KR102251961B1 - Method and apparatus of bidirectional hierarchical temporal memory for bidirectional prediction - Google Patents

Abstract

A bidirectional hierarchical temporal memory method and apparatus capable of bidirectional prediction according to various embodiments, learns input data and, based on the learned data, provides a first direction or a first direction opposite to the first direction according to time. It may be configured to predict data in at least one of the two directions.

Description

A bidirectional hierarchical temporal memory method and apparatus capable of bidirectional prediction {METHOD AND APPARATUS OF BIDIRECTIONAL HIERARCHICAL TEMPORAL MEMORY FOR BIDIRECTIONAL PREDICTION}

Various embodiments relate to a method and apparatus for a bidirectional hierarchical temporal memory capable of bidirectional prediction.

Hierarchical temporal memory (HTM) is a kind of artificial neural network, and it is based on cell connection based on human biological neurons. At this time, the cell structure of hierarchical temporal memory is modeled based on the neuron structure in the brain's neocortex. According to the hierarchical temporal memory, data patterns are learned based on time, and future data patterns are predicted from the learned data patterns. At this time, data patterns are learned based on the cell structure, and future data patterns can be predicted based on the cell structure.

However, the hierarchical temporal memory as described above is only used to predict future data patterns. According to hierarchical temporal memory, patterns of data in the past that have not been learned cannot be predicted. Accordingly, there is a need for a method capable of predicting not only future data patterns but also past data patterns based on a hierarchical temporal memory.

A bidirectional hierarchical temporal memory method capable of bidirectional prediction according to various embodiments includes an operation of learning input data, and a second direction opposite to the first direction according to the passage of time based on the learned data. It may include an operation of predicting data.

A bidirectional hierarchical temporal memory device capable of bidirectional prediction according to various embodiments includes an encoder that detects input data, and learns the detected data, and based on the learned data, a first according to the passage of time. A predictor for predicting data in a second direction opposite to the direction may be included.

According to various embodiments, the bidirectional hierarchical temporal memory device may predict past data based on learned data. That is, the bidirectional hierarchical temporal memory device can predict not only future data but also past data based on the learned data.

1 is a diagram illustrating a cell structure for a bidirectional hierarchical temporal memory according to various embodiments.
2 is a diagram illustrating a bidirectional hierarchical temporal memory device according to various embodiments.
3A and 3B are diagrams for explaining a bidirectional prediction operation of a bidirectional temporal memory device according to various embodiments.
4 is a diagram illustrating a method of operating a bidirectional hierarchical temporal memory device according to various embodiments.
5 is a diagram illustrating a data prediction operation in the first direction of FIG. 4.
6 to 11 are diagrams for explaining a data prediction operation in the first direction of FIG. 4.
12 is a diagram illustrating a data prediction operation in the second direction of FIG. 4.
13 to 16 are diagrams for explaining a data prediction operation in the second direction of FIG. 4.

Hereinafter, various embodiments of the present document will be described with reference to the accompanying drawings.

Various embodiments of the present document and terms used therein are not intended to limit the technology described in this document to a specific embodiment, and should be understood to include various modifications, equivalents, and/or substitutes of the corresponding embodiment. In connection with the description of the drawings, similar reference numerals may be used for similar elements. Singular expressions may include plural expressions unless the context clearly indicates otherwise. In this document, expressions such as "A or B", "at least one of A and/or B", "A, B or C" or "at least one of A, B and/or C" are all of the items listed together. It can include possible combinations. Expressions such as "first", "second", "first" or "second" can modify the corresponding elements regardless of their order or importance, and are only used to distinguish one element from another. It does not limit the components. When any (eg, first) component is referred to as being “(functionally or communicatively) connected” or “connected” to another (eg, second) component, the component is It may be directly connected to the component, or may be connected through another component (eg, a third component).

