41 CLAIMS

• 1. A probabilistic learning element that sequentially receives objects and outputs sequences of recognized states, said learning element comprising:
• means for sequentially receiving objects;
• means for storing received object information, including,
• said received objects, and
• sequences of received objects;
• means for storing items of previously learned information, said items including,
• sequences of states,
• states contained in said sequences of states,
• objects contained in said states contained in said sequences of states,
• sequences of objects contained in said states contained in said sequences of states,
• positional information for each object contained in said states contained in said sequences of states, and
• predetermined types of knowledge relating to said previously learned information, whereby received object information, relating to received objects, is stored as well as previously learned information;
• means for correlating said received object information with said previously learned information for assigning conditional probabilities to possible sequencies of recognized states;
• means, responsive to said conditional probabilities of possible sequences of recognized states, for determining a most likely sequence of recognized states;
• means, responsive to said previously learned information, for detecting that a state has ended and for providing an end of state signal; and
• means, responsive to said end-of-state signal, for outputting said most likely sequence of recognized states as a recognized state sequence.
• 2. A probabilistic learning element as described in claim 1, wherein said positional information stored for each object includes the object's distance to begin and distance to end of a state.
• 3. A probabilistic learning element as described in claim 2, wherein said items of previously learned information may occur a plurality of times and said predetermined types of knowledge include the number of occurrences of each said item of stored previously learned information.
• 4. A probabilistic learning element as described in claim 3, wherein said states each have a length and said predetermined types of knowledge further include the length of each state, the number of occurrences of each length, sequences of state lengths, the number of occurrences of each sequence of state lengths, state-length pairs, the number of occurrences of each state-length pair, sequences of state-length pairs and the number of occurrences of each sequence of state-length pairs.
• 5. A probabilistic learning element as described in claim 4, wherein the means for correlating includes a first means for determining the conditional probabilities that possible states will span an object sequence having particular begin and end times and second means for determining the conditional probabilities of possible state-length pairs given the previous state-length pair context.
• 6. A Probabilistic learning element as described in claim 5, additionally comprising means responsive to the conditional probabilites that possible states will span an object sequence having particular begin and end times and the conditional probabilities of possible state-length pairs given the previous state-length pair context to implement an algorithm known as the Viterbi Algorithm and provide probabilites of possible state-length pairs that span a particular object sequence given the previous state-length pair context of each possible state-length pair.
• 7. A probabilities learning element as described in claim 5, wherein the first means is responsive to two types of conditional probability signals, a first type signal corresponding to the conditional probabilities of object sequences occurring within a state given the state and a second type signal corresponding to the conditional probabilities of states with a particular begin time given an end time, object sequence and a state.
• 8. A probabilistic learning element as described in claim 7, wherein the first type probability signal is derived from the conditional probabilities of an object occurring given the previous object context and state which probabilities are calculated from the stored learned information relating to objects and object occurrences.
• 9. A probabilistic learning element as described in claim 7, wherein the second type probability signal is derived from conditional probabilities of an object occurring in a particular position in a state given the previous object context its position and state which probabilities are derived from the stored learned information relating to objects, object occurrences and the object positional information.
• 10. A probabilistic learning element as described in claim 5, wherein the second means for determining the conditional probabilities of state-length pairs is responsive to the stored learned information relating to previously learned states and their occurrences, lengths of previously learned states and their occurrences, state-length pairs from previously learned states and their occurrences, sequences of previously learned states and their occurrences, sequences of lengths of previously learned states and their occurrences, sequences of state-length pairs and their occurrences.
• 11. A probabilistic learning element as described in claim 1, wherein the means for storing are adapted to store the information in accordance with the context in which the stored information statistically occurred, whereby from any stored information the stored information which statistically occurs next in context is directly accessible and the conditional probabilities may be easily derived from the contextually stored information.
• 12. A probabilistic learning element as described in claim 1, additionally comprising means for providing a rating of confidence in said sequence of recognized states.
• 13. A probabilistic learning element as described in claim 12, additionally comprising means, responsive to said rating of confidence, for causing said means for storing items of previously learned information to store the recognized state sequence, the objects, sequences of objects and states forming the recognized state sequence and the predetermined types of knowledge relating to the objects, sequences of objects, states and sequences of states forming said recognized state sequence as items of previously learned information when the rating exceeds a predetermined threshold level.
