Associative memory and recall model with KID model for human activity recognition
Introduction
In the recent years, there has been unprecedented growth in computer technology and related fields. Most daily activities such as communication, transport, healthy living and education are now mediated by computer technology. This growth can be largely attributed to the affordability of devices such as mobile phone handsets and personal computers. The field has also been awash with new software innovations capable of running on small devices making them accessible to the masses. The pressure to constantly improve these technologies is still on as users keep demanding for more improved solutions to ease their daily lives. As the world moves towards era of the Internet of Things (IoT), mobile devices and other smaller devices will be expected to run complex software systems efficiently. This is largely because users will expect these devices to become their dependable and long term cognitive partners. As such, finding ways of utilizing minimal resources to facilitate learning and recall is paramount if IoT is to be successful. Looking for inspirations from biology, AI researchers have discovered that the mammalian brain is able to work efficiently while using little resources compared to the ordinary computer models.
To understand what makes the mammalian brain tick, we turn to biology, psychology and related fields for guidance. The human brain being the command and control center of the human nervous system contains about 86 billion nerve cells (neurons) [1], [2]. These nerve cells link to each other through axons and dendrites which are believed to be the basic memory units. During stimulation, the neurons influence each other to either fire or not fire (transmit an electrochemical signal). This binary nature of how the neurons behave can be easily simulated in a computer model. The gist of the matter however lies on how memories are formed and recalled by the brain. Psychologists have for long held the view that humans learn and remember through associating often unrelated items through a process known as chunking. Recent empirical evidence from experiments involving observation of experts seem to support this claim.
Chunking is a kind of cognitive compression mechanism where the brain parses information into sub-components that are more memorable and easier to process than the seemingly random bits of which they are composed [3]. For instance, humans learn and recall long sequences of phone numbers by segmenting the sequence into small segments. For example, a phone number 0724942245 would be easier to remember when divided into 3 segments; 0724-942-245. Chunking is the hallmark of the brain’s organization [4]. Numerous psychological experiments have confirmed that experts in given domains accumulate their expertise through chunking. Chess masters acquire skills when constellations of pieces (segments of the game) become associated with moves or strategies and are stored in the long-term memory [5]. It rightly appears that the role of chunking is to move to a given goal through divide and conquer principle. This principle is vivid in most computer paradigms such as object-oriented programming where larger problems are broken down into sub-problems. These realities make chunking a practical approach when building a computer model as well as an effective strategy of improving learning in cognitive learning models.
Drawing from the inspiration of chunking in human cognition, this study focuses on simulating it on a computer model. The model will work with a Hopfield learning model to store, chunk, and filter associative memory.
This paper is structured as follows. In Section 2 the problem statement will be described in detail. In Section 3 the proposed AMR model will be introduced and explained. In Section 4 the role of the KID model will be presented in detail. In Section 5 the associations learning in the learning layer will be described. In Section 6 the chunking mechanisms will be explained in detail. In Section 7 some of the existing studies will be reviewed. In Section 8 an application scenario of the proposed system will be demonstrated. In Section 9 the experiment results will be discussed. Section 10 will conclude this paper.
Section snippets
Problem description
Currently a human activity can be recognized or inferred using different methods. Making use of human experts is one such method where a person analyzes entities or attributes that associate to form a human activity. For instance, to infer that cooking is taking place, ingredients preparation, lighting the stove and cleaning the dishes are associated entities that must be established. More human activities in a home can be inferred this way but some might require in depth knowledge to identify
Proposed human activity recognition system
The proposed human activity recognition system is based on the AMR model in conjunction with the knowledge–information–data (KID) model. It is assumed that a smart home is equipped with sensing devices which can acquire data as the five senses of human beings. Sensor data about human activities will be collected from a smart home and processed in the KID model [9]. Sensor data will be conceptualized as a formal concept of object by the KID model in conjunction with the learning layer in the AMR
The roles of cognitive learning of the KID model
The KID model was initially proposed by A. Sato and R. Huang [9]. It is a cognitive model that was originally designed for transforming data in retail stores into knowledge that can support decision making. The KID was built after a realization that analysis of big data involves using a lot of skilled man power to create meaningful assumption from the data as well as formulate hypotheses. As a cognitive model, the KID model carries the important cognitive functions of perception, learning, and
Associations learning in the learning layer
The learning layer is the initial layer of the AMR model which accepts the KID’s learning results. It is responsible for memorizing the internal relations of an activity from the KID model which is a 5-tuple. The internal relations of a formally represented activity is the Cartesian product between the activity’s entity and property sets [10]. The internal relations can be represented as a vector pattern. The learning layer then uses a Hopfield neural network [11] to memorize the pattern.
Chunking mechanisms
A chunk is a meaningful unit of information built from smaller pieces of information, and chunking is the process of creating a new chunk [12]. In [13], [14], de Groot developed chunking theory in his pursuit of understanding how experts and novices in the game of chess solve the same problem. He chose chess game since it was easy to measure expertise of players precisely. It was concluded that there are clear differences between the search patterns of experts and novices given the same chess
Existing studies
There have been different approaches for recognizing human activities in a smart home. In this section, we review process mining analytics [22], activity mining using COM [21] and activity mining using HMS [23]. The goal of these approaches is to use data from a smart home to deduce activities which is also one of the goals of the proposed AMR model.