The term "module" used in this document includes a unit composed of hardware, software, or firmware, and may be used interchangeably with terms such as, for example, logic, logic blocks, parts, or circuits. The module may be an integrally configured component or a minimum unit that performs one or more functions, or a part thereof. For example, the module may be configured as an application-specific integrated circuit (ASIC).

1 is a diagram illustrating a cell structure for a bidirectional hierarchical temporal memory according to various embodiments.

Referring to FIG. 1, a cell structure for implementing various data patterns may be defined. According to the cell structure, a plurality of cells 100 for representing data patterns may be arranged in a column dimension 101 and a cell dimension 103. The cells 100 may be grouped into a number of columns 102, and the columns 102 may be arranged along the column dimension 101. In addition, the cells 100 may be hierarchically arranged along the cell dimension 103 in each column 102.

2 is a diagram illustrating a bidirectional hierarchical temporal memory device according to various embodiments. 3A and 3B are diagrams for explaining a bidirectional prediction operation of a bidirectional temporal memory device according to various embodiments.

Referring to FIG. 2, a bidirectional hierarchical temporal memory device 200 according to various embodiments may include an encoder 210, a bidirectional predictor 220, and a decoder 230. According to various embodiments, the bi-directional hierarchical temporal memory device 200 includes bidirectional (ti-1, ti+1) data from the current (to-ti) data pattern 311 as shown in FIG. 3A. The patterns 313 and 315 can be predicted. To this end, the bidirectional hierarchical temporal memory device 200 may learn a data pattern from an initial point of time to to a current point of time ti, for example, the data pattern 311 of the present (to-ti). In addition, the bidirectional hierarchical temporal memory device 200 predicts a data pattern 313 in a first direction, for example, a future (ti+1), based on the data pattern 311 of the present (to-ti) or the second The direction, for example, the data pattern 315 in the past (ti-1) can be predicted. In this case, the bidirectional hierarchical temporal memory device 200 may predict bidirectional (ti-1, ti+1) data patterns 313 and 315 based on the cell structure of FIG. 1 described above.

The encoder 210 may convert data v(t) input as multi-channel time series data. In this case, the encoder 210 may convert the input data v(t) from the original form into a binary vector form as shown in (a) of FIG. 3B. Through this, the encoder 210 may transmit the converted data (v _b (t)) to the bidirectional predictor 220.

The bidirectional predictor 120 may perform time series prediction based on the _{transformed data (v b (t)). The bidirectional predictor 220 may learn the transformed data v b} (t) based on a hierarchical temporal memory (HTM) algorithm. In addition, the bidirectional predictor 120 may predict bidirectional data based on the learned data. To this end, the bidirectional predictor 120 may include a spatial pooler and a temporal pooler. That is, the bidirectional predictor 220 may predict data in at least any one of a first direction, such as the future or a second direction, such as the past. In this case, the bidirectional predictor 220 may predict bidirectional data based on the cell structure as shown in (b) of FIG. 3B. For example, the bidirectional predictor 220 _{activates at least one of the cells based on the transformed data (v b} (t)), and based on the activated cell, the data in the first direction or the data in the second direction At least one can be predicted. Through this, the bidirectional predictor 220 may output the predicted data P _b (t) to the decoder 230. _{In this case, the bidirectional predictor 220 may output data P b} (t) predicted in the form of a binary vector as shown in (c) of FIG. 3B.

The decoder 230 may _{retransform the predicted data P b} (t). Here, the decoder 230 may receive the predicted data P _b (t) in the form of a binary vector. In this case, the decoder 130 may _{reconvert the predicted data P b} (t) from the binary vector form to the original form. Through this, the decoder 230 may detect the reconverted data as output data. Here, the output data is at least one of first output data (p(t)) representing data in the first direction, for example, future data, or second output data (i(t)) representing data in the second direction, such as past data. It may include.

4 is a diagram illustrating a method of operating a bidirectional hierarchical temporal memory device according to various embodiments.

