Application of an artificial intelligence algorithm model of memory retrieval and roaming in sorting Chinese medicinal materials

Abstract

In order to capture autobiographical memory, inspired by the development of human intelligence, a computational AM model for autobiographical memory is proposed in this paper, which is a three-layer network structure, in which the bottom layer encodes the event-specific knowledge comprising 5W1H, and provides retrieval clues to the middle layer, encodes the related events, and the top layer encodes the event set. According to the bottom-up memory search process, the corresponding events and event sets can be identified in the middle layer and the top layer respectively; At the same time, AM model can simulate human memory roaming through the process of rule-based memory retrieval. The computational AM model proposed in this paper not only has robust and flexible memory retrieval, but also has better response performance to noisy memory retrieval cues than the commonly used memory retrieval model based on keyword query method, and can also imitate the roaming phenomenon in memory.

Keywords

Artificial intelligence cognitive model autobiographical memory roaming in memory

1. Introduction

Cognitive model is an important research direction in the field of artificial intelligence, of which autobiographical memory (AM) is a system which encodes, stores, and guides the retrieval of all remembered set information related to people’s personal experiences [1]. Although AM is an important part of thinking way, there are few modeling studies on it in China, such as encoding, retrieving and roaming through memories. Memory, usually considered as a function of AM, plays a vital role in self-acceptance and self-change. Roaming in memory refers to recalling a series of related AMs which span different life events, which is also the basis of memory therapy, usually used for the elderly to improve their psychological and cognitive health [2].

Remembered set memory and AM refer to the memory set of past events experienced by a person. However, the latter can be regarded as a special type of remembered set memory, which contains a person’s life experience from an individual point of view. Nevertheless, most existing computational remembered sets and AM modules do not differ significantly in their use and presentation.

Memory module is an important part of various cognitive models. For example, the cognitive models proposed in the references [3, 4, 5, 6] include both short-term working memory module and long-term memory module, which may not specify the exact type of long-term memory module used, such as remembered set [7], semantics [8] or autobiography [9]. The specific model in reference [10] explicitly describes a combination of remembered set memory modules. In the reference [11], a memory model for computing remembered sets is proposed, which is independent of other cognitive modules, and clearly defines the formation, retrieval and forgetting of past events in computer games. However, the use of this model is limited to the recall of historical data, and does not contain emotions [12] as one of the input fields, which is an important element in AM. The AM model named Xapagy proposed in the reference [13] is designed to realize narrative reasoning, and its activities are roughly similar to some psychological processes shown by human beings in stories. Xapagy uses complex natural language processing methods limited to storytelling. An online system is developed in the reference [14] that enables users to construct visual memories based on their mobile data as a form of visual AM for self-reflection and experience sharing. In the reference [15], the interaction between human and robot is stored as AM, so that the humanoid robot can accumulate experience and extract the regularity. However, when retrieving the stored memory, the aforementioned AM models only use the minimum amount of index knowledge to call a simple retrieval only for all the memories, all the memories of specific users, or all the memories that constitute the selected verbs. a computational AM model based on keyword query and used in the reference [16] for memory retrieval.

There is little research on roaming phenomenon in the literature about remembered set memory and AM model. In the reference [17, 18], an attempt was made to use the self-trapping attractor neural network to simulate the roaming effect in short-term and long-term associative memory. They refer to roaming in networks as a mechanism that allows “sparsely connected networks to wander around attractors far away from their initial state”.

Different from the aforementioned computational AM models, the computational AM model proposed in this paper aims to capture memory, including a picture snapshot of a person’s life experience and the related background, i.e., time, place, character, activity and emotion. It can use different types of cues to retrieve the encoded AM, and simulate the roaming phenomenon of human thinking in memory [19].

Autobiographical memory refers to the mixed memory of a person’s complex life events, which is closely related to the self-experience of memory. In the field of memory, research related to the self has a long history. Of course, the early psychologists did not clearly define the category of autobiographical memory, but only involved this field in the discussion of memory theory [20].

