Learning machine learning with young children: exploring informal settings in an African context

ABSTRACT

Background and context

Researchers have been investigating ways to demystify machine learning for students from kindergarten to twelfth grade (K–12) levels. As little evidence can be found in the literature, there is a need for additional research to understand and facilitate the learning experience of children while also considering the African context.

Objective

The purpose of this study was to explore how young children teach and develop their understanding of machine learning based technologies in playful and informal settings.

Method

Using a qualitative methodological approach through fine-grained analysis of video recordings and interviews, we analysed how 18 children aged 3–13 years constructed their interactions with a machine-based technology (Google’s Teachable Machine).

Findings

This study provides empirical support for the claim that Google’s Teachable Machine contributes to the development of data literacy and conceptual understanding across K–12 irrespective of the learners’ backgrounds. The results also confirmed children’s ability to infer the relationship between their own expressions and the output of the machine learning-based tool, thus, identifying the input-output relationships in machine learning. In addition, this study opens a discussion around differentials in emerging technology use across different contexts through participatory learning.

Implications

The results provide a baseline for future research on the topic and preliminary evidence to discern how children learn about machine learning in the African K–12 context.

KEYWORDS:

• Machine learning
• data
• young children
• participatory learning
• informal settings
• Africa

Introduction

Machine learning (ML) has become integral to everyday life through the implementation of its processes into devices and services (e.g. Hitron et al., 2018). In recent times, scientific and public interest in introducing machine learning processes to young children has dramatically increased (Ho & Scadding, 2019; Narahara & Kobayashi, 2018; Sanusi & Oyelere, 2020; Sanusi, 2021b; Vartiainen, Tedre, et al., 2020). In addition, there is accumulating evidence that introducing machine learning concepts to children contributes to children acquiring a better understanding of the world around them (Hitron et al., 2019; Lin et al., 2020; Morris & Fiebrink, 2013; Scheidt & Pulver, 2019). However, the underlying mechanisms of machine learning and the technology behind it have been elusive in education until now. Consequently, there is a need for additional interdisciplinary human-computer interaction and education research investigating how machine learning literacy is currently implemented in learning experiences (Zhou et al., 2020).

Previous research (Sanusi, 2021b) has also pointed out that existing studies are Western-centric (specifically, the US, Europe and some Asian countries) and consideration should be given to research in an African context for inclusive discussion around machine learning in K–12 education. In the literature, inclusive artificial intelligence (e.g. Druga et al., 2019; Xia et al., 2022) has become a critical addition to K–12 education, which is still characterised by limited support for designing tools and curriculums to teach the emerging concept. Furthermore, K–12 learning contexts have unique needs, such as emphasis on engagement and scaffolding, and contextual challenges, such as tight school schedules, which require additional design considerations in order to be overcome (Zhou et al., 2020). The identified issues present a research gap, as there is a clear need to investigate young children’s exploration of machine learning technology, especially in a culturally diverse context. Cultural diversity is operationalised in this study as multiple ethnic groups, languages and dialects, among other differences. The expectation is that the experiences will suggest design ideas and materials and reveal the appropriateness of the tools for different contexts or cultural particularities.

The machine learning-based technology adopted in this study for exploration by the participants is Google’s Teachable Machine (GTM). GTM is a web-based application that facilitates the teaching and learning of basic machine learning principles (Carney et al., 2020). Earlier studies have utilised GTM in a Finnish context (Toivonen et al., 2020; Vartiainen, Tedre, et al., 2020). This study aims to introduce basic concepts of machine learning and further explore young children’s early encounters and insights into machine learning-based technology in Nigeria. In addition, this research presents empirical work and an analysis that illustrates how young children teach and interact with machine learning-based technology in informal settings. The contribution of this study to the body of knowledge stems from providing insights into pedagogical activities and an empirical exploration of how young children can explore and develop their understanding of ML-based technologies in playful and informal settings. In addition, this study replicates earlier research by Vartiainen, Tedre, et al. (2020), investigating how young children explored machine learning concepts using GTM in a Finnish context. The findings of Al-Zubidy et al. (2016) suggest that the lack of replication in the computing education domain is inhibiting the maturation of the field. Randolph and Bednarik (2008) also assert that there is a need for replication studies to identify potentially non-generalisable results. In general, we explored a different situation, context and subjects to determine whether the primary findings of the original study in the Finnish context could be applied to subjects in a different cultural setting, in this case, Nigeria.

The ongoing technological revolution necessitates that all continents and countries of the world are involved in its development, which requires contextual particularities to be considered to ensure relevance across contexts (Gwagwa et al., 2021; Sanusi, Oyelere, et al., 2022). In keeping with this, ML education in K–12 that currently generates interest in developed context needs to be promoted among youths in Africa to allow them to keep pace with the technological advancement and be equipped as active creators and designers of future AI and ML technologies. Even though Nigeria has a centre dedicated to AI and other emerging technologies, known as the Centre for Artificial Intelligence and Robotics, specifically for the transformation of the country’s digital economy, there is no current initiative for youths and AI at compulsory education level. However, efforts to equip the younger generation with AI knowledge have begun through data camps and virtual programmes such as Data Science Nigeria and the work of researchers (Sanusi, Olaleye, et al., 2022).

Although the implementation of computer studies (CS) is plagued with several challenges in Nigerian schools, it is a compulsory subject in the basic educational system (Oyelere et al., 2016). This provides a basis on which the discussion of machine learning literacy can be built, as ML concepts could be introduced at various levels in the K–12 settings. The results obtained from this study add nuance to the existing literature. First, this study provides a foundation for studies on the development of ML conceptual understanding among children and youths in an African setting. It also provides a deeper understanding of how children of different educational and cultural backgrounds interact with machine learning. Second, this study contributes to the corresponding research design and implementation, that is, interactions between siblings, friends and parents/guardians, opening up interesting possibilities for family learning. This approach proposes a powerful way to explore novel topics such as ML in informal contexts. In addition, this paper is proof that children can develop a basic understanding of ML ideas, irrespective of their location and regardless of the limitations and issues around ICT/CS education in the country.

Based on the above information, our understanding is that Nigerian children may have limited access to data-intensive technologies, artificial intelligence (AI), and ML tools. We assume this position based on limited access to the new technologies caused by the digital exclusion that is predominant in Nigeria due to social inequalities. Consequently, this study attempts to illustrate how machine learning could be taught to children outside educational settings using a participatory learning approach and positioning the children’s activities concerning technology use in a different context. The participatory learning approach is relevant to this study as it emphasises children’s active contribution to learning and shared meaning-making and endeavours (Hedges & Cullen, 2012). In addition, this study is arguably one of the first studies concerning the exposure of young children to the unfolding process of machine learning and teaching machine learning in K–12 settings in Africa.