Referring to FIG. 4, the bidirectional hierarchical temporal memory device 200 may learn data input in operation 410. In this case, the encoder 210 may convert input data. Here, the encoder 210 may convert the input data from an original form into a binary vector form. Thereafter, the bidirectional predictor 210 may learn the transformed data based on a hierarchical temporal memory (HTM) algorithm. Through this, the bi-directional hierarchical temporal memory device 200 may learn data from the initial time point t0 to the current time point ti, for example, data of the present time (to-ti).

The bidirectional hierarchical temporal memory device 200 may predict bidirectional data based on the data learned in operation 420. The bidirectional data may represent data in a first direction and data in a second direction. Here, data in the first direction may represent data in the future (ti+1) that has not been learned, and data in the second direction may represent data in the past (ti-1) that has not been learned. In this case, the bidirectional predictor 210 may predict at least one of data in the first direction and data in the second direction based on the cell structure. After that, the decoder 230 may retransform the predicted data. Here, the decoder 230 may retransform the predicted data from a binary vector into an original form and output the reconverted data. Through this, the bidirectional hierarchical temporal memory device 200 may manage the output data. Here, the bidirectional hierarchical temporal memory device 200 may store the output data or provide it to the outside.

5 is a diagram illustrating a data prediction operation in the first direction of FIG. 4. 6 to 11 are diagrams for explaining a data prediction operation in the first direction of FIG. 4.

Referring to FIG. 5, the bidirectional hierarchical temporal memory device 200 may activate at least one of cells based on current (to-ti) data in operation 510. The current (to-ti) data may include first data at a first time point to and second data at a second time point ti later than the first time point.

According to various embodiments, the bidirectional hierarchical temporal memory device 200 determines at least one active cell based on first data of a first time point to, as shown in FIG. 6A, and , At least one prediction cell may be determined based on third data of a second time point (ti) predicted from the first data. In addition, the bidirectional hierarchical temporal memory device 200 may determine at least one active column based on the second data of the second time point (ti) as illustrated in FIG. 6B. At this time, if the prediction cell of the third data is included in the active column of the second data, the bidirectional hierarchical temporal memory device 200 accurately predicts the prediction cell of the second data as shown in FIG. 6C. It is determined that the second data is predicted, and the prediction cell of the second data may be defined as a winner cell and activated. Meanwhile, if the prediction cell of the third data is not included in the active column of the second data, the bidirectional hierarchical temporal memory device 200 deactivates the prediction cell of the second data as shown in FIG. 6B. can do. On the other hand, if the active column of the second data does not include the prediction cell of the third data, the bidirectional hierarchical temporal memory device 200 bursts the active column of the second data as shown in FIG. 6B. It can be determined as a burst column 601. In addition, the bidirectional hierarchical temporal memory device 200 determines all cells of the burst column 601 as active cells, as shown in FIG. 6C, and selects any one of the active cells of the burst column 601. It can be determined by the winner cell.

According to an embodiment, in order to determine a winner cell of a burst column, the bidirectional hierarchical temporal memory device 200 includes a segment based on first data of a first time point to, as shown in FIG. 7. ) A matching score of 710 and 720 may be detected. In this case, the matching score may be detected according to the number of synapses 730 connected to the winner cells of the first data. Thereafter, the bidirectional hierarchical temporal memory device 200 may determine any one of the segments 710 and 720 having the largest matching score as the best matching segment 720. Through this, the bidirectional hierarchical temporal memory device 200 determines the best matching cell 740 among cells of the burst column 601 using the best matching segment 720 and bursts the best matching cell 740. It can be determined as the winner cell of the column 601. Meanwhile, if the best matching segment 720 is not determined, the bidirectional hierarchical temporal memory device 200 may arbitrarily determine a winner cell in the burst column 601.