• 14. A probabilistic learning element as described in claim 1, additionally comprising learning supervision means, responsive to external reinforcement signals, for causing said means for storing items of previously learned information to store the recognized state sequence, the objects, sequences of objects and states forming the recognized state sequence and the predetermined types of knowledge relating to the objects, sequences of objects, states and sequences of states forming the recognized state sequence as items of previously learned information.
• 15. A probabilities learning element as described in claim 14, additionally comprising means for correcting a recognized state sequence prior to initiating an external reinforcement signal.
• 16. A probabilistic learning element as described in claim 1, additionally comprising;
• means for providing a rating of confidence in said sequence of recognized states;
• learning supervision means adapted to receive said rating of confidence and an external reinforcement signal, said means being responsive to the rating of confidence of the recognized state sequence and the external reinforcement signal for providing an output signal when either the rating of confidence exceed a predetermined threshold level or an external reinforcement signal is received; and
• means responsive to the output signal from the learning supervision means to cause said means for storing items of previously learned information to store the recognized state sequence, the objects, sequences of objects, and states forming the recognized state sequence and the predetermined types of knowledge relating to the objects, sequences of objects, states and sequences of states forming said sequence of recognized states as items of previously learned information.
• 17. A probabilistic learning element as described in claim 16, additionally comprising means for correcting a recognized state sequence prior to initiating an external reinforcement signal.
• 18. A probabilistic learning element that sequentially receives objects and outputs sequences of recognized states and includes context driven searching, said learning element comprising:
• means for sequentially receiving objects;
• short term memory means for storing, in sequential context, said received objects;
• a context organized memory means comprising a plurality of tree structures for storing items of previously occurring learned information, said items including,
• states and the number of previous occurrences of said states, said states each having a length,
• objects contained in said states and the number of previous occurrences of said objects,
• lengths of said states and the number of occurrences of said state lengths, and
• state-length pairs in said states and the number of occurrences of said state-length pairs, said items of stored information being stored in accordance with the context in which the items of stored information statistically occurred, whereby from any items of stored information an item of stored information which statistically occurs next in context is directly accessible;
• said tree structures used to store the object information include an alltree structure and a plurality of singletree structures, the alltree structure stores the contextual occurrences of all objects received by the probabilistic learning element and at each node of the alltree there is provided an attribute list which refers to singletrees that include the same object context as the node of the alltree, a singletree is provided for each said state, whereby searching is facilitated by using the alltree as a pointer to the less complex singletrees;
• means for correlating said received objects stored in the short term memory means with information stored in the context organized memory means, said correlation being facilitated by use of the context of said received object stored in the short term memory means as a pointer to the context of the statistically stored information in the context organized memory means, said correlating means assigning conditional probabilities to possible sequences of recognized states;
• means, responsive to said conditional probabilities, for determining a most likely state sequence;
• means, responsive to the stored information, to determine a probability of an end of a state; and
• means, responsive to the end-of-state probability, for outputting said most likely state sequence as a sequence of recognized states.
• 19. A probabilistic learning element as described in claim 18, wherein each singletree contains object information for a state with each node representing an object and having an attribute list, said attribute list including information relating to said object including the objects distance from said states beginning, the objects distance to said states end and the number of times that the object appeared at that particular position within a state.
• 20. A probabilistic learning element as described in claim 18, additionally comprising means for providing a rating of confidence in said sequence of recognized states.
• 21. A probabilistic learning element as described in claim 20, additionally comprising means responsive to said rating of confidence to cause said context organized memory to store the objects and states forming the recognized state sequence, the lengths and state-length pairs of said states and the predetermined types of knowledge relating to the objects, states, state lengths and state-length pairs from said recognized state sequence as items of previously learned information when the rating exceeds a predetermined threshold level.
• 22. A probabilistic learning element as described in claim 18, additionally comprising learning supervision means responsive to external reinforcement signals to cause said context organized memory to store the objects and states forming the recognized state sequence the lengths and state-lengths pairs of said states and the predetermined types of knowledge relating to the objects, states, state lengths and state-length pairs from the recognized state sequence as items of previously learned information.
• 23. A probabilistic learning element as described in claim 22, additionally comprising means for correcting a recognized state sequence prior to initiating an external reinforcement signal.