This study analysed how 18 young children between the ages of 3–13 years constructed their interaction with machine learning using GTM and observed how the responses of GTM moulded their actions and engagements during the exploration. Teaching machine learning concepts through GTM was conducted with the support of the children’s guardians in informal, non-educational settings. The following research questions, which were adapted from a previous study (Vartiainen, Tedre, et al., 2020), are addressed in this study:

1. What do the children teach the machine?

2. What sorts of interactivity procedures are evident when the children teach the machine?

3. What are the children’s views and experiences during the process of teaching the machine?

The paper is structured as follows: Section 2 defines the theoretical framework, while Section 3 discusses the methodology and intervention process (participants, contextual descriptions, ethics, and analysis). Section 4 reports the qualitative data, and Section 5 describes the results. Section 6 discusses the findings and addresses the key elements emerging from the experience in two contexts (Nigeria and Finland). The final section proposes a roadmap for future work.

Context and pedagogical framework

Basic education and computer science in Nigeria

In Nigeria, as in any other part of the world, basic education provides the bedrock upon which other comprehensive school levels rest. The Universal Basic Education (UBE) Programme was introduced in 1999 by the Federal Government of Nigeria as a reform programme to provide greater access to and ensure the quality of basic education throughout Nigeria (UBEC, 2020). Despite the officially free and compulsory basic education, one in every five of the world’s out-of-school children resides in Nigeria (UNICEF Nigeria, 2020). According to Suleiman et al. (2019), the regional synopsis by Education for All (EFA) reports education in Nigeria is at serious risk of not achieving the goal of universal basic education. A recent report from the World Bank also revealed that Nigerian primary and secondary school children attend school but fail to learn (Suleiman et al., 2019). The report confirms the assertion of UNICEF (2021) that schooling does not always lead to learning and that there are more non-learners in school than out of school worldwide. Ololube et al. (2016) predict that if the trend of schooling without learning continues in Nigeria, millions of young Nigerian students will be faced with the prospect of lost opportunities and lower wages in the latter parts of their lives.

Computer science has been integrated into the basic educational level of Nigeria in recognition of its globally proven importance (Egede & Asabor, 2019). A myriad of studies (e.g. Aworanti, 2016; Danner & Pessu, 2013; Isaac et al., 2018; Ogundile et al., 2019) have shown that the teaching and learning of computer science is still very low in Nigeria. This was further corroborated by the Federal Ministry of Education (FME), which stated that the implementation of computer science education is plagued with many challenges, including policy, institutional and administrative capacities, curriculum, equity issues and funding (Federal Ministry of Education, 2019). The need for a standardised and coordinated deployment of computer science education in Nigeria informed the development of a National Policy on ICT Education by the FME. This policy document previously developed in 2010 and revised in 2019 provides the necessary guidance for all stakeholders’ expectations in the whole process of ICT education in the country. While the policy statement recognises that the state of ICT education in Nigeria falls below global standards, this document reinforces the need for focused intervention in ICT education.

This study focuses on samples in informal settings. Informal settings were utilised since ML-related concepts have not been introduced into mainstream education in Nigerian schools. In addition, resources (unstable and interrupted internet connection and erratic power supply) may hinder conducting such activities in the school environment. Moreover, as a new concept for schools, exploring another setting other than the school environment might reveal approaches that could be utilised to effectively learn ML. The children are drawn from private and public schools in order to observe and gather their perspectives on the machine learning-based technology tool explored in this research. This study may help to initiate a discussion around differentials in emerging technology use across cultures and backgrounds and create an understanding of the factors responsible for educational impetus.

Teaching machine learning in informal settings

Informal learning settings depart from formal and non-formal learning in that the learning is rarely organised or guided by a programme or curriculum (Ainsworth & Eaton, 2010; Dubovi & Tabak, 2020). Rather than being guided by a rigid curriculum, it is often considered to be experiential and spontaneous. Irrespective of the learning settings, that is, formal (e.g. school), non-formal (e.g. museum) and informal (e.g. home), all learning is valuable and contributes to an individual`s cognitive, emotional and social growth (Ainsworth & Eaton, 2010). Dierking et al. (2003) and Phamduy et al. (2015) emphasised that informal science education is critical for improving scientific literacy through its spontaneous nature and practically limitless opportunities.

ML as an emerging research area in K–12 education has mostly been introduced through workshops or as an after-school programme (Mahipal et al., 2023 Ma et al., 2023). A few studies have shown that ML has been taught as part of a course in the classroom (Sperling & Lickerman, 2012; Burgsteiner et al., 2016). Although Long et al. (2021) promote AI literacy in museum-like settings to examine learning ML in such unrestricted situations, we only identified *one study (Vartiainen, Toivonen, et al., 2020) that explored the ML process with children at home. While these studies provide a baseline for understanding learning about ML in informal settings, more studies are needed to validate these findings and explore other contexts. Investigating informal context is important because learning is a cumulative process involving connections and reinforcement among the variety of learning experiences people encounter in their lives, such as at home, during schooling and out in the community (Dierking et al., 2003). Exploring different learning settings will provide information about effective approaches to introducing and demystifying ML in K–12 education.

Participatory learning in children’s education

Our approach to exploring how children of different educational and cultural backgrounds interact with machine learning builds on participatory learning. Participatory learning draws from sociocultural and cultural-historical theorising of learning and participation that originate from the intellectual work of Lev Vygotsky (1978) and his followers. In general, participatory learning emphasises that children develop through participation in the everyday activities and practices of communities which change historically and culturally (Hedges & Cullen, 2012; Rogoff, 1990). Moreover, participatory learning emphasises that all learning and knowledge construction is situated (Hedges & Cullen, 2012), that is, an individual’s learning is bound to a cultural context (Sfard, 1998) and mediated by different tools and affordances (Schoultz et al., 2001). In addition, this approach aims to promote equity of participation (Kafai et al., 2020) and a culturally grounded pedagogy that is responsive to students’ identities, values, backgrounds and cultural resources that affect their interest and engagement in learning (Hedges & Cullen, 2012).

Previous work with participatory pedagogy further highlighted positioning children as active agents to provide them with the opportunity to take ownership of their own learning (Kumpulainen & Lipponen, 2012). As children’s interests reflect what they experience in their everyday lives, educators then have to harness children’s own perspectives and cultural resources when designing learning activities and environments that connect with the everyday practices in which they are active (Bulunuz, 2013). Participatory approaches further stress the role of the tools and affordances that are within the reach of children in the “zone of proximal development” (Vygotsky, 1978). Through the mediation of people and cultural tools in settings that promote learning, a child may be able to solve problems or complete tasks that would otherwise be too difficult to accomplish (Vygotsky, 1978). According to Hedges and Cullen (2012), collaborative activities supported by cultural artifacts and tools facilitate inquiry, knowledge construction and the appropriation of complex ideas and cognitive processes. Through scaffolded activities with more competent peers and adults, children may also gain understanding of the use of different tools and technologies and gradually incorporate them into the repertoire of their own learning (Wertsch, 1998). Edwards (2005) pointed out that participatory learning results in the development of “the capacity to interpret the complexity of the world and have the wherewithal to respond to that complexity” (p. 60). Hedges and Cullen (2012) further stressed that meaningful knowledge building occurs in the context of self-motivated participation in authentic activities in which a child may also construct an identity as a competent learner.