For example, the bidirectional hierarchical temporal memory device 200 includes synapses 831 and 820 of the segments 810 and 820 based on the first data and the third data of a first time point as shown in FIG. 8. 833) can be updated. Here, the segments 810 and 820 include the learning segment 820, and the learning segment 820 may be related to the winner cell 840 of the burst column 601. In this case, the bidirectional hierarchical temporal memory device 200 may strengthen the synapse 833 connected to the active cell and weaken the synapse 831 connected to the inactive cell. To this end, the bidirectional hierarchical temporal memory device 200 may increase the weight ρ of the synapse 831 connected to the active cell and decrease the weight ρ of the synapse 833 connected to the inactive cell. .

For example, the bidirectional hierarchical temporal memory device 200 may add a new synapse 935 to the learning segment 920 among the segments 910 and 920 as shown in FIG. 9. Here, the learning segment 920 may be related to the winner cell 940 of the burst column 601. In this case, the bidirectional hierarchical temporal memory 200 may add a new synapse 935 connected from an active cell of another segment 910 to the learning segment 920. Here, the number of new synapses 935 may be determined by subtracting the number of synapses 933 connected to the active cell from the number of synapses 931 connected to the prediction cell for the learning segment 920. For example, if the number of synapses 931 connected to the prediction cell for the learning segment 920 is 4 and the number of synapses 933 connected to the active cell is 3, the number of new synapses 935 is 1 Can be determined. Through this, the bidirectional hierarchical temporal memory device 200 may add one new synapse 935 to the learning segment 920.

The bidirectional hierarchical temporal memory device 200 may predict at least one of the cells based on the cell activated in operation 520. In this case, the bidirectional hierarchical temporal memory device 200 may predict at least one of the cells as data in the first direction, that is, future (ti+1).

According to various embodiments, the bidirectional hierarchical temporal memory device 200 may calculate the activity of all segments 1010 and 1020 as illustrated in FIG. 10. In this case, the activity level may be determined according to the number of active cells connected to each of the segments 1010 and 1020. For example, the bidirectional hierarchical temporal memory device 200 may calculate activity through synapses of each segment 1010 and 1020. The bidirectional hierarchical temporal memory device 200 may calculate a product of a weight for each synapse and a cell state value, and calculate the activity of each segment 1010 and 1020 as a sum of the calculation results for all synapses. In FIG. 10, a dotted line indicates a case in which the weight of the synapse is less than or equal to a predetermined reference value, and the solid line may indicate a case where the weight of the synapse exceeds the reference value.

For example, when the weight is less than or equal to the reference value, it may be determined as 0, and when it exceeds the reference value, it may be determined as 1. The cell state value may be determined as 1 for an active cell and 0 for an inactive cell. Through this, the bidirectional hierarchical temporal memory device 200 determines the activity of the first segment 1010, 1 * 1 + 1 * 0 + 0 * 0 + 0 * 0 + 1 * 0 + 0 * 1 + 0 * 0 It can be calculated as + 1 * 0 = 1. Similarly, the bidirectional hierarchical temporal memory device 200 indicates the activity of the second segment 1020, 1 * 0 + 0 * 0 + 1 * 1 + 0 * 0 + 1 * 1 + 0 * 0 + 1 * 1 = It can be calculated as 3.

According to various embodiments, the bidirectional hierarchical temporal memory device 200 may activate at least one of the segments 1010 and 1020 based on the activity of all the segments 1010 and 1020. At this time, when the activity of at least one of the segments 1010 and 1020 exceeds a predetermined threshold, the bidirectional hierarchical temporal memory device 200 activates the at least one segment 1020 and the remaining segments 1010 Can be disabled. For example, if the threshold value is 2, the bidirectional hierarchical temporal memory device 200 may activate the second segment 1020 and deactivate the first segment 1010. In addition, when at least one of the segments 1010 and 1020 is activated, the bidirectional hierarchical temporal memory device 200 uses the activated segment 1020 to determine a prediction cell corresponding to future (ti+1) data. (1040) can be determined. Through this, the bidirectional hierarchical temporal memory device 200 may determine at least one of the cells as a prediction cell corresponding to future (ti+1) data, as illustrated in FIG. 11.

