Download the full pdf it's not much to read and seriously why would you skimp out.
A Platform Technology For Brain Emulation Updated 9-05-2013
www.academia.edu
PVDM
Thanks Rfta,
This article provides great insights into the working of Akida.
The synapses also serve a memory function.
"
It has been proved that synapses are memory ii that acquire information through learning. Synapses are everywhere in the brain, Therefore memory is everywhere. Every movement, vowel, and shape etcetera has been learned and is stored as a set of variables somewhere in synaptic memory. Learned motor functions are stored in the motor cortex, while learned speech patterns are stored in Broca's and Wernicke's areas. Each module of the brain has been trained early in life and has formed connections that are appropriate to its function."
So each synapse compares a string of input spikes* with its memorized string of spikes and fires when there is sufficient correspondence rather than exact identicality, and retains the differences as a new example of the type of object, enabling the brain to automatically recognize variations of an object it had seen before.
[*Edit: or a group of synapses each compare a corresponding one of the string of input spikes with their single stored spike]
A glimpse into the near future:
"T
he nodes are organized in minicolumns and hypercolumns. Vernon Mountcastle described the columnar organization iii of the somatosensory cortex in his 1957 paper. In his 1978 paper iv he elaborated that this columnar organization is found everywhere in the cortex. A hypercolumn consists of up to 100 minicolumns. Each minicolumn consists of around 80 neurons. A hypercolumn comprises a computational unit, able to learn and to match complex temporal patterns. At least 10,000 digital nodes can be mapped onto a full-custom ASIC, representing one complete hypercolumn. "
So 8000 neurons in a hypercolumn. How many synapses per neuron?
"T
herefore it is possible to train individual hypercolumns for specific tasks, such as the recognition of sounds, syllables and words, or visual image recognition consisting of the identification of line segments. This task is then copied into a function library, where it can be used to upload the function to larger networks consisting of many hypercolumns. This gives the machine sufficient innate knowledge to proceed to learn from subsequent sensory input streams. Increasingly complex functionality is copied as "learning models‟ into the function library."
"C
onsider that intelligence is defined vi as:
a. The capacity to autonomously acquire knowledge and skills, e.g. to learn.
b. To form associations between knowledge.
c. To be aware of the self and the environment, to learn from it and to be able to interact with it.
d. The ability for autonomous adaptability to a new environment.
e. The ability to think, reason and combine knowledge to form new solutions.
f. The ability to comprehend relationships.
g. A capacity for abstract thought.
h. A capacity to create new ideas, philosophy and art. "
Just as well Peter was not around in Gallileo's day.
"A
t an average of 7000 synapses per node this represents a sustainable throughput of nearly 50 Gigabytes of data per second per device. A modest synthetic brain will consist of thousands of devices."
Cortical columns
"T
he design is highly repetitive, with each node an exact replica of every other node. It is therefore expected that the small scale design will scale quite well to a component containing at least 10,000 nodes. The connectome for the larger scale device will be the biological model of a cortical column."
Well now I've got to rethink my assumption that the 2-bit and 4-bit implementations of Akida would require correspondingly scaled MAC circuits. The NPUs (previously single-bit) have now been arranged in groups of 4 to form nodes, so is it possible that a node is capable of handling 4-bit weights and activations without a MAC circuit? After all, adding a MAC circuit to an NPU would be a major redesign. But I think I did see an article from the company which did confirm the MAC hypothesis ...
www.valeo.com
