Dongyan Lin / 林冬妍

To what extent do the temporal representations in the brain contribute to behaviour? Here, we trained deep reinforcement learning agents on timing and time-dependent tasks, and showed the emergence of brain-like time representations in networks trained to calculate time implicitly. With virtual lesion experiments, we showed that the temporal representations largely contribute to behaviour through recurrent dynamics, demonstrating dissociation between temporal tuning curves and timing behaviour.

(Lin et al., 2023 Sci Rep / Lin & Richards, 2021 bioRxiv)

In this work, we set out to answer the questions "what do neurons in the visual cortex like to see? Do neurons in different visual areas like to see different images?" with deep learning models. We show that mouse higher visual areas indeed prefer different visual stimuli, but their tuning landscape is more complex than single-neuron preference!

(Preprint / SfN 2022)

Since the original paper from Barack and Krakauer (2021) came out, the field of cognitive computational neuroscience has been debating if the brain is better understood from a single-neuron ("Sherringtonian") or population geometry ("Hopfieldian") approach, or if they could be merged, or if they should remain distinct but complementary perspectives. We organized this GAC workshop to bring together researchers whose work provides different perspectives on the two paradigms, and the possibility of their reconciliation.

(CCN 2023 workshop recording / Proposal)

I'm grateful to have contributed to the work through my internship at RIKEN. We developed a sequencing protocol for short capped RNAs, applied it to the human cell line THP-1, and compared it with the landscape of long capped RNAs, discovering distinct transcription initiation preferences and associations with disease SNPs. This deep sequencing elucidates the diverse array of transcripts from known and novel promoters and enhancers, revealing new noncoding RNA families.

(de Hoon et al., 2022 Genome Research)