Learning by teaching a computer

The research conducted by Papert (1980), Kafai and Harel (1991), and Ackermann (2004) shows that one of the ways children can learn is by teaching a computer. Teaching a computer can be achieved when children engage the computer in designing, creating and programming artifacts. One innovative work of educational vision that utilised LOGO programming (Papert, 1980) asserted that children’s thinking ability was greatly enhanced when they had to teach a computer based on the guiding principle of Turtle Geometry. In this principle, the children controlled the LOGO turtle using words since the turtle embodied an external representation of the children’s thinking. Papert emphasised that children often aligned their internal mental model with the external representation when they taught and gave directions to the computer.

By teaching the computer this way, the children were able to grasp vital information they would not otherwise have learned through conventional means. The children’s ability to express their thoughts and intentions to the computer for proper execution was also essential (Papert, 1980). The mediated action of Papert’s work, according to Ackermann (2004), tends to align with the learning of sociocultural theories. Hence, highlighting that humans typically use tools to aid reasoning implies that those children will be obliged to understand some concepts in a personified form (Papert, 1980). According to Michelle et al. (2020), when children teach the computer by providing captivating examples, which allow other people to have fun with the techniques, they also add to their knowledge of how ML works behind the scenes. This means that when we create ML models, we typically provide a platform that enables an individual to ask questions regarding the inputs and outputs of any model with which they come into contact as they go about their daily activities.

Programming teachable machines

The advent of teachable machines has seen a significant rise in the number of people who have been empowered to create ML models. In addition, children do not need to communicate with machines via code but rather by using pattern recognition and natural language technologies since AI is now fully accommodated in today’s world (Druga et al., 2019). According to Druga et al. (2019), when working with ML technologies, knowledge of programme syntax is unnecessary to foster children’s participatory learning. They emphasised that rather than engaging in rule-based programming or deductive reasoning, which have been the significant drivers of earlier programming paradigms, we now engage children with data-driven programming to provide training data sets to the machines. From there, machine models are created, which are then used to control the machines.

For instance, in this research, we engaged the modelling capacity of GTM in creating machine learning models. In addition, GTM, with its user-friendly and rich interface, enabled the children to play around with the creation of machine learning models in an intuitive and educative way, thereby paving the way for these youngsters to learn how machine learning works. GTM tool is an entirely web-based application that abstracts the details of the machine learning algorithm from the user. Hence, the user can create ML models without writing a single programming code (Toivonen et al., 2020; Vartiainen, Tedre, et al., 2020). For instance, in this work, the children used their web camera to provide the training data (images) to the computer, thus, enabling the image tool to teach the ML model how to classify images.

Furthermore, the children also taught the teachable machine how to classify their facial expressions and recognise their voices without writing any programmabl2016e code. They did this with the aid of GTM predictive tools, which are used to train predictive models. Instead of writing codes that drive programming language experiments in education, allowing children to utilise well-designed ML tools will introduce them at an early stage to the act of teaching a computer. In the process, they may also learn how ML algorithms work by providing training datasets and then controlling the machine using the trained model. Overall, these novel paradigms of computational approaches and ML tools expand the possibilities and new methods available for children and provide a novel approach to exploring and meaningfully impacting on our present ecosystem (Tedre et al., 2020).

Research design

In this study, the ideas and views of young children were considered to foster inclusive and equitable education in machine learning (Druga et al., 2019). The participatory research philosophy in this study was to empower younger children and introduce them to machine learning at an early stage to support them in subsequently making significant changes to the technology around them. We examined and explored the various encounters these younger children had with machine learning-based technology to achieve this. Specifically, we looked at the multiple ways these children were able to teach machine learning-based technology. Eighteen children aged 3–13 years participated in this study. We analysed how these children constructed their interactions with the machine-based technology with the help of guardian(s). Furthermore, we gathered the machine outputs from the children’s inputs and how they shaped the children’s actions and interactions. The groupings of the children are shown in Table 1.

Table 1. The teaching session and the video data collected.

Settings	Participants (year old)	Groupings	Collected data from the use of the ML tool (min., including interview)
Location A
Peer teaching a computer with the support of a parent or familiar adult	Samal (5y) and Falil (3y) with their mother	Face detection	20:22
	Akin (8y), Grace (5y) and Tolu (7y) with a familiar adult	Face recognition	25:59
	Evelyn (8y), Babajide (9y), Salewa (12y) and familiar adult	Voice recognition	32:30
Location B
Peer teaching a computer with the support of a parent or familiar adult	Samuel (6y), Vincent (5y), Gloria (5y), Enemona (4y)	Face detection	35:49
	Treasure (13y), Deborah (10y),	Face recognition	18:50
	Jessica (9y), Victoria (8y), Goodluck (7y), and Joel (7y)	Voice recognition	29:00

Names have been changed to ensure anonymity.

While we adopted the research design approach used by Vartiainen, Tedre, et al. (2020), a qualitative methodology approach was utilised in this study to glean information from the participants. This approach allows researchers to study things in their natural settings, attempting to make sense of, or interpret, phenomena in terms of the meanings people bring to them (Aspers & Corte, 2019).

Participants

Eighteen children took part in the experiment in two different locations in Nigeria as shown in Table 1. The participants include eight children from Southern Nigeria (location A) and ten children from Northern (location B) Nigeria. A common opinion in Nigeria is that the populace of location A are more educated and academically inclined than their counterparts in location B. This opinion is further corroborated by UNICEF (2021) statistics, which show that Nigeria is home to the largest number of out-of-school children in the world. Location B is primarily affected in this regard, owing to displacements and emergencies over the fear of the attack and abduction of pupils and members of staff. Both locations were chosen in an attempt to address the equity issue recognised by FME (2019), which revealed that a significant dichotomy exists between school locations and between public and private schools in ICT implementation in Nigerian education. While more scientific reports are required to address and explore the educational imbalance between children in both locations as well as literacy levels across the divide, this study selected samples in both locations. Brief descriptions of the participants are provided.

The eight children from location A ranged in age from 3–12 years old. The children were told to describe themselves in their own words including their names. Names were later changed while reporting to ensure anonymity. The youngest participant was Falil, a three-year-old boy who likes eating amala (yam flour) and playing football. He wants to be like his father when he grows up. Samal, a five-year-old boy and a big brother to Falil, likes football and watching cartoon movies. He wants to drive a car like his grandfather when he grows up. Grace is a six-year-old girl who loves dancing and eating. She wants to be a dancer in the future. Tolu is Grace’s best friend; Tolu is a seven-year-old girl who likes to eat bread and play. In the future, she wants to plait hair. Akin, an eight-year-old boy, likes cats. He likes to play football with his friends and wants to be a footballer in the future. Babajide is a ten-year-old boy who says that he likes to draw objects and images in addition to watching cartoons. He sees himself as an artist in the future. Evelyn, a classmate of Babajide, is a nine-year-old girl who loves singing and going to church. She wants to be a teacher in the future. Finally, Salewa, Akin’s elder sister, is a twelve-year-old girl who likes reading and sees herself working as an engineer in the future.

