Chandan Singh | dl for neuro

dl for neuro

view markdown

Ideas for neuroscience using deep learning

list of comparisons:

Modern deep learning evokes many parallels with the human brain. Here, we explore how these two concepts are related and how deep learning can help understand neural systems using big data.


Brief history

The history of deep learning is intimately linked with neuroscience, with the modern idea of convolutional neural networks dates back to the necognitron.

pro big-data

Artificial neural networks can compute in several different ways. There is some evidence in the visual system that neurons in higher layers of visual areas can, to some extent, be predicted linearly by higher layers of deep networks. However, this certainly isn’t true in general.

  • when comparing energy-efficiency, must normalize network performance by energy / number of computations / parameters

anti big-data

  • could neuroscientist understand microprocessor
  • no canonical microcircuit

Data types

  EEG ECoG Local Field potential (LFP) -> microelectrode array single-unit calcium imaging fMRI
scale high high low tiny low high
spatial res very low low mid-low x low mid-low
temporal res mid-high high high super high high very low
invasiveness non yes (under skull) very very non non
  • ovw of advancements in neuroengineering
  • cellular
    • extracellular microeelectrodes
    • intracellular microelectrode
    • neuropixels
  • optical
    • calcium imaging / fluorescence imaging
    • whole-brain light sheet imaging
    • voltage-sensitive dyes / voltage imaging
    • adaptive optics
    • fNRIS - like fMRI but cheaper, allows more immobility, slightly worse spatial res
    • oct - noninvasive - can look at retina (maybe find biomarkers of alzheimer’s)
    • fiber photometry - optical fiber implanted delivers excitation light
  • alteration
    • optogenetic stimulation
    • tms
      • genetically-targeted tms:
    • local microstimulation with invasive electrodes
  • high-level
    • EEG/ECoG
    • MEG
    • fMRI/PET
      • molecular fmri (bartelle)
    • MRS
    • event-related optical signal = near-infrared spectroscopy
  • implantable
    • neural dust

general projects

  • could a neuroscientist understand a deep neural network? - use neural tracing to build up wiring diagram / function
  • prediction-driven dimensionality reduction
  • deep heuristic for model-building
  • joint prediction of different input/output relationships
  • joint prediction of neurons from other areas


  • non-human primate optogenetics datasets
  • vision dsets
    • MRNet: knee MRI diagnosis
  • datalad lots of stuff
  • springer 10k calcium imaging recording:

    • springer 2: 10k neurons with 2800 images

    • stringer et al. data

    • 10000 neurons from visual cortex

  • neuropixels probes
  • allen institute calcium imaging
    • An experiment is the unique combination of one mouse, one imaging depth (e.g. 175 um from surface of cortex), and one visual area (e.g. “Anterolateral visual area” or “VISal”)
  • predicting running, facial cues
    • dimensionality reduction
      • enforcing bottleneck in the deep model
      • how else to do dim reduction?
  • responses to 2800 images
  • overview:
  • keeping up to date:
  • lots of good data:
  • connectome

    • fly brain:
  • models
    • senseLab:
      • modelDB - has NEURON code
    • model databases:
    • comp neuro databases:
  • raw misc data
    • crcns data:
      • visual cortex data (gallant)
      • hippocampus spike trains
    • allen brain atlas:
      • includes calcium-imaging dataset:
    • wikipedia page:
  • human fMRI datasets:
  • Kay et al 2008 has data on responses to images
  • calcium imaging for spike sorting:

    • spikes: