Home

Organization

Mission Statement

Time Line

Research Interests

Neural Numbers

Scientific Advisory Committee

Useful Links

Neural Numbers

The human brain contains over 100 billion or 10¹¹ neurons.

Some neurons have as many as 100,000 synaptic input sites. The average in the human central nervous system is 1000 to 2000.

The volume of the human brain is approximately 1400 to 1500 cm³.

Neural events of interest happen in the 1 millisecond timeframe.

Information Storage

If the information storage of the human brain can be described by the size of synaptic junctions and the neural connectivity then the following is likely true:
The weigth of a synaptic site can likely be described by 2 bytes (16-bits) for a range of -32767 to 32767. (Negative numbers for inhibitory sites and an approximate vesicule count.)
The connectivity information would take another 12 bytes. (Near neighbors can propably be described with fewer bytes of address.)
If we assume another 4 bytes for locality information and we have a total of 20 bytes per connection.

(10¹¹ neurons) x (2,000 connections/neuron) x (20 bytes/neuron) = 4x10¹⁵ bytes

This implies that 4 pedabytes of storage are necessary for describing a human neural network for the purposes of neural emulation.

Information Processing

We will place primary emphasis on a non-spiking neural model where the stimulus is described by a 32-bit floating-point number presented every millisecond at a synaptic site (representing a very short term moving average of the pulse train.) Every millisecond an average of 2000 floating-point by fixed-point multiplies each followed by a floating accumulate would need to be performed for each neuron.

(10¹¹ neurons) x (2,000 FMADs/neuron/ms) x(1000ms/sec) = 2x10¹⁷ FMADs/sec

If a spiking model is found to be necessary then both storage requirements and the processing capability necessary probably double. However, the event driven nature of spiking models can also offer simplifications that may, if handled properly, simplify the amount of processing necessary at any given time.

Brain Slice Image capture

If each image pixel of information is 4 bytes and it contains the relevant morphological information in a 0.1um x 0.1um x 1um volume then that would amount to:

4 x 1500 cm³ x 10⁵ x 10⁵ x 10⁴ = 6x10¹⁷ pixels.

In order to identify features, stereo images may have to be taken and more than one type of device may have to be used. This may increase the amount of data that would need to be captured by another order of magnitude. Rather that require all of this storage space for a single brain database it may be better to reduce this on the fly if possible to create a connectivity net with appropriate weights as described above under information storage.

Copyright © 2003 LokTime Foundation. All Rights Reserved.