There were ten participants aged 4–13 years from location B. The oldest participant amongst them was Treasure, a thirteen-year-old female. According to her, singing and dancing is her preferred hobby. When she grows up, she would like to be a doctor. Deborah is a ten-year-old girl who loves helping her mother with fetching water and other domestic work. According to her, she would like to become a lawyer when she grows up. Jessica is a nine-year-old girl who loves pink colours; she prefers eating food without fish. Her future ambition is to become an artist. Another participant who likes to clap hands anytime she sings is Victoria. Victoria, an eight-year-old girl, who often wants to be addressed as Vicky, likes playing ludo games and would like to become a teacher when she grows up. Goodluck is a seven-year-old boy who prefers playing video games and watching flying objects in the sky. He wants to become a pilot in the future. Joel is a seven-year-old boy who loves eating. According to him, he wants to be like his father when he grows up. Samuel is a six-year-old boy who likes drawing objects; he would like to become a teacher in the future. Gloria, who likes dancing, is a five-year-old girl whose future ambition is to be a dancer. Vincent, who is also five years old, loves going to the swimming pool to swim for fun. He says he would love to be a soldier when he grows up. The youngest participant in location B is a four-year-old boy named Enemona who loves eating noodles mixed with eggs. According to him, he wants to be like his mother when he grows up.

Research ethics

Since young children were the target group of this research, the researchers were aware of the need to take appropriate precautions regarding ethical issues. The participants were recruited through contacts of family and friends without considering whether they had prior experience of computational concepts or the use of technology in any form. After informing the guardians and parents about the aim of the research, they gave their informed consent to conduct the study. At the same time, we only proceeded with the experiment based on the parents’ willingness and children’s interest in being involved. We followed the ethical guidelines of research in the humanities and the social and behavioural sciences, provided by the Finnish Advisory Board on Research Integrity (2009), including parental approval for pictures and videos. The participants were informed that the experiment was solely for academic and research purposes and that they could withdraw their participation at any time during the intervention. The experiment took place in the children’s homes in a quiet, dedicated room. We also sought guidance from the National Health Research Ethics Committee of Nigeria (NHREC, 2016) since working with children requires particular attention to research ethics. In order to avoid revealing the children’s personal data when using the commercial tool (GTM), we anonymised identifiable information and also deleted all the images and sounds of the children captured by the tool after the exploration process.

Data collection

Following the existing method (Vartiainen, Tedre, et al., 2020), the data collection was performed by the parents or a familiar adult and took place at the children’s homes in location A. This was also the case in location B, but all the children converged in one parent’s house owing to limited resources (e.g. computer and internet). A guardian of one of the children facilitated the teaching and exploration process. The children also taught themselves the use of GTM which promote active learning. The experiment was carried out using GTM, while the children’s interaction with GTM was video recorded using a smartphone. Prior to the experiment, participating adults were given a very short introduction to GTM by the researchers with a brief explanation of what was required of them throughout the intervention. We started by showing them a five-minute video (Google, 2020) on how to use the tool and at the same time, explained the idea and function of the tool. We mentioned to them that GTM makes creating ML models fast, easy, and accessible to everyone and the focus of our activity is Classification, a supervised learning technique which is a subcategory of ML. We explained that a classification algorithm is used to identify the category of new observations on the basis of training data. We further illustrate that ML can learn to predict the category data belongs to, as humans do when they need to put things in categories.

The children were divided into groups based on three variables, that is, face detection, emotions (facial expression) and voice recognition. The participants were divided into three groups primarily based on their age or development level, especially in location B. Location A participants were mostly siblings of the same parent and their friends, and the session was conducted in their homes. These groupings were used to explore two project types on GTM (Image and Audio Project). In location A, two children explored the learning tool that detected their faces, three focused on facial expression and three used voice recognition. In location B, one group of four children provided samples of their faces. Another group of four provided examples of their facial expression-based emotions (happy and sad). Finally, one group of two children provided samples of their voices to GTM. Beginning with the face detection group, the guardian facilitated each child teaching the computer how to detect their face. This was achieved by presenting samples of their faces to GTM with the aid of the host computer’s webcam. Then, to train the model, each child was introduced to the model to allow the machine to detect their face.

The next group was the emotions based on facial recognition. In this case, the guardian helped each child to supply samples of their emotions (happy, sad) to the machine using the computer’s webcam. Each child trained their model. The output showed that the machine could identify the various emotions and the particular child who displayed the emotions. The last category was voice recognition. In this group, the guardian assisted the children by teaching the machine how to recognise their voices. This was accomplished by supplying samples of voice recordings to the machine using the computer’s voice recorder application. Prior to this, some background noise was recorded to neutralise interference between the child’s voice and the background voice. The output showed that the machine was able to learn each child’s voice by recognising the voice and calling the child’s name.

Overall, the children were able to explore the input-output relationship with GTM. The children were able to make their observations and interpretations of GTM output to different inputs. After the entire process was completed, the guardian interviewed each category of children about what transpired during the teaching and exploration of the machine. The guardian also supported the children in their expression and narration of events. Furthermore, the guardian posed several questions to the children in each category, such as, what did you teach the computer? How did the computer learn? What else could you introduce to the computer? The children provided answers to all these questions, which were recorded.

Data analysis

First, all the video and audio data were thoroughly scrutinised to gain a general impression and sense of the children’s interaction with the technology. Before analysing the data, we first divided the data into three phases (Vartiainen, Tedre, et al., 2020): teaching, exploration and explanation. The teaching phase entailed the children naming the emotions and other content to be taught to a computer, and then they created the test data and trained the machine learning model. In the exploration phase, the children actively explored the model with the completed test data. Finally, in the explanation phase, the children reflected and explained what transpired before, during, and after the process of interacting with GTM. The data analysis relied on qualitative content analysis (Elo & Kyngäs, 2008). In particular, the analysis followed a deductive analysis approach, in which a coding scheme was developed based on previous empirical research (Vartiainen, Tedre, et al., 2020). As shown in Figure 1, the three resulting categories and subcategories of the interaction timeline of the subjects (children and guardians), the teaching activity object (e.g. emotion) and the tool used (GTM) served as the basis for tables 3, 6, 7, 8, 9 and 10. The timeline provided the information regarding how interactions emerge as well as the duration, verbal and nonverbal communications and their relationship to the computer’s outputs.

Figure 1. Coding and definitions (Adapted from Vartiainen, Tedre, et al., 2020).