Deep learning is to learn the internal laws and representation levels of sample data. The information obtained in the learning process is of great help to the interpretation of data such as text, images and sounds. Its ultimate goal is to enable machines to have the ability to analyze and learn like humans, and to recognize data such as text, images, and sounds. Deep learning is a complex machine learning algorithm that has achieved results in speech and image recognition far surpassing previous related technologies [21].

2. AM model and its function realization

2.1 Psychological basis of AM model

In various AM models established by psychologists, the AM knowledge is divided into three levels, namely, life cycle, ordinary events and event-specific knowledge, from ordinary to specific, as shown in Fig. 1. If the cue is specific and personal, AM can be accessed directly, and if the cue is ordinary, a retrieval process must be generated to get more specific cues of related memory retrieval. The main difference between direct retrieval and generative retrieval is that the control process in generative retrieval adjusts the retrieval process, and the difference between these two memory retrieval methods is supported by neuroimaging evidence.

Figure 1.

Schematic diagram of hierarchical structure of AM.

2.2 AM model and its dynamic analysis

As shown in Fig. 2, the network architecture of AM model proposed in this paper is a top-down three-layer network structure, in which the $F_{3}$ layer, $F_{2}$ layer and $F_{1}$ layer respectively code the life cycle, general events and event-specific knowledge. The corresponding relationship can be drawn from the example given in Fig. 1. For example, the life experience of “working in $A$ ” can be expressed as the code in $F_{3}$ (learning set), and the related events of this remembered set, i.e., the “first day in work”, “in Office $C$ working” and “drinking on Saturday night in $V$ ”, can be expressed as the code in $F_{2}$ (learning event). A specific event, such as “drinking on Saturday night in $V$ ”, can be read out in $F_{1}$ , i.e., being happy (emotional) on Friday night (time) in $V$ (location) and drinking with colleagues (characters), drinking (activities), feeling happy (emotions), and image memory (images).

The AM model proposed in this paper adopts the dynamic nature of self-organizing network, in which the bottom layer encodes event-specific knowledge, which consists of 5W1H, the bottom layer provides retrieval clues, the middle layer associates event specific knowledge and encodes events. According to the bottom-up memory search procedure, the corresponding events and remembered sets can be identified in the middle layer and the top layer respectively.

With reference to the $F_{1}$ layer (consisting of 6 input fields) and $F_{2}$ layer (consisting of 1 associated field) shown in Fig. 2, the dynamics of the AM model herein are presented.

Figure 2.

Network architecture of AM model proposed in this paper.

Input vector: Let ${\bm{I}}^{l}=(I^{l}_{1},I^{l}_{2},\ldots,I^{l}_{M})$ represent the input vector, where $I^{l}_{j}$ represents the input $j$ of channel $l$ , $j=$ 1, 2, $\ldots$ , $L$ and $l=$ 1, 2, $\ldots$ , $N$ , where $M$ represents the length of channel $l$ and $N$ represents the number of channels.

Input field: Let $F^{l}_{1}$ represent an input field for saving the input mode of channel $l$ , and ${\bm{x}}^{l}=(x^{l}_{1},x^{l}_{2},\ldots,x^{l}_{M})$ represent the activation vector of $F^{l}_{1}$ for receiving ${\bm{I}}^{l}$ , where $x^{l}_{i}$ . $\in$ [0, 1]. In order to prevent code diffusion when fuzzy operation (see Eqs (1) and (4)) is adopted, complement coding is adopted, so that each ${\bm{x}}^{l}$ gets a complement [22] vector $\bar{x}^{l}_{i}$ to male $\bar{x}^{l}_{i}=1-x^{l}_{i}$ .

Link field: Let $\bm{y}=(\bm{y}_{1},\bm{y}_{2},\ldots,\bm{y}_{D})$ represent the activation vector of $F_{2}$ , where $D$ represents the number of codes in $F_{2}$ .

