memory profiling for ironclust

goal: expand # channels and durations by tiling and see how the memory use scales per channel and per duration basi

#1. expand datasets by tiling
– cache the original (64ch, 1200s), and tile it

# 2. run the memory benchmark, parameter loop
– parameter loops: {param_set0.prm, param_set1.prm}
try to break the memory limit (256GB in my workstation)
~1MB/chan/min (3.43 MB/chan/min raw, 153GB/380ch/2hrs)

# 3. run the memory benchmark with efficient processing option

# Data storage problem
install sudo apt install bolt
restart and mount

In Ubuntu, the memory profiling precision above 20MB (MB is 10^6 whereas MiB is 2^20)
Dell precision (T7919) could have a thunderbolt 2 card installed.
(compare 1.3 GB/s vs 50 MB/s over the network)

# IronClust remaining work
1. documentation
2. benchmark scalability, memory, accuracy, channel count, density, recording duration
3. make it faster using remote slurm interface

# Ironclust result section
## accuracy
– accuracy comparison
– probe layout comparison. density. columns
– drift handling comparison

## scalabiilty
– scalability plot: memory and runtime scaling
– accuracy vs. speed plot
– speed comparison with other software
– speed scalability using multiple computing cores

## usability
-input and output format

debrief global brain 2019

loren frank talk hippocampus theta sequential firing sharp wave ripple during immobility but sequential firing during theta occurs during movement. the difficulty is lack of firing need smaller spikes thus requires spikesorting free decoding and the animal decides left or right that alternates this is the work by kenneth kay et al on biorxiv 2019. hippocampus and forebrain interaction very interesting you can trigger lfp averaring using out of field spikes. looking at the pfc and hippocampus interaction you get to see how the memory can possibly transfer out of the hippocampus. i wonder if the sequences can be selectively disrupted in real time. the flex probe can scale up. they can soon mount 512 channel chips (flip chip bonder) on the flex cable. interconnects are the hardest as expected. spike gadget collaboration with loren’s lab. very innovative solution using flash memory storage. 8 arm choice poster was interesting it scales up the two choice task based on W arm maze.

Time to think early in the morning i rarely get this moment alone since my son was born i need to wake up earlier and devote time to think. this is precious because i need to access my new neuroplasticity and plan the future.

mri q&a with Charles Epstein and Jeremy Magland over lunch

Mark Henkelman
http://www.sickkids.ca/AboutSickKids/Directory/People/H/Mark-Henkelman.html
high throughput animal mri
simultaneous fmri and electrophysiology

Coils tuned at
DTI for neuromodulator and neurotransmitter movement tracking
ratiometric measurement of sodium and potassium ions. New receiver coils?

Multimodal imaging worl. MRI, ophys and ephys agreement
https://www.jneurosci.org/content/jneuro/28/46/11796.full.pdf
https://www.tsaolab.caltech.edu/

Fine-Scale Spatial Organization of Face and Object Selectivity
in the Temporal Lobe: Do Functional Magnetic Resonance
Imaging, Optical Imaging, and Electrophysiology Agree?https://ieeexplore.ieee.org/document/4664427

http://haxbylab.dartmouth.edu/publications/HHP05.pdf
Integration of EEG/MEG with MRI and fMRI in functional neuroimaging
https://www.researchgate.net/publication/6887849_Integration_of_EEGMEG_with_MRI_and_fMRI_in_functional_neuroimaging

day 2 global brain

# bernardo sabatini.

fiber photometry and dorsal and ventral striatum behaving differently their dopamine concentrations are different in their time courses.

https://www.frontiersin.org/articles/10.3389/fnins.2019.00766/full

how is he contributing to the global brain neurotransmitter measurement?

direct and indirect pathways. from striatum.

uchida lab harvard. datta lab.

http://datta.hms.harvard.edu/lab-year/current/

# carlos brody

calcium imaging. cellular resolution

mouse encoding redundancy.

optimal deciding and encoding directions.

relative angles preserve between linear encoding and decoding vector although the global angles change over time. so there is a stable representation in the brain over time. the updating of the decoding vectors depends on learning reduced set of variables.

global brain talk monday

# misha ahrens: multimodal imaging studies.
five modality imaging
multimodal imaging. gaba voltage glu calcium
Computational correlation framework: optical imaging corelation functional
Radial astrocytes. transition between active and passive behaviral state
glibal imaging Mu cell 2019
noredrenalergic cells. fiber. medula oblongata. glial calcium transients.
Tze Koh from Byron Yu lab.

# Maneesh Sahani. UCL.
population analysis and theory

Gaussian Process Factor Analysis (GPFA) is a method for inferring latent structure that is shared across a population of neurons from single trials [2]. … Thus, inter-trial variability in neural firing that is shared across the population is modelled via the evolution of the latent processes on a given trial.May 27, 2018Temporal alignment and latent Gaussian process … – bioRxiv
h

temporal variability. time warping. additive varition. point process observation
http://papers.neurips.cc/paper/8245-temporal-alignment-and-latent-gaussian-process-factor-inference-in-population-spike-trains

neuron x time x condition: neuronXfactor, conditionXdrive, factorXdriveXtime. Factorization of regression tensor. Soldado-Magraner

DIstribution representation of neural code
A. averaging.

