design as bio-confirmed convergent design. Cool!
This study was non-invasive and didn’t directly measure the brain activity. So it is not known how these processes are realized in the finch brains. The study does constrain the types of neural structures that might perform as observed, though. Future research may be able to document the internal processes.
The study leaves a number of questions open. For one thing, how are the individual segments (“gestures”) learned or, for that matter, identified as units. E.g., for human babies, learning to utter phonemes is relatively simple, but only if you know how to find the beginning and end of a phoneme—which isn’t obvious.
I would note from experience that this problem is endemic in learning algorithms. I have struggled with machine learning attempts to recognize gestures from movements. It’s very hard, if not impossible, to induce where gestures begin and end, at least without supervision.
Further studies of songbirds may offer bioinspired tricks for learning gestures from examples. (E.g., hints about what is a gesture to be learned.)
The researchers speculate that this decomposition might be an evolutionary compromise between that offers efficient learning with relatively little neural resources. The intuition is that the two simpler learning processes require significant neural tissue and energy, but much less than more complicated optimization strategies. Proving this case will require a lot more information about how zebra finch brains actually work, and also what conceivable alternative architectures might be.
- Dina Lipkind, Anja T. Zai, Alexander Hanuschkin, Gary F. Marcus, Ofer Tchernichovski, and Richard H. R. Hahnloser, Songbirds work around computational complexity by learning song vocabulary independently of sequence. Nature Communications, 8 (1):1247, 2017/11/01 2017. https://doi.org/10.1038/s41467-017-01436-0