Weight vector: Let $\bm{w}^{l}_{j}$ represent the weight vector of the $j$ -th code in $F_{2}$ , used to learn the input mode in $F^{m}_{1}$ .

Parameter: The change of AM is affected by the parameters associated with the input fields in the lower layer, namely, selection parameter $\alpha^{m}>$ 0, learning speed parameter $\beta^{m}$ . $\in$ [0, 1], contribution parameter $\gamma^{m}\in$ [0, 1] ( $\Sigma\gamma^{m}=$ 1) and vigilance parameter $\rho^{m}\in$ [0, 1].

Code activation: The knowledge search begins by calculating the activation of all codes in $F_{2}$ (selecting function values). Specifically, given ${\bm{x}}^{m}$ , for each $F_{2}$ code $j$ , the selection function $T_{j}$ is calculated as follows:

$\displaystyle T_{j}=\mathop{\sum}\limits_{m}\gamma\frac{|x^{m}\wedge w_{j}^{m}% |}{\alpha^{m}+\left|{w_{j}^{m}}\right|}$ (1)

where the fuzzy AND operation “ $\wedge$ ” is defined by $p_{j}\wedge q_{j}\equiv$ min ( $p_{j}$ , $q_{j}$ ) and the norm $|\cdot|$ is defined by $|\bm{p}|=\sum_{j}p_{j}$ .

Code competition: The code competition process is as follows, in which the $F_{2}$ code with the highest selection function value is identified, and the winner is marked as index $I$ , wherein:

$\displaystyle T_{i}=\text{argmax}\{T_{i}:\textit{for all}F_{2}\ \textit{code}% \ i\}$ (2)

Template matching: It is checked whether the selected code $I$ resonates during the template matching process. Specifically, if the matching function $m^{l}_{I}$ meets its vigilance rules for each channel $l$ , resonance occurs, thus:

$\displaystyle m_{j}^{l}=\frac{|x^{l}\wedge w_{j}^{l}|}{\left|{x^{l}}\right|}% \geqslant\rho^{l}$ (3)

If any vigilance constraint is violated, a mismatch reset occurs, in which $T_{I}$ is set to 0 during input representation. Therefore, another $F_{2}$ code will be chosen as the new winner. This search and evaluation process will be guaranteed to end as a submitted code or an uncommitted code that meets the vigilance rules (all weights are initialized to 1 to meet the rules) will be identified to encode the new input pattern. Once an uncommitted code is adopted, a new uncommitted code will be automatically added to $F_{2}$ . In this way, AM model self-organizes its network structure.

Template learning: Once the code $I$ is identified as resonating, the weight vector $\bm{w}^{l}_{I}$ can be updated for each channel $l$ by the following learning rules:

$\displaystyle w_{I}^{\textit{l(new)}}=(1-\rho^{l})w_{I}^{\textit{l(old)}}+\rho% ^{l}(x^{l}\wedge w_{I}^{\textit{l(old)}})$ (4)

Knowledge reading: This top-down retrieval process is called when the selected $F_{2}$ code $I$ provides its weight vector to the input field in $F_{1}$ , so $\bm{x}^{\textit{l(new)}}=\bm{w}^{l}_{I}$ .

Template masking: Because of the dynamic nature of adaptive resonance, it is not required to provide all input vector knowledge retrieval. In this case, all values of the missing input vector $\bm{x}^{l}$ (including the complement set) are set to 1.

2.3 Code and search of input field

In the AM model (see Fig. 2), the input fields in $F_{1}$ are coded 5W1H respectively. In order to make the activation vector $\bm{x}^{k}$ compact and universal, t the normalized is used to encodes the place and time, the explicit value is used to represent the characters, activities, emotions and images.

Time vector ( $\bm{x}^{1}$ ): It uses six values to represent the time of the event, namely, normalized day ( $x^{1}_{1}=I^{1}_{1}$ /31), month ( $x^{1}_{2}=I^{1}_{2}$ /12) and year ( $x^{1}_{3}=(I^{1}_{3}-1900)/200)$ , and their complementary sets.