12 is a diagram illustrating a data prediction operation in the second direction of FIG. 4. 13 to 16 are diagrams for explaining a data prediction operation in the second direction of FIG. 4.

Referring to FIG. 12, the bidirectional hierarchical temporal memory device 200 may activate at least one of cells based on current (to-ti) data in operation 1210. The current (to-ti) data may include first data at a first time point ti and second data at a second time point to earlier than the first time point.

According to various embodiments, the bidirectional hierarchical temporal memory device 200 determines at least one active cell based on first data of a first time point as shown in FIG. 13A and , At least one backward prediction cell may be determined based on the third data of a second view point (to) predicted from the first data. In addition, the bidirectional hierarchical temporal memory device 200 may determine at least one active column based on the second data of a second point in time to, as illustrated in FIG. 13B. At this time, if the reverse prediction cell of the third data is included in the active column of the second data, the bidirectional hierarchical temporal memory device 200 includes the reverse prediction cell of the second data as shown in FIG. 13C. It is determined that it is accurately predicted, and the backward prediction cell of the second data may be defined as a winner cell and activated. On the other hand, if the reverse prediction cell of the third data is not included in the active column of the second data, the bidirectional hierarchical temporal memory device 200 is a reverse prediction cell of the second data as shown in FIG. 13B. Can be disabled. On the other hand, if the active column of the second data does not include the backward prediction cell of the third data, the bidirectional hierarchical temporal memory device 200 selects the active column of the second data as shown in FIG. 13B. It may be determined as a burst column 1301. In addition, the bidirectional hierarchical temporal memory device 200 determines all cells of the burst column 1301 as active cells, as shown in (c) of FIG. 13, and selects any one of the active cells of the burst column 1301. It can be determined by the winner cell.

According to an embodiment, in order to determine the winner cell of the burst column, the bidirectional hierarchical temporal memory device 200, as shown in FIG. A matching score of (segment) 1410 and 1420 may be detected. In this case, the matching score may be detected according to the number of synapses 1430 connected to the winner cells of the first data. Thereafter, the bidirectional hierarchical temporal memory device 200 may determine any one of the segments 1410 and 1420 having the largest matching score as the best matching reverse segment 1420. Through this, the bidirectional hierarchical temporal memory device 200 uses the best matching reverse segment 1420 to determine the best matching cell 740 among cells of the burst column 1301, and selects the best matching cell 1440. It may be determined as a winner cell of the burst column 1301. Meanwhile, if the best matching reverse segment 1420 is not determined, the bidirectional hierarchical temporal memory device 200 may arbitrarily determine a winner cell in the burst column 1301.

For example, the bidirectional hierarchical temporal memory device 200 includes synapses 1531 of the reverse segments 1510 and 1520 based on the first data and the third data of a first time point as shown in FIG. 15. , 1533). Here, the segments 1510 and 1520 may include a learning reverse segment 1520, and the learning reverse segment 1520 may be related to the winner cell 1540 of the burst column 1301. In this case, the bidirectional hierarchical temporal memory device 200 may strengthen the synapse 1533 connected to the active cell and weaken the synapse 1531 connected to the inactive cell. To this end, the bidirectional hierarchical temporal memory device 200 may increase the weight ρ of the synapse 1531 connected to the active cell and decrease the weight ρ of the synapse 1533 connected to the inactive cell. .

The bidirectional hierarchical temporal memory device 200 may predict at least one of the cells based on the cell activated in operation 1220. In this case, the bidirectional hierarchical temporal memory device 200 may predict at least one of the cells as data in the second direction, that is, the past (ti-1).