While the first author was responsible for conducting the analysis, the results of the analysis were negotiated by three researchers to ensure the reliability of the study. As including a positionality statement is important, especially in qualitative research, to understand the researchers’ relationships to the context and data, the first, second and third authors are Nigerian-born scholars who lived in and had experience of the context being studied. We replicated a Finnish study to understand whether the primary findings of the original study could be applied to Nigerian children with a different cultural setting and particularities. However, we modified this study to consider two different regions in the country as well as the data collection process across locations inspired by the context of the study.

Results

As highlighted in the data collection subsection, the children were divided into three groups. In the first group, that is, face detection, the children all wanted to teach the computer how to recognise their faces. In contrast, the second group of children (facial recognition-based emotions), all wanted to teach the computer how to identify whether someone is happy or sad. Finally, in the last category (voice recognition), the children all wanted to teach the computer how to recognise their voices. Based on the children’s exploration with the GTM tool, we infer they were introduced to ideas of classification and, by extension, concepts, and the process of supervised learning. We deduce that the children understood some of these ML concepts and practices since they can describe the input and output of a classification model and explain how ML classifies data by recognizing patterns. Below is a detailed explanation of the procedure.

Face detection group (3–5-year-olds) in location A

The first example of fine-grained process analysis derives from the youngest set of participants, 3-year-old Falil and his 5-year-old big brother, Samal. GTM was introduced to both children by their mother for approximately 10 minutes and concluded with detailing how to create a training data set to train a classifier. During the ten-minute introduction to the tool by their mother, she explained the idea and function of the tool to them. She started by a live demonstration of GTM using the image project to show that a computer can recognize her face and a stuffed animal held up to it differently. The lesson began with gathering data and grouping image samples (her face and stuffed animal) into categories she wants the computer to learn. Before she proceeds to train the model, Samal asks why she had to group the image samples into different classes. She responded that there will be no alternate sample to which the trained embeddings will be compared when testing the model. The following dialogue ensued between the parent and children during the tutorial:

Samal: Mom, why do you create more than one class?

Mother: Each of the classes represents a different category we want to teach the computer to recognize. So, if we add data to only one class, we won’t be able to ascertain if the computer recognizes my face and the stuffed animal differently.

Samal: Okay! Can we have three classes?

Mother: Yes, we can have three or more classes.

As the tutorial progresses, Samal was also interested in why the sample images needed to be trained. As shown in the parent-child conversation below, Samal was informed that in the process of training, the ML model finds patterns and makes predictions.

Samal: Why are you training the samples?

Mother: The dataset needs to be trained because it teaches the ML models to identify desired patterns. After the training, we can then test how the model correctly classifies new examples.

After training the model, their mother instantly tests it out to see whether it correctly classifies her face and the stuffed animal. GTM successfully recognized her face and the stuffed animal with the highest confidence ratings. The children were eager to explore the web-based tool after the tutorial and the accompanying conversations. Falil and Samal chose to teach the computer to detect their faces after they were asked whether they would like to teach a computer to recognise their faces and communicate who is on the camera. Samal, filled with excitement, commenced the exploration process and repeatedly clicked on the computer mouse cursor to record his face. Even without instruction, Samal kept making different facial and bodily expressions as he tried to teach the computer to recognise him. After supplying a variety of samples, he trained the model and was amazed that the computer could detect his face. With considerable excitement, he rushed to tell his mother that the computer could recognise his face. At that point, Falil showed up in front of the computer, but his face could not be recognised by the computer.

Interestingly, after several attempts, Samal informed Falil that the kind of inputs he made must have generated the type of outputs that he got and directed Falil to supply a sample of his face to the computer. As he explored the learning tool, Samal supported and guided his brother through the process since their mother withdrew into the background a few minutes into the intervention session. Falil took the cue from his brother and began supplying different facial expression samples to the system. A large amount of data was generated and trained in a few minutes, which resulted in the computer recognising the expressions on his face. The children expressed their joy at their discovery – that they could teach a computer. This led them to explore dancing and holding various objects to further test teaching them to the computer. Figure 2 illustrates the emergence and nature of the child-initiated exploration of GTM. The interaction results show how the children built and tested an ML-based face detection classifier with their facial and bodily expressions. It can be concluded that the ML tool, based on its easy-to-use features, enabled collaborative processes among the children and allowed exploration as well as familiarisation with machine learning-based technologies. Figure 3 depicts the exploration process among the children.

Figure 2. Co-teaching a teachable machine.

Figure 3. Children exploring GTM.

Face detection group (4–6-year-olds) in location B

Our fine-grained process analysis began in this group with four-year-old Enemona. Enemona was extremely curious about teaching the computer how to recognise his face. At first, he could not believe the computer could learn to recognise his face until the guardian properly guided him. The guardian instructed him to supply samples of his face to GTM using the computer’s webcam. At this point, Enemona positioned his face in front of the webcam and produced different samples of his face as guided by the guardian. The next participant was five-year Vincent, who also supplied samples of his face to the computer using the same medium as Enemona. Vincent was followed by Gloria, who also provided examples of her face to the machine. Finally, Samuel, a six-year-old boy, also supplied a sample of his face to the machine to investigate how the machine could learn their faces. The oldest participant in this group served as the group leader; thus, the leader facilitated the model’s training with the other participants to help the machine to learn how to recognise their faces. When the training process was completed, each of the participants presented their face to the machine for detection. At this point, the machine was able to detect the face of each of them by showing the percentage level of accuracy in face matching. It is worth mentioning that the machine was able to detect the faces of all four children. There was a scenario in which a child presented his face to GTM for detection, and the machine was unable to detect this child because the child failed to supply samples of his face to the machine. The children’s teaching and exploration phase is shown in Figures 4 and 5. Figure 6 illustrates the emergence and nature of the child-initiated exploration of GTM.

Figure 4. The first child teaches the machine a facial expression of sadness while the second child laughs at her sad look.

Figure 5. Children teaching the machine the facial expression of happiness by taking turns.

Figure 6. Co-teaching a teachable machine.

Facial recognition group (6–8-year-olds) in location A

The next set of children who explored GTM were 6-year-old Grace, her best friend, 7-year-old Tolu, and 8-year-old Akin. This group illustrates how the three children collaboratively interface with GTM. Within approximately 5 minutes, an adult familiar to the children introduced the teaching process, how to create a training data set and how to train the completed model. Since the focus was to teach the computer facial expressions of happiness through “smiling or laughing” and anger using “a sad-looking face”, Akin, the oldest of the children, started the process. Akin had several non-verbal interactions with the learning tool expressing various signs of happiness with laughter or a smile. Several emotions depicting anger were also supplied to the system by the self-directed Akin exploring the tool. Sometime later Akin confirmed he could teach computers by demonstrating emotions through facial expressions, although he encountered a challenge as one of his inputs could not be recognised. Tolu then commenced her exploration process. Tolu’s GTM exploration and interaction process was expeditious because she had learned from Akin’s experience. The youngest in the group, Grace, took her cue from Tolu and created a chain of quick input-output interactions exploring how the machine responded to her multi-modal input.