Place vector ( $\bm{x}^{2}$ ): It represents the location of the event with four values, i.e., normalized latitude ( $x^{2}_{1}=(I^{2}_{1}+90)/180)$ and longitude ( $x^{2}_{2}=(I^{2}_{2}+180)/360)$ ( $\bm{I}^{2}$ represents the input location vector).

Character vector ( $\bm{x}^{3}$ ):It’s a vector representing the people involved in the event, with the length corresponds to the classification of people in relation.

Activity vector ( $\bm{x}^{4}$ ): It’s a vector represents the content in an event, with the length corresponding to the classification of activity similarly.

Emotion vector ( $\bm{x}^{5}$ ): It’s a vector represents how it feels during in an event. Emotion, affects the encoding and retrieval of AM, which is divided into nine types, i.e., neutral, surprised, excited, happy, satisfied, tired, sad, painful and angry, as shown in Fig. 3. Therefore, $\bm{x}^{5}$ has a length of 18. This classification follows the happiness-arousal model established in the reference.

Image vector ( $\bm{x}^{6}$ ): It’s a vector t represents the picture memory related in an event. Its value encodes the file path where the image is stored, and it is not used as a part of the retrieval cue together with other vectors in the process of memory retrieval.

The $F_{2}$ layer coding event of AM model. The pseudo code of the event encoding and retrieval process is shown in Algorithm 1.

Algorithm 1: Pseudocode for event encoding and retrieval process
1. For a given input mode $\bm{I}^{k}$ , encode $\bm{x}^{k}$ in $F_{1}$
2. Activate all codes in $F_{2}$ //see Eq. (1)
3. Repeatedly select the winner code $J$ // see Eq. (2)
4. until resonance occurs//see Eq. (3)
5. if coding is required then execute learning // see Eq. (4)
6. end if
7. if retrieval is required then read $\bm{w}^{k}_{J}$ in $F_{1}$
8. end if

2.4 Coding and retrieval of remembered sets

Assuming that the related events of a remembered set occur at $t_{0}$ , $t_{1}$ , $\ldots$ , $t_{n}$ , let $y_{t_{i}}$ represent the activation value of events occurring at $t_{i}$ . In order to encode the event sequence, it is necessary to keep $y_{t_{n}}>y_{t_{n-1}}>\ldots>y_{t_{0}}$ all the time. Therefore, the decay parameter $\tau\in$ (0, 1) is used to adjust the activation attenuation.

The $F_{3}$ layer of the AM model encodes a remembered set of events to correlate the related events encoded in $F_{2}$ . The pseudo-code of event set coding and retrieval process is given in Algorithm 2.

Algorithm 2: Pseudocode of remembered set encoding and retrieval process
1. for all subsequent events do of a remembered set
2. select the winner code $J$ about $\bm{x}^{k}$ in $F_{1}$ in $F_{2}$
3. $\bm{y}_{J}\leftarrow$ 1
4. for all pre-selected code do in F2
5. $\bm{y}_{i}^{\textit{(new)}}\leftarrow\bm{y}_{i}^{\textit{(old)}}(1-\tau)$
6. end for
7. end for
8. Select the winner code $I^{\prime}$ about $\bm{y}$ in $F_{3}$
9. if code is required then learn the weight vector $\bm{w}^{\prime}_{I}$ in $F_{3}$ : $\displaystyle w_{I}^{\textit{l(new)}}=(1-\rho^{l})w_{I}^{\textit{l(old)}}+\rho% ^{l}(x^{l}\wedge w_{I}^{\textit{l(old)}}).$ 10. end if
11. if retrieval is required then read $w^{\prime}_{I^{\prime}}$ in $F_{2}$
12. end if

3. Roaming in memory

In the AM model proposed in this paper, roaming refers to the mode of rule memory search.