Application scenario
A smart home application scenario is used to test the proposed AMR model. It is based on the experiment conducted by F. Javier [26] where two participants’ daily activities of living were observed for a period of 35 days in their houses. The data generated from this experiment, the “Activities of Daily Living (ADLs) dataset [26]” is publicly available. The established human activities were Leaving, Toileting, Showering, Sleeping, Breakfast, Lunch, Dinner, Snack, Spare time, TV, and Grooming.
Experiment results
Multiple iterations of experiment were conducted throughwhich it is possible to discover the evolution of associations among activities as entities and the computational complexity of the AMR model increase in comparison to other related models. A total of 393 (80%) activities in the dataset were fed into the AMR model.
Conclusions and future work
This study aims at creating a memory structure that mimics the internal working of human brains. To ensure memory items could be easily associated, they were formally expressed using formal rules in denotational mathematics. A formal activity consists of internal relations, pre-relations and post relations. With these kinds of relations defined in an activity, chunking mechanisms could be applied to group associated activities together.
Chunking allowed for packaging of related activities as a
Acknowledgment
This work is partially supported by the Japan Society for the Promotion of Science Grants-in-Aid for Scientific Research (No. 18K11408).
Runhe Huang is a full professor in the Faculty of Computer and Information Sciences at Hosei University, Japan. She received a Sino-Britain Friendship Scholarship for her study in U.K and received her Ph.D in Computer Science and Mathematics from University of the West of England in 1993. Her research fields include cognitive computing, brain modeling, computational intelligence computing, big data, machine learning. She is an IEEE senior member (SM’17).
References (32)
- et al.
Perception in chess
Cogn. Psychol.
(1973) - T. Lewis, S. Writer, ‘Human Brain: Facts, Functions & Anatomy’, March 2016.[Online]Available:...
- R.C. Mohs, ‘The Science of How Human Memory Works’. [Online]. Available:...
- M. Popova, ‘The Science of “Chunking,” Working Memory, and How Pattern Recognition Fuels Creativity’. [Online]....
- et al.
Learning of chunking sequences in cognition and behavior
PLoS Comput. Biol.
(2015) Computational Chunking in Chess
(2011)- et al.
Formal properties and rules of concept algebra
The universe of discourse
JSTOR
(1975)In search of denotational mathematics: Novel mathematics Means for contemporary intelligence, brain, and knowledge sciences
J. Adv. Math. Appl.
(2012)- A. Sato, R. Huang, From data to knowledge: A cognitive approach to retail business intelligence, in: IEEE International...
Concept algebra: A denotationalmathematics revisited for formal knowledgerepresentation and cognitive robot learning
J. Adv. Math. Appl.
Neural networks and physical systems with emergent collective computational abilities
Proc. Natl. Acad. Sci. USA
Chunking mechanisms and learning
Het Denken Van Den Schaker
Thought and Choice in Chess
Expert and “novice” problem solving strategies in chess: Sixty years of citing de Groot 1946
Think. Reason.
Cited by (5)
Editorial: Smart Cyber–Physical Systems: Toward Pervasive Intelligence systems
2020, Future Generation Computer SystemsFuzzy associative memories with autoencoding mechanisms
2020, Knowledge-Based SystemsCitation Excerpt :Various ways on how to improve the capabilities of the recall of associative memories have been a subject of many studies on associative memories. These studies focus mainly on conceptual, architectural, and algorithmic developments [1–33]. Two-level fuzzy associative memories [17–19] and multilevel fuzzy associative memories [20,34] have been presented and discussed in various applications.
The storage capacity of a directed graph and nodewise autonomous, ubiquitous learning
2023, Frontiers in Computational NeuroscienceQuantum Inspired High Dimensional Conceptual Space as KID Model for Elderly Assistance
2020, Advances in Intelligent Systems and ComputingA ship target location and mask generation algorithms base on mask RCNN
2019, International Journal of Computational Intelligence Systems
Runhe Huang is a full professor in the Faculty of Computer and Information Sciences at Hosei University, Japan. She received a Sino-Britain Friendship Scholarship for her study in U.K and received her Ph.D in Computer Science and Mathematics from University of the West of England in 1993. Her research fields include cognitive computing, brain modeling, computational intelligence computing, big data, machine learning. She is an IEEE senior member (SM’17).
Mr. Peter Kimani Mungai earned a degree in B.Sc. in Information Technology from Dedan Kimathi University of Technology, Kenya, in 2015 and is currently pursuing his Master of Science degree in Computer and Information Sciences at Hosei University Japan. Prior to enrolling for the master’s degree program, he worked as a web designer at Excia East Africa Ltd.
Jianhua Ma is a full professor in the Faculty of Computer and Information Sciences at Hosei University, Japan. He received the B.S. and M.S. degrees from National University of Defense Technology, China in 1982 and 1985, respectively, and the Ph.D. degree from Xidian University, China in 1990. His research interests include multimedia, networks, ubiquitous computing, social computing, and cyber intelligence. He is a founder chair of IEEE CIS Smart World Technical Committee (granted in 2016) and IEEE SMC Cybermatics Technical Committee (granted in 2016)
Dr. Kevin I-Kai Wang received the Bachelor of Engineering (Hons.) degree in Computer Systems Engineering and PhD degree in Electrical and Electronics Engineering from the Department of Electrical and Computer Engineering, the University of Auckland, New Zealand, in 2004 and 2009 respectively. He is currently a Senior Lecturer in the Department of Electrical and Computer Engineering, the University of Auckland. He was also a research engineer designing commercial home automation systems and traffic sensing systems from 2009 to 2011. His current research interests include wireless sensor network based ambient intelligence, pervasive healthcare systems, human activity recognition, behaviour data analytics and bio-cybernetic systems.