According to various embodiments, the bidirectional hierarchical temporal memory device 200 may calculate the activity of all reverse segments 1610 and 1620 as illustrated in FIG. 16. In this case, the activity level may be determined according to the number of active cells connected to each of the reverse segments 1610 and 1620. For example, the bidirectional hierarchical temporal memory device 200 may calculate activity through synapses of each of the reverse segments 1610 and 1620. The bidirectional hierarchical temporal memory device 200 may calculate a product of a weight for each synapse and a cell state value, and calculate the activity of each segment 1610 and 1620 as a sum of the calculation results for all synapses. In FIG. 16, a dotted line indicates a case where the weight of the synapse is less than or equal to a predetermined reference value, and the solid line may indicate a case where the weight of the synapse exceeds the reference value.

For example, when the weight is less than or equal to the reference value, it may be determined as 0, and when it exceeds the reference value, it may be determined as 1. The cell state value may be determined as 1 for an active cell and 0 for an inactive cell. Through this, the bidirectional hierarchical temporal memory device 200 determines the activity of the first segment 1010, 1 * 1 + 1 * 0 + 0 * 0 + 0 * 0 + 1 * 0 + 0 * 1 + 0 * 0 It can be calculated as + 1 * 0 = 1. Similarly, the bidirectional hierarchical temporal memory device 200 indicates the activity of the second segment 1020, 1 * 0 + 0 * 0 + 1 * 1 + 0 * 0 + 1 * 1 + 0 * 0 + 1 * 1 = It can be calculated as 3.

According to various embodiments, the bidirectional hierarchical temporal memory device 200 may activate at least one of the reverse segments 1610 and 1620 based on the activity of all the reverse segments 1610 and 1620. have. At this time, when the activity of at least one of the reverse segments 1610 and 1620 exceeds a predetermined threshold, the bidirectional hierarchical temporal memory device 200 activates the at least one reverse segment 1620, and the remaining reverse segments (1610) can be disabled. For example, if the threshold value is 2, the bidirectional hierarchical temporal memory device 200 may activate the second reverse segment 1620 and deactivate the first reverse segment 1610. And when at least one of the reverse segments 1610 and 1620 is activated, the bidirectional hierarchical temporal memory device 200 uses the activated reverse segment 1620 to correspond to past (ti-1) data. The prediction cell 1640 may be determined. Through this, the bidirectional hierarchical temporal memory device 200 may determine at least one of the cells as a prediction cell corresponding to data of the past (ti-1), as illustrated in FIG. 17.

According to various embodiments, the bidirectional hierarchical temporal memory device 200 may predict past data based on learned data. That is, the bidirectional hierarchical temporal memory device 200 may predict not only future data but also past data based on the learned data.

Although various embodiments of the present document have been described, various modifications may be made without departing from the scope of the various embodiments of the present document. Therefore, the scope of the various embodiments of the present document is limited to the described embodiments and should not be defined by the scope of the claims to be described later, as well as the scope and equivalents of the claims.

Claims (20)