In this section, how to make AM model roam in memory is proposed. Roaming includes two main processes, namely, using changing cues at each iteration, changing search cues and iteratively searching AM.

In Algorithm 3, the pseudo-code of changing the process of searching cues is given, which is similar in concept to chromosome variation in genetic algorithm [14]. The change process is summarized as adding adjusted noise to a given search cue consciously at a randomly determined position, regulated by following parameters, namely, the variation rate $U\in$ [0, 1] and the noise level $M\in$ [0, 1]. The former is used to control the variation probability of the search cue value in a certain field, and the latter is used to control the change amount.

Algorithm 3: Variation pseudo-code of memory search cues

1. given a memory search cue

\bm{x}=\{\bm{x}_{1},\bm{x}_{2},\ldots,\bm{x}_{5}\}

2. for all

\bm{x}^{j}\in\bm{x}

3. if

\textit{rand}()\leqslant T

then //select the

j

-th field as the variation

4. for all

x^{j}_{i}

j=

1, 2,

\ldots

\bm{x}^{i}

/2 do

5. if

j\leqslant

2&&rand()

\leqslant

\bm{x}^{j}

then // select the

j

-th value for the normalized vector to mutate

x^{j}_{j}\leftarrow(1+(\textit{rand}()-0.5)L)\cdot x^{j}_{i}

7. limit

x^{j}_{i}

to [0,1]

8. else if

j\geqslant

3&&rand()

\leqslant

1+M

|\bm{x}^{j}|

then

9. cancel the binary value of

x^{j}_{i}

10. end if

11. end for

12. for all

x^{j}_{i}

i=|\bm{x}^{j}|/2+1

|\bm{x}^{j}|/2+2

\ldots

\bm{x}^{j}

x^{j}_{j}\leftarrow 1-x^{j}_{i-|xj|/2}

//calculate the complementary set

13. end for

14. end if

15. end for

In Algorithm 4, the pseudo-code of AM model roaming in memory is given. The roaming process is summarized as follows: based on the given memory search cues at the beginning of each iteration, the most relevant events are retrieved by using the given cues, which are not included in but will be added to the retrieved remembered set, and then the next iteration is carried out by using the changed retrieval cues. The termination rule of Algorithm 4 is to retrieve a predetermined number of events $N$ .

Algorithm 4: Pseudo-code implemented by AM model in roaming process

1. given a memory retrieval clue

\bm{x}=\{\bm{x}^{1}

\bm{x}^{2}

\ldots

\bm{x}^{5}\}

\rho^{m}\leftarrow

m=

1, 2,

\ldots

, 7 //remove all vigilance rules during memory search

M=\emptyset

4. repeat

5. repeat identifying

E

about

\bm{x}

F_{2}

//search for the especially event

\bm{y}_{E}\leftarrow

0 // inhibit

7. until

E\notin M

K\leftarrow K\cup\{E\}

// retrieval order saved in

K

9. variation

\bm{x}

// see Algorithm 3

10. until

|K|=N

11. retrieve all events in

K

4. Experimental results and discussion

A set of data was collected for the experiment, which consisted of 53 event snapshots of Ding Junhui, the champion of billiards in China, and the corresponding background [19]. The 53 events consisted of 12 event sets, each containing 3 to 7 events. All features except emotion were extracted directly from online web pages, and emotion was extracted manually from pictures and their backgrounds.

Based on the collected data set, eight types of relationships were defined, namely, strangers, colleagues, friends, acquaintances spouses, neighbors, family and classmates, and 15 types of activities were also defined, namely, catering, leisure, tourism, vacations, shopping, night outings, entertainment, sports, exercise, work, parties, socializing, celebrations, weddings and schools. After standardizing the input vectors, samples data were submitted to the AM model for encoding 53 events in $F_{2}$ (by Algorithm 1) and 12 event sets in $F_{3}$ (by Algorithm 2).

