News & Analysis
Memristors mimic human brain
R Colin Johnson
1/15/2013 1:04 PM EST
HRL Labs breakthru
Today that insight has been revitalized by HRL Laboratories LLC (formerly Hughes Research Laboratories) and its Center for Neural and Emergent Systems (CNES) which has embraced the use of memristors as artificial synapses with funding from DARPA's SyNAPSE program. In conjunction with the University of Michigan, HRL recently outlined its progress toward such cognitive computers in a paper is entitled A Functional Hybrid Memristor Crossbar-Array/CMOS-System, for Data Storage and Neuromorphic Applications.
Click on image to enlarge.
Unlike the common titanium-dioxide material, HRL Labs has created a refined material stack with intrinsic rectification properties that it claims solve the "sneak path" problem that slowed early development efforts of memristive materials in crossbar arrays.
"These memristors have an intrinsic diode like behavior that prevents currents from sneaking through in the reverse bias direction and affecting the stored memory values," said program manager and principal investigator Narayan Srinivasa.
HRL's formulation storea precise analog values corresponding to the synaptic strength between brain-like neurons, then put sthe neurons themselves on a companion chip.
"Our neuromorphic architecture uses a rich programmable brain-like connectivity," said Srinivasa. "To enable this connectivity, we have decoupled the memristor array from the neuromorphic architecture, putting the crossbar arrays on one chip to store synaptic conductances which are then driven by a separate neuromorphic chip."
In its first iteration, HRL Labs two-chip solution emulates a single layer of a real brain. However, the most interesting aspects of neural learning use up to six layers. For instance, in the visual cortex each layer of processing progressively associates different orientations from which an object might be viewed. In other words, a single layer could learn to recognize a two-dimensional silhouette of an object, but would fail to identify that same object if it is turned sideways. However, after passing through all six layers that same object can be recognized regardless of from which direction it is being viewed. But to emulate all six layers of the human cortex, a 3-D memristive array with six layers will needed.
"At present we are focused on building 2-D crossbars--to minimize risk and maximize our probability of success," said Srinivasa. "But eventually we want to scale the arrays in the third dimension to emulated the true synapse geometries found in the brain.
Today that insight has been revitalized by HRL Laboratories LLC (formerly Hughes Research Laboratories) and its Center for Neural and Emergent Systems (CNES) which has embraced the use of memristors as artificial synapses with funding from DARPA's SyNAPSE program. In conjunction with the University of Michigan, HRL recently outlined its progress toward such cognitive computers in a paper is entitled A Functional Hybrid Memristor Crossbar-Array/CMOS-System, for Data Storage and Neuromorphic Applications.
Click on image to enlarge.
HRL's memristor crossbar array fabricated atop a CMOS chip
can store 10 Gbits per square centimeter.
can store 10 Gbits per square centimeter.
Source: HRL
Unlike the common titanium-dioxide material, HRL Labs has created a refined material stack with intrinsic rectification properties that it claims solve the "sneak path" problem that slowed early development efforts of memristive materials in crossbar arrays.
"These memristors have an intrinsic diode like behavior that prevents currents from sneaking through in the reverse bias direction and affecting the stored memory values," said program manager and principal investigator Narayan Srinivasa.
HRL's formulation storea precise analog values corresponding to the synaptic strength between brain-like neurons, then put sthe neurons themselves on a companion chip.
"Our neuromorphic architecture uses a rich programmable brain-like connectivity," said Srinivasa. "To enable this connectivity, we have decoupled the memristor array from the neuromorphic architecture, putting the crossbar arrays on one chip to store synaptic conductances which are then driven by a separate neuromorphic chip."
In its first iteration, HRL Labs two-chip solution emulates a single layer of a real brain. However, the most interesting aspects of neural learning use up to six layers. For instance, in the visual cortex each layer of processing progressively associates different orientations from which an object might be viewed. In other words, a single layer could learn to recognize a two-dimensional silhouette of an object, but would fail to identify that same object if it is turned sideways. However, after passing through all six layers that same object can be recognized regardless of from which direction it is being viewed. But to emulate all six layers of the human cortex, a 3-D memristive array with six layers will needed.
"At present we are focused on building 2-D crossbars--to minimize risk and maximize our probability of success," said Srinivasa. "But eventually we want to scale the arrays in the third dimension to emulated the true synapse geometries found in the brain.
|
|
|
|
|
Navigate to related information
iniewski
1/16/2013 10:47 AM EST
Memristor is just a component, like resistor, inductor or capacitor...saying it mimics the brain stretches reality too much...overall an exciting research component but we have not seen any commercial application from HP yet, still waiting for the memristor revolution!
Sign in to Reply
vandamme
1/16/2013 11:25 AM EST
I'm still waiting for the tunnel diode revolution.
Sign in to Reply
selinz
1/16/2013 5:42 PM EST
I think the brain tie in is that instead of black and white, you have shades of grey... matter...
Sign in to Reply
A Sceptic
1/17/2013 9:15 AM EST
The R & D smoking around the "memristor" stuff seems to be continuing in the New Year. That's really funny, because - up to now - no one has been able to show that Chua's hypothetical concept of nonvolatile memristor/memristive systems can be realized, i.e., to propose a reasonable physical model that would satisfy the mathematical state equations of such a system.
Sign in to Reply
Kopernikus
1/19/2013 11:53 AM EST
Memristors have been conceived much Earlier than 1971, where only the name has been changed from "Memistor" bei adding an "r" behind "Mem". Stanford-Professor Bernie Widrow has founded the "Memistor Corporation" already in the early sixties. However, the implementation technology of this early memistor could not follow Moore's law.
Sign in to Reply
Sparky_Watt
1/22/2013 2:02 PM EST
The idea of a memory that is non-volatile, as fast as DRAM (I doubt the claim it is faster, the address logic will dominate that) and smaller is great. It is nice to hear that is coming. Of course, they said the same thing for FRAM.
However, people keep hyping this idea that we can mimic the human brain. Let's be clear about this:
- We don't know how the brain works. We know how one component, the neuron works. We don't have more than a tiny clue as to how that combines to make the neural part of a brain. The impact of the thousands of neurotransmitters and non-neural cells from a processing point of view is unknown. How can we mimic something we don't understand?
- We can't control an artificial brain, any more than we can control a human brain. How would you like to deal with a psychopath that is 10 times smarter than you, and thinks 1000 times faster? Until we understand this we'd better keep our neural network experiments very small scale.
Sign in to Reply
A Sceptic
4/11/2013 8:16 AM EDT
It seems now that the fourth fundamental circuit element - referred to as the "memristor" - cannot exist in physical reality. In principle, no realistic physical model can be proposed for a solid state memory device which would operate in accordance with Chua's mathematical concept of genuine non-volatile memristors/memristive systems. When analyzed under physical aspects, the hypothetical mathematical state equations defining non-volatile memristors are by themselves in severe conflict with fundamentals of physics as discussed in "Fundamental Issues and Problems in the Realization of Memristors" by P. Meuffels and R. Soni (http://arxiv.org/abs/1207.7319) and "On the physical properties of memristive, memcapacitive, and meminductive systems" by M. Di Ventra and Y. V. Pershin (http://arxiv.org/abs/1302.7063).
Sign in to Reply