Brain learning and recognition emulation circuitry and method of recognizing events
First Claim
1. An apparatus for associating an input event with a previously learned event comprising an association matrix having a sparse forward matrix characterized by a plurality of input lines randomly coupled by programmable contacts to a plurality of target lines and further comprising a plurality of programming lines, where each programmable contact is comprised of a select transistor coupled to said input line and to said target line and a floating gate MOS contact transistor coupled to said select transistor and to said programming line, and further comprising means coupled to each of said input lines, programming lines and target lines to place predetermined voltages on these lines during a learning mode to cause the charge stored on the floating gates of selected contact transistors to be altered in accordance with a forward matrix learning rule and for placing predetermined voltages on said input, target and programming lines during a recognition mode for allowing selected select and contact transistors to drain amounts of charge out of selected target lines in accordance with the amount of charge stored on the floating gates of said selected contact transistors.
3 Assignments
0 Petitions
Accused Products
Abstract
There is disclosed herein a method and apparatus for learning the characteristics of an event and for recognizing events which have been previously learned. The apparatus is comprised of forward and feedback matrices which process input signals characterizing events to be learned by comparing the convergence responses generated by the input signals at successively different levels of convergence threshold. The convergence threshold is altered until an acceptable range of convergence is reached, and then the pattern of convergence is learned. The pattern of convergence is stored for future reference. In recognition mode, the input signals are again processed at successively different convergence threshold levels until an acceptable range of convergence is reached. The pattern of convergence is then compared to the stored convergence patterns of known events to determine if the event is known. Also disclosed are methods of operating the operatus to recognize events whose input signals are cluttered by noise or which have input patterns which are incomplete. Also disclosed are a method and apparatus for learning the differences between a first event and another event with which the first event was confused during the learning mode.
225 Citations
38 Claims
- 1. An apparatus for associating an input event with a previously learned event comprising an association matrix having a sparse forward matrix characterized by a plurality of input lines randomly coupled by programmable contacts to a plurality of target lines and further comprising a plurality of programming lines, where each programmable contact is comprised of a select transistor coupled to said input line and to said target line and a floating gate MOS contact transistor coupled to said select transistor and to said programming line, and further comprising means coupled to each of said input lines, programming lines and target lines to place predetermined voltages on these lines during a learning mode to cause the charge stored on the floating gates of selected contact transistors to be altered in accordance with a forward matrix learning rule and for placing predetermined voltages on said input, target and programming lines during a recognition mode for allowing selected select and contact transistors to drain amounts of charge out of selected target lines in accordance with the amount of charge stored on the floating gates of said selected contact transistors.
- 3. An apparatus for learning of events and recognizing an input event as a previously learned event comprising an association matrix means for using a convergence rule to create an output vector and further comprising means for automatically, repetitively altering the convergence threshold used in implementing said convergence rule while attempting to recognize the same input event until an acceptable level of convergence is achieved and establishing the output vector which results at said acceptable level of convergence as the final output vector on a plurality of output lines, and wherein said association matrix means includes a plurality of target lines for which convergence individually occurs or does not occur where convergence on a target line is defined as a predetermined level of signal on said target line, and further comprising a feedback matrix means including a plurality of feedback lines and further comprising a plurality of programmable contacts between said target lines and said feedback lines and wherein said feedback lines are coupled to said output lines, said feedback matrix means for selectively programming all the contacts in said feedback matrix which couple target lines which have converged to an active feedback line where an active feedback line is defined as a feedback line coupled to an output line corresponding to a target line which has converged, where said output vector is defined as a collection of output lines coupled to a corresponding collection of target lines where some of said target lines may have converged and where some of said target lines may not have converged, and where said programming is done so as to make it easier for the same output vector to occur the next time the same input event which created the input vector occurs even if the input vector is not identical because of missing components.
- 6. An apparatus for associating an input event with a previously learned event comprising an association matrix and further comprising false positive learning means for causing learning to correctly occur in the case of false positive identifications and wherein said association matrix is comprised of a forward matrix means for learning the characteristics of an input event by unidirectional programming of the transfer characteristics of programmable transducers which randomly couple a plurality of input lines to a plurality of target lines and wherein each said transducter causes a response signal on the corresponding said target line of a programmable magnitude when the corresponding said input line is active, said programming occurring in accordance with a predetermined learning rule based upon convergence of the response magnitude above a predetermined, variable convergence threshold of a sufficient number of target lines where said sufficient number lies within a predetermined range, and wherein said learning rule implemented in said forward matrix means by varying said convergence threshold until said sufficient number of converged target lines occurs and then programming all said transducers coupled to converged target lines to create even greater response signal magnitude the next time each said corresponding input line is active, and further comprising a feedback matrix means coupled to all of said target lines by a plurality of programmable transducers coupling each of a plurality of feedback lines where each feedback line is coupled to one of said target lines at an input end and has an a plurality of output ports coupled by said feedback matrix transducters to each of said target lines and having the same structure as said programmable transducers in said forward matrix means, said feedback matrix means for sensing convergence on said target lines and for programming said feedback matrix transducers such that cross-linking between target lines occurs such that if programming of said feedback matrix transducers is based upon a first set of target lines converging said first set or some first subset of said first set of target lines can be made to converge in the future when the same input event which caused said first set of target lines to converge causes some second subset of said first set of target lines to converge.
