Ravi Mill

The effects of task similarity during representation learning in brains and neural networks Nature Communications - Here, the authors show learning tasks with similar structures can initially cause interference and slow down learning, but both the brain and artificial networks gradually...

Our new paper, now published in @natcomms.nature.com , asks a simple question: when two tasks share a common structure, does the brain learn them more efficiently? Surprisingly, this was not the case. Thread below (1/7)
rdcu.be/eSwvU

Dynamically shifting from compositional to conjunctive brain representations supports cognitive task learning - Nature Communications Learning shifts multi-task representations from compositional to conjunctive formats. Cortical conjunctions uniquely associate with effects of practice, and index switch costs. Subcortex is critical f...

10) Please check out the full paper here: “Dynamically shifting from compositional to conjunctive brain representations supports cognitive task learning”, doi.org/10.1038/s414...

9)Our results suggest the shift from compositional-conjunctive representations serves as a general neurocomputational signature of learning. So early experience of learning the cello is reliant on retrieval of compositional reps, whilst later practice encodes more specialized cortical conjunctions.

Matrix of cross-network interactions, with violin plot summary of distal (off-diagonal) interactions

8) Despite the observed importance of cortical conjunctions, subcortical regions (hippocampus and cerebellum) have a critical role in orchestrating the cross-network interactions that bind compositional representations into conjunctions.

Results graphs depicting association between cortical conjunction strength and increased pattern separation, and practice-induced resistance to task switching costs

7) The strength of conjunctions specifically in cortex index the extent/robustness of practice, associating with behavioral improvement, increased pattern separation and resistance to task switching costs.

Results plot showing method of estimating compositional and conjunctive representations, block-averaged strength estimates for each type, and timecourse trajectories for each over practice

6)We find that novel tasks rely on task-general compositional representations (allowing transfer of abstract knowledge), whereas practiced tasks are reliant on task-specific conjunctions (allowing for specialization/expertise). Conjunctions originate in subcortex before slowly spreading to cortex.

Design schematic for multi-task fMRI paradigm

5) We tackled this question here, using functional MRI to scan healthy young adults as they performed a battery of complex tasks, from first novel presentation through hours of practice.

4) Although the effects of practice in improving behavior are well-established, what happens in the brain to allow us to perform tasks when they are novel through repeated practice remains mysterious.

Schematic showing necessity/utility of practice for cello skill development

3)However, practice is required to move from mere competency towards proficiency/expertise. Whilst transfer of prior knowledge may allow you to play “Mary had a little lamb” on the cello within days-weeks, it will require months-years of practice to play more complex pieces, like Bach’s cello suites

Schematic showing compositional sensory and motor transfer when the task of playing the cello is new

2) We all perform a multitude of cognitive tasks every day. Many tasks are novel, requiring rapid transfer of prior knowledge to achieve basic competency. E.g., when learning to play the cello, you may benefit from prior knowledge of sheet music or prior experience with another bowed instrument...

Schematic depicting cortical-subcortical interactions during multi-task learning

Excited to see our paper with @mwcole.bsky.social finally out in peer-reviewed form @natcomms.nature.com! We examine how the human brain learns new tasks and optimizes representations over practice…1/n

Lab’s latest is out in Imaging Neuroscience, led by Kirsten Peterson: “Regularized partial correlation provides reliable functional connectivity estimates while correcting for widespread confounding”, where we demonstrate a major improvement to standard fMRI functional connectivity (correlation) 1/n

Never a bad moment to remind everyone, if you have a union, join it and get involved. If you don't, found one. No one is coming to save us, the law won't protect us, and you can't shame the shameless. There is organizing and solidarity.

This Mayor seems willing to sell out all of us to save himself.

New Yorkers won’t allow it.