What kinds of algorithms would it take for a neuroscientist to - - PowerPoint PPT Presentation

▶
what kinds of algorithms would it take for a
SMART_READER_LITE
LIVE PREVIEW

What kinds of algorithms would it take for a neuroscientist to - - PowerPoint PPT Presentation

Feb 05, 2024 255 likes •516 views

What kinds of algorithms would it take for a neuroscientist to understand a microprocessor? with Eric Jonas BioRXiv: Could a neuroscientist understand a microprocessor? Error messages are useful Reverse engineer a big biological distributed

slide-1
SLIDE 1

What kinds of algorithms would it take for a neuroscientist to understand a microprocessor?

with Eric Jonas

BioRXiv: Could a neuroscientist understand a microprocessor?

slide-2
SLIDE 2

Error messages are useful

slide-3
SLIDE 3

Reverse engineer a big biological distributed algorithm

slide-4
SLIDE 4

MOS 6502

Courtesy http://visual6502.org

slide-5
SLIDE 5

How it actually works

Instruction Fetch Instruction Decoder Registers Memory Interface ALU

Control Signals Data Signals

Main Memory

slide-6
SLIDE 6

Multi scale

A B Cin S Cout

1-bit Adder

AND XOR OR

A B Y 0 0 0 0 1 0 1 0 0 1 1 1 A B Y 0 0 0 0 1 1 1 0 1 1 1 0 A B Y 0 0 0 0 1 1 1 0 1 1 1 1

logic gate primitives

Vdd Vdd A B A B Vss Out

AND gate

B A

METAL1 POLY CONTACT N DIFFUSION N-WELL P DIFFUSION

VSS- OUT VSS VDD+

AND gate (silicon)

I/V for single gate

slide-7
SLIDE 7

3 Behaviors

slide-8
SLIDE 8

Lesion studies

slide-9
SLIDE 9

How to make it work

  • Problem: Complex game instead of targeted

instructions

  • Same as for brain
  • But could work if one activated/inactivated
  • And optimized stimulation so that effects are

sparse

slide-10
SLIDE 10

“Spike data”

slide-11
SLIDE 11

Tuning curves

slide-12
SLIDE 12

How to make it work

  • Problem: not having understanding of “instructions”
  • Same as for brain
  • Run lots of programs. Relate instructions to

activities.

slide-13
SLIDE 13

Strong global correlations

slide-14
SLIDE 14

LFPs and power law spectra

slide-15
SLIDE 15

Granger causality

slide-16
SLIDE 16

How to make these work?

  • No idea!
slide-17
SLIDE 17

Whole chip

slide-18
SLIDE 18

Nonnegative matrix factorization finds something

slide-19
SLIDE 19

How to make these work?

  • Need far more different states to be meaningful
  • Far more data
  • Nonlinear dimensionality reduction
slide-20
SLIDE 20

Souped up Stochastic block model finds some network structure

slide-21
SLIDE 21

How to make it work

  • Problem: The network is far more complicated
  • Same for the brain
  • Solutions hierarchical structure inference
  • MCMC is too slow, clustering too unspecific,

needs something in between

  • Big systems
slide-22
SLIDE 22

Kasthuri and Lichtman

slide-23
SLIDE 23

with Kasthuri, Xiao, Jacobsen

0.1 cubic mm 200 microns 0.1 cubic mm 200 microns cubic mm

slide-24
SLIDE 24

Conclusion

  • We know little about how the brain works
  • Data by itself won’t solve the problem
  • Need to ask the fundamental questions
  • Countless big computational problems
slide-25
SLIDE 25