Data processing: artificial intelligence – Neural network – Structure
Reexamination Certificate
2000-09-22
2004-07-13
Starks, Jr., Wilbert L. (Department: 2121)
Data processing: artificial intelligence
Neural network
Structure
C706S033000, C706S040000
Reexamination Certificate
active
06763340
ABSTRACT:
STATEMENT OF GOVERNMENT INTEREST
The invention described herein may be manufactured and used by or for the Government for governmental purposes without the payment of any royalty thereon.
BACKGROUND OF THE INVENTION
This invention relates to an artificial neural network device (ANN) and, in particular, to an ANN embodied in a microelectromechanical system (MEMS). In general, ANNs may be thought of as data processing devices that extract advanced knowledge from complex data. In particular, ANNs can be designed with training methods that use a number of known test cases to effectively teach the ANN to perform a task, for example, to extract features from images, recognize printed text (optical character recognition or OCR) and phonemes/morphemes (speech recognition). They are also used to solve optimization problems and to perform complex, non-linear control functions.
The prior-art ANNs, whether hardware or software, have limitations. Hardware ANNs can be either digital or analog circuits. The basic ANN ‘node’ element or device produces a non-linear function of a sum of weighted inputs, for example, a sigmoid function of the sum of several products, where each product is an ‘input’ multiplied by a ‘weight’. ANN node elements that perform the base function are often used in configurations consisting of highly interconnected layers of elements.
In a digital hardware ANN network, the inputs and weights are binary numbers stored in registers. Ways of using the network can be micro-coded as operations, such as transfer, multiplication, summing, and other computations, on the binary numbers that represent the inputs and weights. Their output is a binary number stored in a register.
In an analog hardware ANN, the inputs and weights take the form of voltages, currents, or charge packets that vary in magnitude. Special-purpose analog or digital circuits that produce an output voltage perform the multiplications and other computations. The prior art also includes software ANNs implemented as a simulation on a general-purpose digital computer.
One disadvantage of all the prior-art ANNs described above is that they require silicon transistors, either in the digital or analog circuits of hardware ANNs or in the general-purpose digital computer of software ANNs. Embodiments using transistors have restricted operating temperature, typically within a range of 0-70 or −55-+125 degrees Celsius. As the power dissipation of silicon circuitry increases with complexity, supplying power to and removing the heat from very large chips containing millions of transistors is troublesome. Thus routing power supply conductors to such a system and packaging and cooling it are difficult and costly, thereby limiting the system's complexity.
Another disadvantage of prior-art ANNs is that their silicon transistors incorporate insulating dielectric layers, such as silicon dioxide, that are subject to degradation. For example, bulk- and surface-charge generation and trapping may occur in the dielectric layers, especially in devices placed in outer space, where they are subject to continuous ionizing radiation. This degradation can lead to shifts in transistor threshold voltages, and in turn to problems in timing. Eventually, the circuits fail. Likewise, charges generated in these layers can become trapped, causing surface inversion and stray leakage that in turn increase power dissipation and cause catastrophic latch-up, particularly in analog circuits.
Still another limitation has been reported for analog hardware ANNs that represent weights with charge packets of variable size stored internally on a small oxide capacitor. At high temperature the amount of charge may decay slightly with dwell time, necessitating a complex “chip in the loop, bake re-training” procedure that may have to be repeated several times to ensure the ANN's stability. Such a device would be difficult to use in a system that requires built-in training modes to enable it to learn from experience.
OBJECTS AND SUMMARY OF THE INVENTION
An object of the present invention is to provide ANNs at lower cost with simpler design and fabrication, greater temperature range and radiation tolerance, and lower operating and standby power requirements.
A further object of the present invention is to provide ANNs based on MEMS.
Still a further object of the present invention is to provide MEMS ANNs made from micro-machined polysilicon as the operative computational element in the MEMS.
Yet a further object of the present invention is to provide MEMS ANNs whose micro-machined polysilicon or composite beam structure is the operative computational element.
Briefly stated, a novel microelectromechanical system artificial neural network (MEMS ANN) device performs the function of a conventional artificial neural network node element. Micro-machined polysilicon or high aspect ratio composite beam micro-resonators replace as computational elements the silicon transistors and software simulations of prior-art ANNs. The basic MEMSANN device forms a non-linear (e.g., sigmoid) function of a sum of products. Products of the magnitudes of sine waves, applied to the input drive comb and shuttle magnitudes, are formed in the frequency domain and summed by coupling a plurality of resonators with a mechanical coupling frame, or by integrating them into one resonator. A sigmoid function is applied to the sum of products by shaping the overlap capacitance of the output comb fingers of the resonator. Methods of building and using various single MEMS ANN devices and multi-layered arrays of MEMS ANN circuits are also described. These novel MEMS ANNs exhibit an attractive combination of performance characteristics, compared to conventional hardware ANNs that use silicon transistors or simulations of ANNs running in software on digital computers, including lower cost, simpler design, wider temperature range, greater radiation tolerance, and lower operating and standby power. These advantages favorably impact system weight and size because of reduced shielding, cooling, and power requirements.
According to an embodiment of the invention, [1
ST
INDEPENDENT CLAIM]
According to a feature of the invention, [2
ND
INDEPENDENT CLAIM]
REFERENCES:
patent: 3947828 (1976-03-01), Meijer
patent: 5172204 (1992-12-01), Hartstein
patent: 5204937 (1993-04-01), Minnaja
patent: 5315162 (1994-05-01), McHardy et al.
patent: 6507828 (2003-01-01), Leonard et al.
Makohliso, S.A. et al, A Blomimetic Materials Approach Towards the Development of a Neural Cell?Based Biosensor, Engineering in Medicine and Biology Society , 1996. Bridging Disciplines for Biomedicine. Proceedings of the 18th Annual International Conferen.*
Triffet, T. et al, Transient potentials transmitted by an artificial electrochemical synapse, System Sciences, 1991. Proceedings of the Twenty?Fourth Annual Hawaii International Conference on, Vol: i, 8?11 Jan. 1991, p.:283?289 vol. 1.*
Kurzweil, Raymond, The emergence of true machine intelligence in the twenty?first century (abstract), Proceedings of the 1993 ACM conference on Computer science, Mar. 1993, p. 507.
Burns Daniel J.
Pronobis Mark T.
Williamson David C.
Burstyn Harold L.
Mancini Joseph A.
Starks, Jr. Wilbert L.
The United States of America as represented by the Secretary of
LandOfFree
Microelectromechanical system artificial neural network device does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Microelectromechanical system artificial neural network device, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Microelectromechanical system artificial neural network device will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3234104