,
17 tweets,
3 min read
Read on Twitter
Next up, I'm at Tatiana Engel's talk on "Discovering dynamic computations from large scale neural activity recordings" @USBrainAlliance @CSHL
understanding brain function for a long time has been through single-neuron responses to stimuli or spontaneous activity -->build theories off of these recordings
but...trail average responses is not really what we see in individual trials (aka...reality)...so what gives here?
@anne_churchland lab data showing heterogeneity of single-neuron responses (Raposo et al., Nat Neurosci 2014) ...now technology has grown enough to allow us to examine neural circuits at large scales
Neural population dynamics can be plotted in multidimensional space, but individual neurons can still be plotted in low dimensions....
Approaches for inferring neural dynamics, need to take into account the capacity of the model and the interpretability...e.g., parametric, overparametrized...
non-parametric approach for inferring neural dynamics has both high capacity and interpretability!
...Heterogeneous firing-functions are inferred simultaneously with the population dynamics, ...model complexity analysis allows for selecting among a continuum of models
How do you build a non-parametric framework? (desired: discovering dynamics from data); diversity of single neurons responses (idiosyncratic, non linear, non monotonic; with single spike resolution...
Model discovery: searching the entire space of possible dynamics, need the 'laws of motion' and 'curl flux' in 2 dimensional space. Learn the shape of continuous functions from data
Each population state maps to the firing rate of individual neurons (non linear, arbitrary shape, learn this at the same time as the landscape of the model is discovered)
The challenge of discovering continuous dynamics from discrete spikes; evaluation of synthetic datasets (rasters of synthetic neurons); comparison to hidden markov model, gaussian process model...both of these methods fail on very fast timescales (i.e., summed data, not averaged)
With non-parametric approach; fast dynamics can be resolved...because dynamics are uncovered from individual spikes, mapped onto population state (continuous or discontinuous...?)
Inference of diverse firing functions; non parametric approach allows for accurate recovery of complex non monotonic data
Overfitting to spiking noise needs to be accounted for....seems like big problem...as you can falsely conclude that there are multiple stable states when there is only 1 or 2
Pick the model that validates dataset best....how do you decide which one is interpretable? model complexity analysis; simple way to choose best model is to compare cross validation error to model complexity (distinguishes between learning features and learning noise)
going forward we want to use it on decision making data to get ideas on how individual neuron activity contributes to activity of the entire circuit/network
