US5087826A
Expired / Lapsed
US
⏱️ 0.0 years remaining
Multi-layer neural network employing multiplexed output neurons
Assignee: Intel Corp
•
Filing: 1990-12-28
•
Granted: 1992-02-11
•
Expires: 2010-12-28
Abstract
Abstract A multi-layer electrically trainable analog neural network employing multiplexed output neurons having inputs organized into two groups, external and recurrent (i.e., feedback). Each layer of the network comprises a matrix of synapse cells which implement a matrix multiplication between an input vector and a weight matrix. In normal operation, an external input vector coupled to the first synaptic array generates a Sigmoid response at the output of a set of neurons. This output is then fed back to the next and subsequent layers of the network as a recurrent input vector. The output of second layer processing is generated by the same neurons used in first layer processing. Thus, the neural network of the present invention can handle N-layer operation by using recurrent connections and a single set of multiplexed output neurons.
Inventors
- 👤 Mark A. Holler
- 👤 Simon M. Tam
Bibliographic Details
Application Number
—
Jurisdiction
US
Data Source
google_patents
Last Synchronized
2026-11-02 01:36
CPC Classifications
G; G06; G06N; G06N3/00; G06N3/02; G06N3/04; G06N3/044; G06N3/0442
Claims Analysis
Total Claims: 13
Claim 1
Independent Claim
62 words
A multi-layer electrically trainable neural network comprising: a first synaptic layer for multiplying an input vector with a first stored weight; a second synaptic layer for multiplying a recurrent vector with a second stored weight; a plurality of neurons deriving Sigmoid output from either or both of said synaptic layers; means for multiplexing said neurons between said first and second synaptic layers.
Claim 2
Dependent Claim
62 words
A multi-layer electrically trainable neural network comprising: a first synaptic layer for multiplying an input vector with a first stored weight; a second synaptic layer for multiplying a recurrent vector with a second stored weight; a plurality of neurons deriving Sigmoid output from either or both of said synaptic layers; means for multiplexing said neurons between said first and second synaptic layers.
Claim 3
Dependent Claim
34 words
The network of claim 1 wherein each of said synaptic layers stores their respective weights as an analog transconductance value and each produces an analog output current summed along dendrites coupled to said neurons.
Claim 4
Dependent Claim
61 words
The network of claim 2 wherein said multiplexing means comprises: a first buffer means for buffering said input vector; a second buffer means for buffering said recurrent vector; control means for controlling said first and second buffers to decouple said input vector from said first synaptic layer during second layer processing, and to decouple said recurrent vector during first layer processing.
Claim 5
Dependent Claim
51 words
The network of claim 3 further comprising: a first fixed-bias weight array associated with said first synaptic layer for altering the sum of currents generated by said first synaptic layer; a second fixed-bias weight array associated with said second layer for altering the sum of currents generated by said second layer.
Claim 6
Dependent Claim
121 words
A parallel distributed processing system for correlating the match between two or more vectors comprising: an input synapse array for computing the inner product of an input vector and a first stored weight vector as a sum of currents; a feedback synapse array for computing the inner product of a recurrent vector and a second stored weight vector as a sum of currents; a plurality of neurons implementing a Sigmoid function; means for sequentially multiplexing said neurons between said input and feedback arrays such that said neurons first operate on the sum of currents generated by said input array to produce said recurrent vector, after which time said neurons then operate on the sum of currents generated by said feedback array.
Claim 7
Dependent Claim
121 words
A parallel distributed processing system for correlating the match between two or more vectors comprising: an input synapse array for computing the inner product of an input vector and a first stored weight vector as a sum of currents; a feedback synapse array for computing the inner product of a recurrent vector and a second stored weight vector as a sum of currents; a plurality of neurons implementing a Sigmoid function; means for sequentially multiplexing said neurons between said input and feedback arrays such that said neurons first operate on the sum of currents generated by said input array to produce said recurrent vector, after which time said neurons then operate on the sum of currents generated by said feedback array.
Claim 8
Dependent Claim
74 words
The system of claim 5 wherein said multiplexing means comprises: a first buffer means for buffering said input vector from said input synapse array; a second buffer means for buffering said recurrent vector from feedback synapse array; control means for controlling said first and second buffer means to decouple said recurrent vector from said feedback array during first layer processing, and to decouple said input vector from said input array during second layer processing.
Claim 9
Dependent Claim
52 words
The system of claim 6 further comprising a first fixed-bias weight array associated with said input synapse array for altering the sum of currents generated by said input array, and a second fixed-bias weight array associated with said feedback synapse array for altering the sum of currents generated by said feedback array.
Claim 10
Dependent Claim
34 words
The system of claim 7 wherein said input and feedback arrays each comprise an array of electrically programmable floating gate devices which store said first and second weights in the form of electrical charges.
Claim 11
Dependent Claim
153 words
An N-layer electrically trainable analog neural network in which inputs are organized into two groups, external and recurrent, each layer containing a synaptic array storing a weight vector and an associated fixed-bias weight, said fixed-bias weight altering the sum of currents generated along the dendrites of each array, said sum of currents representing the dot product between either of said external or recurrent inputs and said stored weight, said network also including a set of neurons for computing the overall dot product associated with each layer and a means for multiplexing use of said neurons during processing beginning with the first layer and proceeding sequentially through the N th layer, wherein processing associated with a given layer proceeds independent of the inputs applied to other layers of said network, the neural response associated with said given layer being provided as the recurrent input to the next layer to be processed in said sequence.
Claim 12
Dependent Claim
153 words
An N-layer electrically trainable analog neural network in which inputs are organized into two groups, external and recurrent, each layer containing a synaptic array storing a weight vector and an associated fixed-bias weight, said fixed-bias weight altering the sum of currents generated along the dendrites of each array, said sum of currents representing the dot product between either of said external or recurrent inputs and said stored weight, said network also including a set of neurons for computing the overall dot product associated with each layer and a means for multiplexing use of said neurons during processing beginning with the first layer and proceeding sequentially through the N th layer, wherein processing associated with a given layer proceeds independent of the inputs applied to other layers of said network, the neural response associated with said given layer being provided as the recurrent input to the next layer to be processed in said sequence.
Claim 13
Dependent Claim
41 words
The network of claim 9 wherein said multiplexing means comprises a plurality of input buffers associated with each said layer of said network, said buffers being controlled to selectively disable said external and/or recurrent inputs in accordance with said processing sequence.
References Cited (Backward Citations)
US4760346A
Cited
US4873661A
Cited
US4985703A
Cited
US4926064A
Cited
US4951239A
Cited
US4972363A
Cited
US4972157A
Cited
US5039871A
Cited
USPTO Annuity & Maintenance Windows
| Annuity Stage | Window Opens | Payment Deadline | Surcharge Period Until | Estimated Fee | Status |
|---|---|---|---|---|---|
| 3.5 Years (Year 4) | 1995-02-10 | 1995-08-12 | 1996-02-11 | $2000.00 | Window Closed / Lapsed |
| 7.5 Years (Year 8) | 1999-02-10 | 1999-08-12 | 2000-02-11 | $3760.00 | Window Closed / Lapsed |
| 11.5 Years (Year 12) | 2003-02-10 | 2003-08-12 | 2004-02-11 | $7700.00 | Window Closed / Lapsed |