The listed in Table 1. Most of them adopt standard parameter values and do not need to be adjusted in the experiments.

Table 1
List of AM-parameters used in the experiment

Parameters	Values	Remarks
Selection parameter $\alpha^{m}$	0.2	Mainly used to avoid non-numerical values in Eq. (1)
Learning rate $\beta^{m}$	1	Not applicable during memory retrieval
Contribution parameter $\gamma^{m}$	0.167	Equally distributed so that $\Sigma\gamma^{m}=$ 1
Vigilance parameter $\rho^{m}$	2	Differed in different retrieval settings
Decay rate $\tau$	0.1	For encoding a sequence of events

4.1 AM retrieval using exact, partial, and noisy cues

After the AM model encodes the memory, the performance of memory retrieval was first tested using the following three types of cues:

(1)
Exact cues: An event was arbitrarily selected from the data set and its representation vector $\bm{y}=\{\bm{y}^{1},y^{2},\ldots,\bm{y}^{5}\}$ was used as the exact retrieval cue;
(2)
Partial cue: In a partial cue $\bm{y}_{\textit{Pa}}$ , for each field $m(m=1,2,\ldots,5)$ , an exact cue $\bm{y}$ and a cue integrity percentage $P$ were given. If rand() $\leqslant P$ /100, $\bm{y}^{m}_{\textit{Pa}}=\bm{y}^{m}$ ; Otherwise, $\bm{y}^{m}_{\textit{Pa}}=\{1,\ldots,1\}$ . $\bm{y}_{\textit{Pa}}=\bm{y}$ when $P=$ 100.
(3)
Noisy cue: In the selected $\bm{y}_{\textit{Pa}}$ field ( $\bm{y}_{\textit{Pa}}\neq\{1,\ldots,1\}$ ), given a partial cue $\bm{y}_{\textit{Pa}}$ and a noisy cue level $M^{\prime}$ , noise was introduced to get noisy cue $\bm{y}_{\textit{No}}$ . The process of introducing noise into search cues follows the process from step 4 to step 13 described in Algorithm 3, where $M=M^{\prime}$ .

As a benchmark, a keyword-based query approach was selected. A set of events will be retrieved after executing a keyword-based query to respond to a given search cue. If the event that was originally used to generate the given cue can be found in the search set, the search is considered to be successful; Otherwise, it is considered to be failed.

The vigilance parameter $\rho^{k}$ may be reduced for AM model to handle noisy cue so that the user can retrieve certain memories even if incomplete cues are provided, the AM model retrieves a set of memories and plays them feedback to the user. For each experiment, 10 repetitions were performed, with each retrieval cue randomly selected or generated, and the results summarized for further analysis.

Since the criteria of generating partial cues and calculating successful retrieval were adopted, both the AM model and the keyword-based query method in this paper achieved 100% success rates in response to both exact and partial retrieval cues. But it is challenging when dealing with noisy cues because uncertainty, while well handled by the human brain, is not necessarily handled by many computational models. Figure 3 shows the query method for memory retrieval response of noisy cues with different percentages of cue integrity $P$ and different levels of noisy cues $M^{\prime}$ .

Table 2
Naming conventions for setting experimental parameters

Name setting ( $U$ , $M$ ) Name setting ( $U$ , $M)$

s Sequential search r Random retrieval

Wss Roaming (0.0, 0.0) Wrr Roaming (1.0, 1.0)

W11 Roaming (0.4, 0.3) W12 Roaming (0.4, 0.4)

W21 Roaming (0.6, 0.3) W22 Roaming (0.6, 0.4)

Table 3
Partial subsequences of AM retrieved by AM model (W22) during roaming

Events in the event sets with time, place, characters, activities, emotion

Cues April 8, 2017 Beijing Friends Festival Night Happy Random initialization

April 8, 2017 Shanghai Friends and Festival Night Happy Event 1 in set 5

Acquaintances

August 10, 2017 Dalian Friends and Social Gathering Happy Event 3 in set 10

Strangers

August 10, 2017 Dalian Friends and Social Gathering Surprised Event 1 in set 10

Strangers

April 8, 2017 Shanghai Friends and Festival Night Happy Event 2 in set 5

Acquaintances

Figure 3.
The success rate of memory retrieval using AM model and keyword-based query methods for response to noisy cues.