-
9. An apparatus for recognizing an unknown event if it has been previously learned comprising:
-
means for learning one or more events by comparing the convergence on a plurality of targets caused by input signals characterizing said event which are randomly coupled to said targets by a variable threshold by varying said threshold and reading the same input signals until the amount of convergence is acceptable and then altering the characteristics of said means for learning so as to make it easier for the same convergence pattern to emerge the next time the same input signals characterizing one of the events learned appear; and means for storing the pattern of convergence on said targets when learning of an event occurs; means for receiving and storing the input signals from an unidentified event, altering the convergence threshold in successive readings of the input signals from said unidentified event until a predetermined amount of convergence occurs and then for comparing the convergence pattern which emerges when the acceptable level of convergence occurs to the collection of stored patterns and signalling when a match is found. - View Dependent Claims (10, 11)
-
-
12. An apparatus for recognizing an event comprising:
-
means for receiving a plurality of input signals and for causing convergence responses on a fraction of a plurality of target lines; means for comparing the convergence response on each target line to a variable convergence threshold and for generating a signal for each target line which converged; means for altering said convergence threshold until the number of converged target lines is in an acceptable range or stopping the learning process if an acceptable range cannot be reached and for altering said means for receiving so as to make it easier for the target lines to converge the next time the same set of input signals appear in order to learn said event when the number of converged target lines is within an acceptable range; means to store the pattern of said output signals for learned events; and means for recognizing unknown events by receiving the input signals characterizing said events and altering said convergence threshold until the number of converged target lines either does or does not reach an acceptable range ad for signalling no recognition if the acceptable range cannot be reached or no match to a stored pattern is found or for identifying the event if a match is found.
-
-
13. A recognition matrix for identifying unknown events from a class of events learned by said matrix comprising:
-
means for receiving and storing a plurality of input signals charactering an event or object which has been sensed; a forward matrix means sparsely populated by convergence response generators coupled between randomly selected intersections between a plurality of input lines and a plurality of target lines for receiving from said means receiving and storing a plurality of signals on any of said input lines and for causing convergence responses on selected target lines; convergence threshold means for establishing a variable convergence threshold at a predetermined level; comparison means coupled to receive said convergence threshold for comparing the total convergence response on each said target line to said convergence threshold and for generating an output signal for each target line which had a total convergence response which exceeds the current level of said convergence threshold; means for programming the response circuits in said forward matrix means which are coupled to active input lines and target lines whose convergence response exceeded said convergence threshold so as to alter the convergence responses said response circuits cause the next time said response circuits receive an active input signal when the number of target lines which have sums of convergence responses which exceed the current level of said convergence threshold is within an acceptable range; and means for altering said convergence threshold under predetermined conditions and for causing said means for receiving and storing to transmit the stored input signals into said forward matrix again and to cause the number of target lines which have sums of convergence responses which exceed said current level of said convergence threshold to again be calculated.
-
-
14. A method of learning the characteristics of an event comprising the steps of:
-
sensing the event using a plurality of sensors each of which senses a particular portion or characteristic of the event; receiving a set of input signals on a plurality of sensor lines from said sensors and passing them through a matrix characterized by a plurality of said sensor lines and a plurality of target lines where some or all of said sensor lines are coupled to one or more said target lines by response circuits having programmable attributes which define the magnitude of the response on the target line to which said response circuit is coupled when said response circuit receives an active input signal having a predetermined characteristic and where the response circuits are coupled between random pairs of said sensor lines and said target lines and where all said response circuits have their programmable attributes initially set to the same predetermined value, said input signals causing a random spatial distribution of responses on said target lines; convergence detecting the magnitudes of all the responses on each said target line; comparing the convergence for each said target line to a convergence threshold value for each target line; comparing the number of target lines for which the response magnitude sum exceeds said convergence threshold value to a predetermined acceptable range; altering the convergence threshold value and repeating the above steps until the number of target lines for which the response magnitude sum exceeds the convergence threshold value is within said predetermined acceptable range; recording the spatial distribution of said responses which characterize the event by altering the programmable attribute of each said response circuit which caused a response on a target line so as to increase the magnitude of response which will be caused by each said response circuit so altered the next time said response circuit receives an active input signal. - View Dependent Claims (15, 16)
-
-
17. A method of learning an event comprising the steps of:
-