Each of the children reported an issue with the machine being unable to recognise their facial expressions during different attempts. In their attempts to solve the puzzle or unravel the challenge the children encountered at different times, they made several efforts at inputting various samples of “happiness” and “sadness” expressions. Tolu finally screamed, “I got it”. During the discussion among the children, they concluded that supplying insufficient samples of their emotions to the tool must be why the machine failed to recognise some of their attempts. This scenario shows the ensuing effect of a collaborative effort, as it leads to discovery and, in this case, exploration of the machine learning concept. Figure 7 depicts the unfolding process of engaging with the learning tool in a supportive social setting.

Figure 7. Learning the limits of the classifier that Akin trained.

Facial recognition group (7–9-year-olds) in location B

The following example shows how Joel, a 7-year-old boy, Goodluck, also a 7-year-old boy, Victoria, an 8-year-old girl, and Jessica, a 9-year-old girl, explored the teachable machine. First, the guardian gave a brief talk and demonstrated how to create a training data set and how to train the developed model to the leader of this group, Jessica (the oldest participant). The first to try this technology in this group was Jessica, who supplied her sample data, that is, facial expressions of happiness-based emotions using the computer webcam. Various expressions of happiness were displayed by Jessica. This was followed by sample data of sadness, which was also supplied to the machine in varying styles. Sample data representing happiness and sadness were then collected from Joel, Goodluck and Victoria with the help of Jessica. The leader of this group guided her peers in how to train the created model and how to name these emotions “happy” or “sad”. At this point when a peer is teaching peers, the guardian usually withdraws into the background, and the leader assumes the role of a peer teacher. After the model was trained, each of the children presented their various emotions to GTM to explore the input-output relationship. The machine was able to correctly identify the happy or sad person to the cheers of the children. At one point during the process, the machine was unable to recognise one of Jessica’s facial expression inputs of sadness. According to the guardian, this was because no such input was trained by the model, which is why he encouraged the children to provide varying samples of “sadness” to the machine so as not to confuse it with “happiness”. Figure 8 depicts the unfolding process of engaging with the learning tool in a supportive social setting.

Figure 8. Learning the limits of the classifier that Victoria trained.

Voice recognition group (9–12-year-olds) in location A

The next set of children introduced to the process of teaching a computer were 9-year-old Evelyn, 10-year-old Babajide and 12-year-old Salewa. Their encounter with the learning tool further illustrates how the children explored GTM collaboratively. Before the commencement of the exploration process, Evelyn’s mother, who is familiar with the other children, gave a ten-minute introduction explaining the idea and function of the tool to the children and showing them a demonstration video. Evelyn’s mother suggested the children begin with voice recognition as they were eager to explore the seemingly new tool. With Evelyn initiating the learning process – teaching the machine to recognise her voice – she speaks with different voices and sounds. Next to interact with the machine was Babajide. Babajide was quick to understand the general idea of the teachable machine, as he was able to detect why the classifier he trained sometimes failed to recognise his input. Salewa took a cue from Babajide and trained the model with a beaded gourd. She was excited as GTM stated with 100% confidence that it had identified the sound was from the beaded gourd. Figure 9 illustrates the co-teaching process.

Figure 9. Co-teaching a computer.

Voice recognition group (10–13-year-olds) in location B

The final example in this category involves two participants, Deborah, a 10-year-old girl, and Treasure, a 13-year-old girl. The guardian introduced GTM to this set of children. After this brief introduction, the children were excited to teach the computer their voices. Deborah was the first to teach the computer how to recognise her voice. She did this by activating GTM on her own before labelling her voice on the machine, after which she recorded some background vocals. Then her voice was recorded using the computer’s voice recorder application. The guardian emphasised the need for them to continue making sound until they exceeded the minimum threshold set out by GTM. This task was replicated by Treasure, who provided samples of her voice to the machine. After they had both succeeded in creating the training model, they trained the model to recognise which voice was Deborah and which voice was Treasure. After the training was completed, they both explored the input-output relationship of GTM, and to their amazement, when either of them made sounds toward the machine, the machine recognises the sound by showing a voice matching accuracy level of 100%. Figure 10 illustrates the co-teaching process.

Figure 10. Co-teaching a computer.

Prototypical trajectories

As our grouping results may illustrate, interpreting the children’s interactions with the tools alone may not provide a complete understanding of how the process relates to the real-world measure. An important aid towards understanding the multi-age groupings across age brackets is to find the prototypical trajectories of the students through the timeline representations in Figures 2, 6 , 7 , 8 , 9 and 10. This will reveal how the explorations and interactions with the teachable interface relate to the children’s age or previous experience. The figures also show how the guardians were supporting the children in teaching the computer before the exploration began. The times indicated in the figures shows the duration of the teaching or the exploration after the guardian had briefed the children.

In Figures 2 and 6, the interaction was mostly non-verbal by nature. The children explored the interface by showing their faces to the system, cheered themselves and tried out other things. They were actively exploring GTM within a few minutes. The figures show that very young children approached the environment playfully with body language. This supports the finding that guided play can be used for teaching preschool children (Weisberg et al., 2015). Figures 7 and 8 show active exploration with minimal guidance from adults. The students were able to unravel the issues encountered while exploring GTM tool. They collaboratively discovered the input versus output process that constitutes an ML project. Figures 9 and 10 show active exploration with almost no guidance from adults. The output of their explorative process inspired the other children in the cohort to provide their input. They quickly understood the general idea of how GTM was able to detect input and why the trained classifier sometimes failed to recognise input data.

While all the children regardless of their age group were able to explore the learning tool, the participants’ backgrounds, such as the expectations of the children in those cohorts, may influence the learning processes during the exploration phase. An example could be how children within the ages of 9–12 years display an understanding of training data and the input versus output relationship more rapidly than 3–6-year-old children. These and other variables related to interaction time, the type of conversation with the guardian, siblings’ presence and facilitation could be important factors that contribute to learning. Overall, the groupings are meaningful as they relate to other real-world measures.

Children’s teaching experiences and explanations

This section reports the experiences of the children during the teaching and learning process. The exploration of machine learning basics utilising GTM as a point of departure has provided some initial findings and pedagogical insights for future research and development, especially in K–12 education. This is evidenced in the self-reported experiences gathered from the children after the intervention and exposure to some of the basics of machine learning. Some of the children’s experiences with the tool are selected and detailed as excerpts below. Some excerpts have been chosen because they are representative of the similar experiences of the children.

The input-output interaction with the computer in the first case of the 3-year-old child shows that the child quickly identified that the bodily expression fed into the computer resulted in the output and response. The child was able to describe the interaction that ensued between him and the computer; the feedback from the computer that followed his bodily actions was of interest to him. This can be deduced from Falil’s explanation to his mother after interfacing with the tool:

Mother: How do you teach [the computer]?

Falil: I smile, I laugh.

—

Mother: So, why does the computer detect that expression?

