Analog VLSI Neural Circuits

1
Analog VLSI Neural Circuits

• CS599 computational architectures in biological
  vision

2
Charge-Coupled Devices

• Uniform array of sensors
• Very little on-board processing
• Very inexpensive

3
CMOS devices

• More onboard processing
• Even cheaper!
• Example ICM532B from www.ic-media.com
  single-chip solution includes photoreceptor
  array, various gain control and color adjustment
  mechanisms, image compression and USB interface.
  Just add a lens and provide power!

4
The challenge

• Digital processing is power hungry
• Analog processing is much more energy efficient
• But so much variability in the gain of
  transistors obtained when fabricating highly
  integrated (VLSI) chips that analog computations
  seem impossible
• nearly each analog amplifier on the chip should
  be associated with control pins, analog memories,
  etc to correct for fabrication variability.
• Hopeless situation?

5
A VLSI MOS transistor
6
An analog chip layout the wish
7
An actual chip the cold reality
8
Biological motivation

• Well, there is also a lot of variability in size
  and shape of neurons from a same class
• But the brain still manages to produce somewhat
  accurate computations
• Whats the trick? online adaptability to
  counteract morphological and electrical
  mismatches among elementary components.

9
Remember? Electron Micrograph of a Real Neuron
10
Mahowald Meads Silicon Retina

• Smoothing network allows system to adapt to
  various light levels.

11
Andreou and Boahen's silicon retina

• See http//www.iee.et.tu-dresden.de/iee/eb/
  analog/papers/mirror/visionchips/vision_chips/
• andreou_retina.html

12
Diffusive network

• dQn/dt is the current supplied by the network to
  node n, and D is the diffusion constant of the
  network, which depends on the transistor
  parameters, and the voltage Vc.

13
Full network

• Two layers of the diffusive network upper
  corresponds to horizontal cells in retina and
  lower to cones. Horizontal N-channel transistors
  model chemical synapses.
• The function of the network can be approximated
  by the biharmonic equation
• where g and h are proportional to the
  diffusivity of the upper and lower smoothing
  layers, respectively.

14
Full network
15
VLSI sensor with retinal organization
16
Carver Mead the floating gate
www.cs.washington.edu/homes/hsud/fg_workshop.html
17
Spatial layout
18
Electron tunneling
19
Electron tunneling
20
Hot electron injection
21
Hot electron injection
22
Spatial layout
23
A learning synapse circuit