(1) calculating the convergence on each of a plurality of targets upon which a plurality of input signals randomly cause convergence responses through the action of convergence response circuits coupled to said targets by comparing the total convergence response for each target to a variable convergence threshold; (2) repeating step 1 for the same input signals and for succesively different convergence thresholds until the number of converged targets is either within an acceptable range or not within an acceptable range and a predetermined limit for the convergence threshold is reached; and (3) altering the characteristics of the response circuits when an acceptable number of converged targets is found such that, the next time an input signal pattern appears for the same event, the same convergence pattern is more likely to occur even if the input signal pattern includes spurious inputs or even if part of the input signals in the input signal pattern which characterize the event are missing; and storing the convergence pattern for any event which is so learned. - View Dependent Claims (18, 19, 20)
-
-
21. A circuit for a recognition system emulating brain cortex comprising:
-
an input conductor; a column conductor adjacent to or crossing but not electrically contacting said input conductor; a capacitor coupled between said column conductor and a first reference voltage source; switching means for selectively coupling said capacitor to a second voltage source to precharge it to a predetermined value; a forward matrix contact structure comprising an electrically programmable non volatile memory cell including a select transistor having its select gate coupled to said input conductor and having its drain coupled to said column conductor and a floating gate transistor having its drain coupled to the source of said select transistor and its source selectively coupled to said second reference voltage source through the channel of a programming transistor which has a gate for receiving a signal to render the channel non conductive during programming; a Schmidt trigger having a variable trigger threshold and having an input for receiving a theta signal controlling the level of said variable trigger threshold and having a trigger input coupled to said column conductor and having an output; and a monostable multivibrator having its trigger input coupled to the output of said Schmidt trigger and having an output. - View Dependent Claims (22, 23, 24, 25, 26, 27, 28, 29, 30)
-
-
31. A circuit for emulating the function of certain areas of the brain comprising:
-
a plurality of input conductors; an output conductor; charge storage means coupled to said output conductor for storing charge; a convergence memory means associated with random pairs of said input conductors and said output conductor for storing a predetermined convergence value in response to a learn signal; coupling means coupled between the same random pairs of said input conductors and said output conductor as said convergence memory means each for draining a variable amount of charge out of said charge storage means when an input signal appears on the associated said input conductor of the pair of conductors to which said coupling means is coupled where the amount of charge drained depends upon said convergence value stored in the associated one of said convergence memory means; trigger means for triggering when the charge on said charge storage means reaches a certain level and outputting a predetermined output signal; and programming means for supplying said learn signal to only the ones of said convergence memory means which are associated with an input conductor which received a predetermined input signal and only if said trigger means has triggered.
-
-
32. An array comprising:
-
a plurality of input lines; a plurality of column conductors adjacent to or crossing said input lines but not electrically contacting them; a plurality of capacitors each having one terminal coupled to one of said column conductors and another terminal coupled to a first reference voltage source; switching means for selectively coupling each said capacitor to a second voltage source at a predetermined time; a plurality of first means each coupled between ramdomly selected input conductors and randomly selected column conductors for draining a predetermined amount of charge out of the capacitors coupled to the selected column conductors when input signals arrives on the corresponding input conductor or conductors coupled to said column conductor by one or more first means; a plurality of Schmidt triggers each having a variable trigger threshold each having a trigger input coupled to one of said column conductors and each having an output; and a plurality of monostable multivibrators each having its trigger input coupled to the output of one of said Schmidt triggers and each having a recognition output. - View Dependent Claims (33, 34, 35, 37, 38)
-
-
36. An apparatus for recognizing degraded information characterizing an event for which characterizing information has been previously learned by said apparatus comprising:
-
an input buffer for storing a plurality of data characterizing an event and for converting said data into input signals which reflect the values of said data on a plurality of output lines; a plurality of convergence target conductors; a plurality of input conductors coupled to said output lines; a forward matrix of a plurality of convergence response generating means coupled to randomly selected pairs of input conductors and target conductors for receiving incoming signals on each input conductor and causing convergence responses on one or more target conductors; convergence detecting means for each said convergence target conductor for convergence detecting the convergence responses for each said target conductor; comparison means for comparing the convergence sum for each target conductor to a convergence threshold signal at a threshold input and for generating an output signal on an output line for each said convergence target conductor which indicates whether its convergence sum exceeded said threshold that is has converged; and a feedback matrix comprised of a plurality of feedback lines each coupled to one of said output lines from said comparison means to a plurality of feedback convergence response generating means coupled to each target conductor for causing feedback convergence responses on each of the target conductors coupled by a feedback convergence response generating means to a feedback line carrying an active output signal; control means coupled to said threshold input of said comparison means for causing said input signals to be repeatedly input at different levels of said threshold for said comparison means from said input buffer and to alter said threshold from an initial value which minimizes the number of target conductors which have converged to successively different values which tend to cause successively more target conductors to converge until an predetermined range of convergence among target conductors is reached, and then for causing the pattern of converged target conductors to be compared to known stored patterns of events which have been previously learned.
-
Specification