• 24. A probabilistic learning element as described in claim 18, additionally comprising;
• means for providing a rating of confidence in said sequence of recognized states;
• learning supervision means adapted to receive said rating of confidence and an external reinforcement signal, said means being responsive to the rating of confidence of the recognized state sequence and the external reinforcement signal for providing an output signal when either the rating of confidence exceeds a predetermined threshold level or an external reinforcement signal is received; and
• means responsive to the output signal from the learning supervision means to cause said context organized memory to store the objects and states forming the recognized state sequence, the lengths and state-length pairs of said states and the predetermined types of knowledge relating to the objects, states, state lengths and state-length pairs from said sequence of recognized states as items of previously learned information.
• 25. A probabilistic learning element as described in claim 24, additionally comprising means for correcting a recognized state sequence prior to initiating an external reinforcement signal.
• 26. A probabilistic learning element that sequentially receives objects and outputs sequences of recognized states, said learning element comprising:
• means for sequentially receiving objects;
• means for storing,
• said received objects;
• sequences of received objects,
• sequences of previously learned states,
• states contained in said sequences of previously learned states,
• objects contained in said states contained in said sequences of previously learned states,
• sequences of said objects contained in said states contained in said sequences of previously learned states, and
• predetermined types of knowledge relating to
• said sequences of previously learned states,
• states contained in said sequences of previously learned states,
• objects contained in said states contained in said sequences of previously learned states, and
• sequences of said objects contained in said states contained in said sequences of previously learned states, so that current object information relating to said received objects and sequences of objects is stored as well as statistical information relating to said previously learned sequences of states, said states, objects and sequences of objects contained in said previously learned sequences of states;
• means for correlating said current object information with stored statistical information relating to previously learned sequences of states for assigning conditional probabilities to possible sequences of recognized states;
• means, responsive to said conditional probabilites of possible sequences of recognized states, for determining a most likely state sequence;
• means, responsive to the stored current object information and statistical information, to determine a probability of an end of a state;
• means, responsive to the probability of an end of a state, for outputting the most likely state sequence as a sequence of recognized states; and
• means for providing a rating of confidence in said sequence of recognized states said means including means for deriving support coefficients relating to how much information was available when calculating the conditional probabilities, said confidence rating being a function of the conditional probabilities and the support coefficients for the conditional probabilities used to determine the most likely state sequence.
• 27. A probabilistic learning element as described in claim 26, additionally comprising learning supervision means responsive to external reinforcement signals to cause said means for storing to store the recognized state sequence, the objects, sequences of objects and states forming the recognized state sequence and the predetermined types of knowledge relating to the objects, sequences of objects, states and sequences of states forming the recognized state sequence as items of previously learned information.
• 28. A probabilistic learning element as described in claim 27, additionally comprising means for correcting a recognized state sequence prior to initiating an external reinforcement signal.
• 29. A probabilistic learning element as described in claim 26, additionally comprising;
• learning supervision means adapted to receive said rating of confidence and an external reinforcement signal, said means being responsive to the rating of confidence of the recognized state sequence and the external reinforcement signal for providing an output signal when either the rating of confidence exceed a predetermined threshold level or an external reinforcement signal is received; and
• means responsive to the output signal from the learning supervision means to cause said means for storing to store the recognized state sequence, the objects, sequences of objects, and states forming the recognized state sequence and the predetermined types of knowledge relating to the objects, sequences of objects, states and sequences of states forming said sequence of recognized states as items of previously learned information.
• 30. A probabilistic learning element as described in claim 29, additionally comprising means for correcting a recognized state sequence prior to initiating an external reinforcement signal.
• 31. A probabilistic learning element as described in claim 26 additionally comprising means responsive to said rating of confidence to cause said means for storing to store the recognized state sequence, the objects, sequences of objects and states forming the recognized state sequence and the predetermined types of knowledge relating to the objects, sequences of objects, states and sequences of states forming said recognized state sequence as items of previously learned information when the rating exceeds a predetermined threshold level.
