🪴 Jacky Zhao's Quartz 4.0

Search

SearchSearch

Neuroscience Papers - Part 2

Dec 04, 20245 min read

Meta

Generalization in tasks

https://www.nature.com/articles/s41586-024-08145-x

Neural networks

https://arxiv.org/pdf/2405.07987 - Platonic representations - all neural networks have same representation representation: Here representation is considered as the common nearest neighbours paraidigm. Two models are said to have same/similar representations if they have large number of common neighbours for a particular data point

how do u see common neighbours?

  • see the embeddenings of a datapoint(call P) u want to check in both models(output of penultimate layer).
  • see embeddings of all other data points, find “k” nearest datapoints to the embedding of datapoint P
  • see number of common among k nearest neighbours interesting:
  • supervised networks like alexnet and unsupervised nets like CLIP, both have same represntations
  • But it obviously depends on metric of represntations. Stricter metrics give less common representations

To think:

  • Is seeing common neighbours the way to check representation between 2 models?

  • In neuroscience, we take it implicity that two different mice have same representation. Hence we club the neurons.

  • https://www.pnas.org/doi/full/10.1073/pnas.1403112111 - Visual models matching human performance and neural responses

Autism

https://www.nature.com/articles/s41593-024-01800-6#Sec2 ASD has high dynamic range mapping from inputs to outputs. In figure 5, they model ASD and typical data with same Leaky compeititve accumulator model, but change the mapping(high dynamic range in ASD) from stim params to drift between control and ASD

Decision making

Superior Colliculus in Decision making

  • Genetically Defined Functional Modules for Spatial Orienting in the Mouse Superior Colliculus

  • A neural mechanism for terminating decisions - https://www.sciencedirect.com/science/article/pii/S0896627323004002?via%3Dihub#fig5 Showing Sup. Colliculus is responsible for threshold, Intrapareterial region is for decision variable

  • Cortico–basal ganglia circuit mechanism for a decision threshold in reaction time tasks : A comp model for calculating thresholds.

  • CONCURRENT, DISTRIBUTED CONTROL OF SACCADE INITIATION IN THE FRONTAL EYE FIELD AND SUPERIOR COLLICULUS

Fractals in brain

https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=fractals+in+brain+networks&btnG=

Others

Rewiring retinal projections to Medial Geniculate Nucleus in ferrets, now visual inputs go from retina to MGN to Auditory cortex. Retinotopic map observed in A1. This rewiring is achieved by ablate LGN so retina can’t go to LGN. And nearby to LGN is MGN. But MGN has inputs from Sup. Colliclus and Branchium of Inferior Colliculus. Ablate those regions to so that retina projections don’t have competition. - https://chatgpt.com/share/67505755-83dc-8002-969b-193d8e30a6a4

Bayesian Inference:

Max Neural Likelihood(MNLE) - https://elifesciences.org/articles/77220#content

  • Finding Likelihood function using Neural networks

Seq. Neural Posterior Estimation(SNPE) - https://elifesciences.org/articles/56261

  • Finding Posterior directly from Neural Network

Drift Diffusion Models in 2 Alternate choice task:

Race between Proactive and reactive process- https://www.nature.com/articles/s41467-021-27302-8#Sec11

”Proactive responses are generated when the AI threshold is reached first; the choice is then defined as a direct read-out of the sign of the EA process 𝑥(𝑡) after the interrupted stimulus is integrated”

  • Model parameters were estimated for each animal only based on RTs for all trials including FBs

  • Slowing of responses due to fatigue or satisfied: account for the slowing of the responses within each session (Fig. 6a, b), AI drift scaled linearly with trial index k as V_A,k_ = _ν_A0 + _ν_trial·k, and we fitted the parameters _ν_A0 and _ν_trial

  • Bias towards left or right: For biased trials, the value of the starting point was signed depending on the trial-dependent expectation of rewarded side b__k = ±1 , so that _Z_E = _z_E·b__k, where we fitted the parameter _z_E representing the magnitude of the animal’s expectation

  • Not used choice data ??? “MLE values were obtained using RTs data only, and not choices”

  • Extended DDM: evidence accumulations before the onset of stim, directly after fixation

  • Wiener Process: https://galton.uchicago.edu/~lalley/Courses/313/BrownianMotionCurrent.pdf

Weber’s law and DDM - https://www.nature.com/articles/s41593-019-0439-7#Equ1

DDM review by Roger Ratcliff - https://www.cell.com/trends/cognitive-sciences/abstract/S1364-6613(16)00025-5?_returnURL=https%3A%2F%2Flinkinghub.elsevier.com%2Fretrieve%2Fpii%2FS1364661316000255%3Fshowall%3Dtrue

Original DDM paper - Roger Ratcliff: A Theory of Memory Retrieval

DDM in value based tasks instead of sensory tasks - https://elifesciences.org/reviewed-preprints/96997

-Nested Sampling

Stats

Tools

Graph View

Backlinks