B. Distributed distributional code (DDC). encodes distribution.
– expected value int(p(x)x)
– probablistic computations. beliefe propagation. decidsion making (reward expectation), variational learning (latent statistics expected value)
– so neural firing rate is the coefficient of the basis function

Easy to learn, representative power, …
Helmholtz machine: unsupervised learning by bootstrapping generation and recognition
adversarial? generative and recognition network.
wake phase and sleep phase.

Infrerence in time-postdiction (wenliang & sahani 2019 NIPS) same as NeurIPS
carry belief from the past. dynamical DDC


# Session dinner
Ahrens and Ziquin. voltron imaging.
https://www.biorxiv.org/content/10.1101/436840v1

certain behaviors are phased locked to other behaviors. this is related to the behavioral hierarchy. distributed CPG. this sets the mechanism of multi-timescale behavioral hierarchy (Mark Zimmer)
impute corrupted behavior. probablistic genrative model
Scott Linderman biorxiv, behavioral decoding work.
https://web.stanford.edu/~swl1/#publications

NeuroPAL. deterministic brainbow. neuralID and connectome.

# Svoboda Mesoscale activity project (MAP)
memory guided flexible mbehavior. Jorge Jaramillo
David Lui. Susu Chen. premotor circuits in the medulla. medulla delay period firing rate increases. ALM and basal ganglia.
in-cage training setup. track localization. ALM projection to either thalamic or medulla projection types. differentiate those two. basal ganglia disinhibition. release movement. disinhibition. movement initiation

# Mante (ETH zurich)
dynamics attractor can be modelled by different combination of potential landscape + input dynamics combination. compare saddle point, line attractor, point attractor types.
compare residuals to compare likelihood of model correctness
eigenvalue, singular values. point attractor is stable. 2D dimensional projeciton. see how the system responds to the perturbation.

# Sussilo
Interpretable neural dynamics: GOLD. goal oriented learning of dynamics
Encoder RNN. stimulus to behavior to Neural. Encoder RNN.
dynamical mechanism.
C. Chandrasekeran. PMd neuropixels task.

# xiao-jing Wang. distributed persistent acivity in multi-regional brain circuits
persistent activity. can’t be stimulated directly.
mouse model whole brain. Allen institute Julie Harris
https://www.nature.com/articles/s41593-019-0417-0
keeping track of information. mechanism of short term memory.
long firing or firing chain?
sequential activation

Mom’s kidney problem

enoxaparin can damage kidney and CrCl

Over-anticoagulation can occur in patients with moderate (CrCl 30–50 mL/min) to severe renal dysfunction (CrCl <30 mL/min)

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4379338/

Thus, an LMWH with a lower molecular weight are is more dependent on renal clearance and therefore may accumulate in patients with renal dysfunction and may be more pronounced with the smaller LMWH [29].

reduce to 1mg/kg once a day if eGFR < 30mL/min . now she gets 1.5 mg twice a day. 3x of the recommended dose if her CrCl is low.

Cockcroft and Gault equation should be used for low therapeutic index and high-risk drugs. In other cases, eGFR is an adequate estimate of the CrCl.

https://academic.oup.com/ndt/article/32/12/1967/3091492

CKD patients have higher risks of bleeding

https://onlinelibrary.wiley.com/doi/full/10.1046/j.1365-2141.2003.04196.x

Differences between LMWH.

Enoxaparin has 4300 dalton and tin ape run has 5800.

Use of low doses of low‐molecular‐weight heparin (< 3500 anti‐FXa units/d) is associated with equal efficacy and better safety compared with low‐dose unfractionated heparin, whereas use of higher doses of low‐molecular‐weight heparin is associated with significantly greater bleeding (Koch et al, 2001; Mismetti et al, 2001).

action items

1. extent of her kidney damage

2. choice of her anticoagulation regimen

3. her housing situation when discharged

Sep 2 Mon Labor day @ Flatiron

Learning React. why? it ties model view controller framework. Imagine what I can do if I can expose the controls through the web. My work will be completely visible and usable by the users through the web interface. react is the fastest way to get the scientific tools to the users. JSX + react allows easy integration with the python backend. this is why I must learn react. it’s tied to Jeremy’s new visualization effort (reactopya) and it’s a very scalable way to design scientific visualization. Combined with kb-snapshot (model), spikeforest (controller), reactopya (view) it allows full integration of the flatiron pipeline. Now it’s time to build a strong well integrated system.

The ultimate goal is to implement all neural processing routines on the web and work with Jeremy to realize this dream. Break the boundary and make the data available anywhere everywhere. Let the web technology transform the way science is conducted. I bet CZI has similar vision.

Built-in Javascript functions

  • setInterval(fh, milliseconds)
  • new Date().toLocaleTimeString()

Built-in react functions

  • ReactDOM.render(element, document.getElementById(‘root’))
  • const element = (); #declare
  • React.Component doc