According to Fig. 3, when $P=$ 0, all events in the dataset can be retrieved, so the success rate in Fig. 3a is 100%; However, as the $P$ value increases, the search cue is composed of more and more input fields. As a result, the memory retrieval performance of AM models and keyword-based query methods generally degrades as the values of cue integrity $P$ and level of noisy cue $M^{\prime}$ increase.

The comparison of Fig. 3 shows that the AM model in this paper obviously performs preferably in the successful search rate of response to noisy cues compared with the keyword-based query method, because it can preferably deal with the uncertainty in the retrieval process by reducing the vigilance value when dealing with noisy cues. Thus, in dealing with incomplete information, the AM model in this paper is more inclined to pursue human intelligence.
4.2 Test on roaming performance in memory

Name setting ( $U$ , $M$ )	Name setting ( $U$ , $M)$
s	Sequential search	r	Random retrieval
Wss	Roaming (0.0, 0.0)	Wrr	Roaming (1.0, 1.0)
W11	Roaming (0.4, 0.3)	W12	Roaming (0.4, 0.4)
W21	Roaming (0.6, 0.3)	W22	Roaming (0.6, 0.4)

In order to test the roaming performance of AM model in memory, different combinations of change rate $U$ and noise level $M$ given in Table 2 are used in different settings as the comparison. $T=$ 0.2 or 0.4 was selected, so that in the variation process, one or two fields of the cue can be expected to change according to the average value, and $M=$ 0.3 or 0.4 was selected to generate relatively small noises in the roaming process. Extreme cases (Wss and Wrr) were also tested, i.e. $U$ and $M$ were set as boundary values, so that the experimental results are compared with each other.

Table 3 shows some memory sequences retrieved by AM model during W22 configuration roaming process, as shown in Fig. 4 illustrates the roaming of AM model in this paper from set 5 (including 5 events) to set 10 (including 7 events), but roaming back to set 5 after 2 steps.

Figure 4.

Image playback of events retrieved in the order shown in Table 3.

Therefore, the AM model in this paper can effectively retrieve a modest subset of a person’s AM and can model the roaming in memory.

5. Conclusions

This paper proposes a computational model AM for autobiographical memory encoding and retrieval in online autonomous subjects that model the life experience of users. It is a three-layer network structure model that can imitate the mental wandering phenomenon in human autobiographical memory; For the memory retrieval of noisy clues, the performance of the AM model proposed in this paper is better than the keyword-based query method, because the latter cannot handle many existing photos or noisy clues in the memory repository; The AM proposed in this paper can also realize the function of roaming in memories, it can imitate a person’s memory sequence before and after different sets of events, that is, a moderate subset of a person’s autobiographical memory, and can imitate the phenomenon of roaming in memories.

Footnotes

Funding

Key project of natural science research in Colleges and universities of Anhui Province (KJ2018A0881); Bozhou “Artificial Intelligence $+$ ” Manufacturing Technology Application Research Center (ypzx002); Mechatronics professional teaching resource library (2020zyk28).

References

Chen

G.D.

and Li

L.H.

, Review on the research of random autobiographical memory, Psychological Science (1) (2014), 245–252.

Xia

, The positive effect of emotional memory and the influence of self reference in the elderly, Guangzhou: Jinan University, 2016.

Jiang

Y.H.

Fang

H.W.

H.Y.

and Liu

Y.X.