In the bidirectional data prediction method performed by a bidirectional hierarchical temporal memory device including an encoder, a predictor and a decoder,
Detecting, in the encoder, input data according to the passage of time;
Learning the detected data by the predictor;
Predicting, in the predictor, data in a second direction opposite to the first direction according to the passage of time based on the learned data; And
The decoder outputs data in the second direction
Including,
The encoder converts the input data from the original form into a binary vector form,
In the predictor, the operation of predicting data in the second direction,
Expressing the converted data by activating at least one of the cells in a cell structure in which a plurality of cells are arranged in a column dimension and a cell dimension according to a hierarchical temporal memory (HTM) algorithm;
Representing data in the second direction based on the cell structure; And
Outputting the data in the second direction in the form of a binary vector
Including,
The decoder converts the data in the second direction into an original form,
The operation of expressing the data in the second direction,
Determining at least one active cell in the cell structure based on first data of a first view;
Determining at least one prediction cell in the cell structure based on third data predicted from the first data according to the hierarchical temporal memory algorithm;
Determining at least one active column in the cell structure based on second data of a second view in the second direction from the first view according to the hierarchical temporal memory algorithm;
If the prediction cell is included in the active column, activating all cells of the active column; And
If the prediction cell is not included in the active column, deactivating all cells of the active column
Containing,
Two-way data prediction method.
delete In the bidirectional data prediction method performed by a bidirectional hierarchical temporal memory device including an encoder, a predictor and a decoder,
Detecting, in the encoder, input data according to the passage of time;
Learning the detected data by the predictor;
Predicting, in the predictor, data in a first direction according to the passage of time based on the learned data; And
In the decoder, outputting data in the first direction
Including,
The encoder converts the input data from the original form into a binary vector form,
In the predictor, the operation of predicting data in the first direction,
Activating at least one of the cells in a cell structure in which a plurality of cells are arranged in a column dimension and a cell dimension according to a hierarchical temporal memory (HTM) algorithm to represent the converted data;
Representing data in the first direction based on the cell structure; And
Outputting the data in the first direction in the form of a binary vector
Including,
The decoder converts the data in the first direction into an original form,
The operation of expressing the data in the first direction,
Determining at least one active cell in the cell structure based on first data of a first view;
Determining at least one prediction cell in the cell structure based on third data predicted from the first data according to the hierarchical temporal memory algorithm;
Determining at least one active column in the cell structure based on second data of a second view in the first direction from the first view according to the hierarchical temporal memory algorithm;
If the prediction cell is included in the active column, activating all cells of the active column; And
If the prediction cell is not included in the active column, deactivating all cells of the active column
Containing,
Two-way data prediction method.
delete delete delete delete delete delete delete In the bidirectional hierarchical temporal memory device capable of bidirectional prediction,
An encoder that detects input data over time;
A predictor for learning the detected data and predicting data in a second direction opposite to the first direction over time based on the learned data; And
A decoder that outputs the data in the second direction
Including,
The encoder,
Converting the input data from the original form to a binary vector form,
The predictor,
In a cell structure in which a plurality of cells are arranged in a column dimension and a cell dimension according to a hierarchical temporal memory (HTM) algorithm, at least one of the cells is activated to represent the converted data,
Based on the cell structure, expressing the data in the second direction,
Outputting the data in the second direction in the form of a binary vector,
The decoder,
Converting the data in the second direction into its original form,
The predictor,
Determine at least one active cell in the cell structure based on the first data at a first time point,
According to the hierarchical temporal memory algorithm, based on third data predicted from the first data, at least one prediction cell is determined in the cell structure,
According to the hierarchical temporal memory algorithm, based on second data of a second view in the second direction from the first view, at least one active column in the cell structure is determined,
When the prediction cell is included in the active column, all cells of the active column are activated,
If the prediction cell is not included in the active column, deactivating all cells of the active column,
Bidirectional hierarchical temporal memory device.
delete In the bidirectional hierarchical temporal memory device capable of bidirectional prediction,
An encoder that detects input data over time;
A predictor for learning the detected data and predicting data in a first direction over time based on the learned data; And
A decoder that outputs the data in the first direction
Including,
The encoder,
Converting the input data from the original form to a binary vector form,
The predictor,
In a cell structure in which a plurality of cells are arranged in a column dimension and a cell dimension according to a hierarchical temporal memory (HTM) algorithm, at least one of the cells is activated to represent the converted data,
Based on the cell structure, expressing the data in the first direction,
Outputting the data in the first direction in the form of a binary vector,
The decoder,
Converting the data in the first direction into its original form,
The predictor,
Determine at least one active cell in the cell structure based on the first data at a first time point,
According to the hierarchical temporal memory algorithm, based on third data predicted from the first data, at least one prediction cell is determined in the cell structure,
According to the hierarchical temporal memory algorithm, based on second data of a second view in the first direction from the first view, at least one active column in the cell structure is determined,
When the prediction cell is included in the active column, all cells of the active column are activated,
If the prediction cell is not included in the active column, deactivating all cells of the active column,
Bidirectional hierarchical temporal memory device.
delete delete delete delete delete delete delete

Images