• 32. A probabilistic learning element that sequentially receives objects and outputs sequences of recognized states, said learning element comprising:
• means for sequentially receiving objects;
• short term memory means for storing said received objects in sequential context;
• context organized memory means, for storing items of previously occurring learned information, including a plurality of tree structures, each tree having a plurality of connected nodes, said plurality of tree structures including,
• an alltree structure having objects stored at the nodes of the tree along with the number of previous occurrences of each object, said alltree storing all objects contained in previously learned states in context so that from any stored object, objects which statistically occur next in context are directly accessible, each node of the alltree including an attribute list pointing to nodes of singletrees having objects stored therein in the same context as the context of the alltree node,
• a plurality of singletrees, one for each previously learned state, each node of the singletrees storing an object in context along with the number of previous occurrences of said object and an attribute list including positional information relating to the position of the object within the state and the number of previous occurrences of the object in that position,
• a tree structure for storing learned states in context so as to include states, the number of previous occurrences of each state, sequences of states and the number of previous occurrences of each state sequences,
• a tree structure for storing lengths of learned states in context so as to include state lengths, the number of previous occurrences of each state length, sequences of state lengths and the number of previous occurrences of each state length sequence, and
• a tree structure for storing state-length pairs of learned states in context so as to include the number of previous occurrences of each state-length pair, sequences of state-length pairs and the number of previous occurrences of each state-length pair sequence;
• means for correlating said received objects stored in the short term memory means with information stored in the context organized memory means, said correlation being facilitated by use of the context of said received objects stored in the short term memory means as a pointer to the context of the stored information in the context organized memory means, said correlating means assigning conditional probabilities to possible sequences of recognized states;
• means, responsive to said conditional probabilities, for determining a most likely state sequence;
• means, responsive to the stored information, to determine a probability of an end of a state; and
• means, responsive to the end-of-state probability, for outputting said most likely state sequence as a sequences of recognized states.
• 33. A probabilistic learning element as described in claim 32 additionally comprising means for providing a rating of confidence in said recognized state sequence said means including means for deriving support coefficients relating to how much information was available when calculating the conditional probabilities, said confidence rating being a function of the conditional probabilities and the support coefficients for the conditional probabilities used to determine the most likely state sequence.
• 34. A probabilistic learning element as described in claim 32, additionally comprising:
• means for providing a rating of confidence in said sequence of recognized states;
• learning supervision means adapted to receive said rating of confidence and an external reinforcement signal, said means being responsive to the rating of confidence of the recognized state sequence and the external reinforcement signal for providing an output signal when either the rating of confidence exceeds a predetermined threshold level or an external reinforcement signal is received; and
• means responsive to the output signal from the learning supervision means to cause said context organized memory to store the objects and states forming the recognized state sequence, the lengths and state-length pairs of said states and the predetermined types of knowledge relating to the objects, states, state lengths and state-length pairs from said sequence of recognized states as items of previously learned information.
• 35. A probabilistic learning element as described in claim 34, additionally comprising means for correcting a recognized state sequence prior to initiating an external reinforcement signal.
• 36. A probabilistic learning element as described in claim 32, additionally comprising means for providing a rating of confidence in said sequence of recognized states.
• 37. A probabilistic learning element as described in claim 36, additionally comprising means responsive to said rating of confidence to cause said context organized memory to store the objects and states forming the recognized state sequence, the lengths and state-length pairs of said states and the predetermined types of knowledge relating to the objects, states, state lengths and state-length pairs from said recognized state sequence as items of previously learned information when the rating exceeds a predetermined threshold level.
• 38. A probabilistic learning element as described in claim 32 additionally comprising learning supervision means responsive to external reinforcement signals to cause said context organized memory to store the objects and states forming the recognized state sequence the lengths and state-lengths pairs of said states and the predetermined types of knowledge relating to the objects, states, state lengths and state-length pairs from the recognized state sequence as items of previously learned information.
• 39. A probabilistic learning element as described in claim 38, additionally comprising means for correcting a recognized state sequence prior to initiating an external reinforcement signal.
• 40. A probabilistic learning element as described in claim 32, wherein the means for correlating comprises:
• means for correlating the object information stored in the context organized memory means with the object information stored in the short term memory means for determining conditional probabilities that possible states will span an object sequence having a particular begin time and end time;
• means for correlating the state, length and state-length pair information stored in the context organized memory means for determining conditional probabilities of state-length pairs given the previous state-length pair context; and
• means, responsive to the two previously mentioned conditional probabilities, for implementing an algorithm known as the Viterbi Algorithm and for providing probabilities of possible states, with a particular length that spans an object sequence given the previous state-length pair context.
• 41. A probabilistic learning element as described in claim 32, wherein the means responsive to stored information comprises means responsive to the object information stored in short term memory means and the object information stored in the context organized memory means for providing a probability signal corresponding to the probability that a state has ended.