, A method of question semantic relation recognition based on bidirectional long short memory recurrent neural network, Journal of Fuzhou University (Natural Science Edition) 46(1) (2018), 52–58.

and Laird

, Spontaneous Retrieval from Long-Term Memory for a Cognitive Architecture, In Proceedings of Twenty-Ninth AAAI Conference on Artificial Intelligence, Austin, Texas, USA, 2015, pp. 44–550.

Lieto

Lebiere

and Oltramari

, The knowledge level in cognitive architectures: current limitations and possible developments, Cognitive Systems Research 48 (2018), 39–55.

Starzyk

J.A.

and Graham

, MLECOG: motivated learning embodied cognitive architecture, IEEE Systems Journal 11(3) (2017), 1272–1283.

Bisby

J.A.

Horner

A.J.

Bush

and Burgess

, Negative emotional content disrupts the coherence of episodic memories, Journal of Experimental Psychology General 147(2) (2018), 243–256.

Fan

Y.X.

Guo

J.F.

Lan

Y.Y.

and Cheng

X.Q.

, Context based deep semantic sentence retrieval model, Acta Sinica Sinica 31(5) (2017), 156–162.

Conway

M.A.

, A cross-cultural investigation of autobiographical memory, Journal of Cross-Cultural Psychology 36(6) (2005), 739–749.

10.

Frost

and Phillips

B.J.

, An Episodic Memory Retrieval Algorithm for the Soar Cognitive Architecture, In Proceedings of Australasian Joint Conference on Artificial Intelligence, Springer, Cham, 2015, pp. 343–355.

11.

Dietsche

Backes

and Stratmann

, Altered neural function during episodic memory encoding and retrieval in major depression, Human Brain Mapping 35(9) (2014), 4293–4302.

12.

Kotov

, A computational model of consciousness for artificial emotional agents, Psychology in Russia State of the Art 10(3) (2017), 57–73.

13.

Boloni

, Integrating perception, narrative, premonition and confabulatory continuation, Biologically Inspired Cognitive Architectures 8 (2014), 120–131.

14.

Thudt

Baur

and Huron

, Visual mementos: reflecting memories with personal data, IEEE Transaction on Visualization and Computer Graphics 22(1) (2016), 369–378.

15.

Pointeau

Petit

and Dominey

P.F.

, Successive developmental levels of autobiographical memory for learning through social interaction, IEEE Transactions on Autonomous Mental Development 6(3) (2014), 200–212.

16.

Harris

and Gittens

, Modeling believable agents using a descriptive approach, Biologically Inspired Cognitive Architectures 14 (2015), 10–21.

17.

Dorronsoro

and Pinel

, Combining Machine Learning and Genetic Algorithms to Solve the Independent Tasks Scheduling Problem, In Proceedings of 2017 3rd IEEE International Conference on Cybernetics, Exeter, UK, 2017, pp. 1–8.

18.

Pavloski

and Karimi

, The self-trapping attractor neural network. I. Analysis of a simple 1-D model, IEEE Transactions on Neural Networks 14(1) (2003), 58–65.

19.

Wang

and Tan

A.H.

, Creating autonomous adaptive agent in a real-time first-person shooter computer game, IEEE Transactions on Computational Intelligence and AI in Games 7(2) (2015), 123–138.

20.

Johnson

A.J.

and Morley

E.G.

, Sharing personal memories on ephemeral social media facilitates autobiographical memory cyberpsychology, Behavior and Social Networking 2021.

21.

Liu

H.H.

Nai

Y.H.

Saridin

Tanaka

O’ Doherty

Hilal

Gyanwali

Chen

C.P.

Robins

E.G.

and Reilhac

, Improved amyloid burden quantification with nonspecific estimates using deep learning, European Journal of Nuclear Medicine and Molecular Imaging 48 (2021), 1842–1853.

22.

Sharar

S.R.

Alamdari

Hoffer

Hoffman

H.G.

Jensen

M.P.

and Patterson

D.R.

, Circumplex model of affect: a measure of pleasure and arousal during virtual reality distraction analgesia, Games for Health Journal 5(3) (2016), 197–202.