Falil: Because I smiled and laughed.

—

Further along in the discussion, Falil explained the process to his mother using bodily gestures to imitate the interactions he had with the computer.

Akin evaluated Grace’s understanding after the teachable machine exploration process, which was basically concerned with how much she understood about what had been taught to the computer. This was accomplished specifically by asking questions about recognition based on emotions, as shown in the excerpt below:

Akin: How did you express your emotion?

Grace: I squeezed my face.

Akin: What again?

Grace: I also frowned or showed a happy face.

Akin: What did the computer learn?

Grace: What I did.

Akin: Do you mean your emotion?

Grace: Yes.

Akin: What else can you teach?

Grace: Anything.

For the other participants, after the exploration phase with the teachable machine, the guardian interviewed each group representative to share their experiences with the machine. Below is the transcription of one of the audio recordings which involved nine-year-old Jessica.

Mother: Can you explain what you learned today?

Jessica: Waoo, interesting! I learned about the teachable google machine and how a person shows themselves to the computer, and the computer will differentiate you and someone else.

Mother: So what did you teach the computer?

Jessica: How do I feel? Am I happy or sad?

At this point, the mother wanted to know how the computer-related information was either happy or sad and how Jessica managed to teach the computer facial expressions of happiness and sadness.

Mother: So how did you teach the computer to know when you are sad or happy?

Jessica: Hmmm, Mummy, this is amazing! I pressed down the record button to supply images of myself smiling to the computer, after which I trained the pictures collected.

Mother: Trained the images collected? You mean you trained the model?

At this point, the mother believed Jessica was referring to the trained model and had to explain the model by telling her daughter that a model represents something, in this case, a group of images.

Jessica: Yes, I trained the model. That was why it could recognise my emotions of happiness

The following interview involved four-year-old Enemona who could not wait to be asked a question before telling his mum what he experienced.

Enemona: Mummy! Mummy! I saw myself on the computer

He was referring to his image that was captured using the computer’s webcam

Mother: So what happened?

Enemona: The computer then called out my name; that means the computer knows me.

Mother: That was because you taught the computer your name.

Enemona: That means I can teach the computer other things apart from my face?

Mother: Yes, you can; please go ahead and teach the computer other things.

At this point, the child was excited to begin another phase of exploration by using bodily gestures. While the children’s expressions of their experiences with GTM show they understood the input-output relationships, none of them used ML-related vocabularies such as “training”, “data” or “model” or other computing concepts during their explanations. The children focused on describing their use of bodily expressions during the exploration process and how the machine recognised the inputted samples. Even though these ML-related and computing terms were mentioned to the children during the exploratory phase, the more mature members of the community (parents, familiar adults and more experienced peers) may not have provided the guidance and help required to develop this conceptual understanding.

What are the children learning with GTM?

Using an informal and playful interaction with GTM, we created a rich and positive exposure for multi-age groups of children in Nigeria to learn the basic concepts of machine learning. Adopting playful learning with a teachable machine interface to expose these sets of children to machine learning concepts will prepare them for further learning in this emerging field. Utilising such an approach is valuable, as several studies have shown that playful learning is a useful approach for using novel tools and learning environments because it supports engagement and hands-on learning, which usually produces joy in learning and is also an effective way to foster students’ learning, creativity, and imagination (Baker et al., 2021; Kangas et al., 2017; Kangas, 2010). We chose to explore machine teaching with children because it can be an effective vehicle for exposing children to machine learning and AI concepts at an early stage (Dwivedi et al., 2021).

The literature has explored the use of GTM to introduce ML concepts to students from early childhood education through to high school. For instance, the studies of Vartiainen, Tedre, et al. (2020), Vartiainen, Tedre, et al. (2020), and Vartiainen et al. (2021) show that the ML-based technology supported students in developing various kinds of design ideas that harnessed face recognition, gestures, or voice recognition for solving real-life problems. Correspondingly, Lee and Chun (2021) utilised GTM to teach elementary students AI and have found that both the students’ interest and understanding of AI has improved since the application of the programme. The study by Lee et al. (2021) introduced supervised learning (e.g. concepts, processes, and bias) with GTM. Actua, a Canadian STEM education non-profit (Actua, n.d.), is reported to have found it a web application that introduces key AI concepts of “classification” using images and sound. In addition, using GTM, Payne explored supervised learning and explained concepts of bias in her MIT AI Ethics Education Curriculum (Payne, 2019). In the study by Dwivedi et al. (2021), the teachable machine interface was used to introduce ML metrics (confidence scores), model swapping for pattern recognition and the enablement of quick data inspection (e.g. images vs. gestures). The usage and the projects created by students with GTM indicate that the tool is a resource for students’ creative projects in addition to being useful for learning ML concepts. Carney et al. (2020, p. 4) assert that, based on the use of the tool among teachers and researchers, “GTM facilitates active learning of AI concepts by requiring students to interact with those concepts by making models themselves”.

Building on prior studies, we explored a teachable machine interface with children and opted for gesture, face and voice recognition tasks. With this interface, we believe students were introduced to classification and, by extension, concepts and the process of supervised learning. In this way, the children have begun to understand data and its relation to machine learning, the relationship between input and output and why training data matters. Even though the children interfaced with GTM in short sessions with the goal of exposing them to some machine learning ideas, their engagement with GTM, the interviews and anecdotal information suggest that the students have begun to understand some basic ML concepts. Such concepts include the awareness that “training data matters”. Data is the most important part of machine learning because no model can be trained without data. Realising that the performance of an ML algorithm is only as good as the data used to train it is a powerful step in laying the foundation for more sophisticated ideas concerning how ML works and operates.

Discussion

This study explored the deployment of machine learning technology on K–12 children in informal settings in Nigeria and the unfolding teaching and learning process. We introduced 18 children between the ages of 3 to 13 years to machine learning technology such as GTM. Teachable machines have helped many people obtain a sense of how machine learning algorithms work. GTM, a web-based system for teaching and learning basic machine learning principles, has been utilised in earlier studies (Toivonen et al., 2020; Vartiainen, Tedre, et al., 2020). Hence, in our study, we carefully selected children from two different locations within the same country to enable them to explore the workings of machine learning without the need to write code. We feel that early exposure to these technologies will help to bridge the digital gap that exists in Nigeria, facilitate new design ideas and materials and reveal the appropriateness of the tools for different contexts or cultural particularities. We found that some of the children were initially sceptical about whether they could teach computers. This finding was deduced during some of the explorations of GTM from children who could not believe the computer could learn to recognise their faces (e.g. Enemona and Samal). This was mainly due to the children’s backgrounds and lack of knowledge about computers and the internet. The children involved had no previous interactions with computers, even though they know and are aware of their existence.

