Multimodal integrationHow do animals integrate information from multiple sensory modalities to improve their estimate about the external world? Recent works showed that insect brain contains neurons that act as a compass, which can keep track of animal's current heading. For my postdoctoral work, we have characterized a mechanosensory pathway in the fruit fly brain which conveys wind direction information detected in the antennae to these compass neurons (Okubo et al., Neuron, 2020). This work suggested that the compass system is an multimodal system that can flexibly combine self motion inputs, visual inputs, and mechanosensory inputs.
More broadly, I am interested in how to combine information from multiple datasets to increase the robustness and accuracy of statistical estimates. |
Neural mechanism of multi-task learningLike speech acquisition in humans, songbirds learn their song by listening to their tutors and imitating them. How do young birds learn to produce multiple different types of syllables? By combining quantitative behavioral analysis, single-unit recordings, and neural network modeling, we found that the protosyllables are generated by a rhythmic neural sequence, which subsequently splits into multiple different sequences corresponding to different syllables (Okubo et al., Nature, 2015). This allows efficient learning since patterns that are common to all syllables are learned first so that they don't need to be learned from scratch for each syllable, demonstrating a biological example of a multi-task learning in neural networks.
|
'Gap-free' neural circuits
How do neural circuits process sensory information and produce behavioral output? I am interested in addressing this question using neural circuits where individual neurons are identifiable, and synaptic connectivity between them are measurable. This allows us to follow the flow of information in a 'gap-free' manner, helping us understand the principles of neural circuit function.
I have co-taught a 10-day course on this topic with Nikhil Bhatla (MIT IAP, 2013). |