BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Europe/Stockholm
X-LIC-LOCATION:Europe/Stockholm
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20210916T132450Z
LOCATION:Henry Dunant
DTSTART;TZID=Europe/Stockholm:20210707T110000
DTEND;TZID=Europe/Stockholm:20210707T113000
UID:submissions.pasc-conference.org_PASC21_sess143_msa266@linklings.com
SUMMARY:Approximating Single Neurons with Deep Convolutional Neural Networ
 ks
DESCRIPTION:Minisymposium\n\nApproximating Single Neurons with Deep Convol
 utional Neural Networks\n\nBeniaguev\n\nUtilizing recent advances in machi
 ne learning, we introduce a systematic approach to characterize neurons' i
 nput/output (I/O) mapping complexity. Deep neural networks (DNNs) were tra
 ined to faithfully replicate the I/O function of various biophysical 
 models of cortical neurons at the millisecond (spiking) resolution. A Temp
 orally-Convolutional-DNN with 5-8 layers was required to capture the I/O m
 apping of a realistic model of layer 5 cortical pyramidal cell (L5PC), whi
 ch includes nonlinear voltage-dependent dendritic currents, and was activa
 ted by AMPA, GABAa, and NMDA synapses. This DNN generalized well when pres
 ented with inputs widely outside the training distribution. We find t
 hat the main contribution to the model's complexity stems from the NMDA sy
 napses - when NMDA receptors were removed, a much simpler network (Fully-C
 onnected-NN with one hidden layer) was sufficient to fit the model. Analys
 is of the DNNs' weight matrices revealed that synaptic integration in
  dendritic branches could be conceptualized as pattern-matching from a set
  of spatiotemporal templates. This study provides a unified characterizati
 on of the computational complexity of single neurons and suggests that cor
 tical networks have a unique architecture, potentially supporting their co
 mputational power.\n\nDomain: CS and Math, Chemistry and Materials, Physic
 s, Life Sciences, Engineering
END:VEVENT
END:VCALENDAR