This study divided the children into three groups, that is, face detection, facial recognition-based emotions and voice recognition. The experiment was conducted with the help of familiar adults in the children’s homes. The data analysis was divided into three phases of teaching, exploration and explanation. The results show that the children perfected teaching the teachable machine how to recognise whether they were sad or happy. In addition, they taught the computer how to recognise their faces. Finally, the computer was taught how it could decipher their various voices by calling each of the names corresponding with the voices. The children achieved this by supplying enough test data samples to the teachable machine to create their classification model. The children also accomplished the exploration of the output-input relationship and were surprised to discover the high accuracy rate of GTM. Furthermore, the children were able to explain in their own terms what transpired before, during and after the interaction process with GTM.

This study provides empirical support for the claim that GTM, a machine learning environment used to introduce machine learning concepts, contributes to the development of data literacy and conceptual understanding across K–12 irrespective of the cultural background. The children’s learning gains increased, and the potential of participatory learning for the development of data literacy was also corroborated by the analyses of the children’s interviews after the exploration of the machine learning tool, as the majority of the children were able to interact with the tools. The findings suggest that irrespective of the children’s location in Nigeria, they were able to explore the technological tool and report increased awareness of the teaching and learning process of the machine. The study findings also suggest that learning and development were perceived to occur among the children through interaction and participation with others, including the teaching assistants (familiar adults) and the children’s peers. This could be observed through the guidance and feedback provided by parents and familiar adults when the children were exploring GTM. This finding supports the idea of participatory learning, which emphasises that children develop through active participation in everyday activities with their families, communities and local cultures (Hedges & Cullen, 2012).

There is a large body of literature examining cultural differences in technology use (Jackson & Wang, 2013) and comparing technology acceptance levels (Lee et al., 2013; Yoo & Huang, 2011). According to Gallivan and Srite (2005), the information technology community has a long-standing interest in cultural differences in technology attitudes and use. However, studies exploring and understanding how technology works and how to introduce the inner workings of emerging technology, such as AI to young children considering cultural differences is rare and should be investigated. Understanding the development process, involvement in co-design and experimentation can deepen children’s knowledge of machine learning techniques and empower them to build their own machine learning applications.

In a time when many elementary, primary and high school classrooms, especially in public schools, include almost no digital literacy instruction in an African context, especially Nigeria, we believe this work could provide a model for considering how to engage pupils in project-based and design thinking issues, which may even foster youth activism through machine learning concepts. While the study was carried out in informal settings, interventions in informal settings can serve to challenge abstract knowledge of the inner workings of technology by disrupting children’s conception of what it means for young learners to engage in digital literacy through understandings of computers, code and the internet (Mertala, 2019) and meaning-making practices of emerging technology for new forms and functions. Studies in informal settings have attracted a considerable amount of research. The findings have mainly indicated that such activities or programmes demonstrate the potential to develop enriching contexts in which youth can build their academic, social, emotional and civic skills (Durlak & Weissberg, 2007; Vance, 2010). A recent study by Efstathiou et al. (2018) utilising the augmented reality mobile learning approach to teach primary school students in non-formal settings improved students’ conceptual understanding and historical empathy. However, existing research on teaching machine learning basics in informal and school settings is scarce, with a notable lack of studies comparing contexts and learning approaches.

Limited studies have attempted to research children’s interactions with AI devices in various social, economic and cultural settings. The study by Druga et al. (2019), which focuses on the context within developed countries, identified some differences along the lines of social-economic status and geographies. However, this study explored two different locations (with different literacy levels) within a developing country and found no differences in their interactions with ML-based technology. As a replication study, we also attempted a comparison with an earlier study conducted with Finnish samples (Vartiainen, Tedre, et al., 2020). In Finland, equal access to new technology is widespread, while digital exclusion is predominant in Nigeria due to social inequalities in income distribution. As a result, the Finnish children have the advantage of access to technological devices embedded with ML applications. Surprisingly, the comparison of the findings showed that young children’s interactions with ML-based technology were similar in both countries, across teaching, exploration and explanation phases. The children were able to create training data and train the machine learning model using GTM through several attempts using different facial and bodily expressions. In the exploration phase, it was revealed that the children identified the input-output relationship with GTM. However, in the explanation phase, when the children in each context described their interaction process and experiences with the computer, the Finnish children used internet-related terminologies (e.g. Wikipedia) to support their claims unlike their Nigerian counterparts. This suggests that exploring young children`s knowledge of computing concepts in Nigeria is necessary. Overall, the social and cultural differences make this comparison a little difficult, and any conclusions concerning the two contexts must be treated with caution. While this paper provides some useful insights, more studies considering the homogeneity of samples drawn from both contexts should be conducted. Exploring the children’s conception, attitudes and interest in ML will be another interesting perspective to consider in future experiments.

Although machine learning introduces learning pathways that are very different from traditional computing education (Vartiainen, Tedre, et al., 2020), several studies have been conducted with the aim of understanding the use of computing systems and the conceptions of young children, especially in Western or developed countries. For instance, Mertala’s (2019) study explored five- to seven-year-old children’s concepts of computers, code and the internet utilising a Finnish sample. Chaudron et al. (2018) surveyed young children and the role(s) played by digital technologies across Europe. The study by Bird and Edwards (2015) also presented a framework based on the sociocultural concept to understand how children learn to use technologies through play utilising a sample of Australian children. However, little is known about technology exploration among young children in African settings. Few studies have attempted to examine the impact of technology in teaching and learning (Sanusi et al., 2017) or determine the factors influencing the assistive technology selection process for young children (Van Niekerk et al., 2019). Hubber et al. (2016) position paper also considered whether touch screen tablets could improve educational outcomes in primary school children in Malawi. While more studies are needed reporting children and young learners` exposure to emerging technology, we are not aware of any studies in Nigeria that examine the consequences of exposing children to machine learning concepts and processes. The authors have also found no studies that compare the unfolding process of teaching and learning using computer devices in a Western and non-Western context.

Conclusion

We built our qualitative and exploratory study on top of a study conducted by Vartiainen, Tedre, et al. (2020). We explored young children’s early encounters and insights into machine learning-based technology. We analysed the intervention in the African context and attempted to compare Vartiainen, Tedre and Valtonen’s findings of the interaction process against our newly studied process. While the empirical exploration of children’s learning of machine learning is still in its infancy, these findings support the potential that learning about machine learning displays in non-formal settings. However, the teaching of ML in informal settings can be further improved and include interventions in formal school settings. According to Stromholt and Bell (2018), the coordination of learning across settings creates opportunities for engagement in overlapping and iterative social and scientific practices that reinforce content, relationships, interest, engagement and, ultimately, the uptake of identifying resources for pupils in science and data literacy. In addition, introducing the unfolding process of the machine learning concept can be explored across different settings (Sanusi, 2021a). The inference is that children will draw cultural resources from one setting into another. It will also introduce and merge new social and scientific meaning-making practices into settings inside and outside of the classroom (Stromholt & Bell, 2018). This study represents a baseline exploring machine learning concepts and the democratisation of data literacies, including machine learning-based technology, which requires additional research to be fully understood.

Despite the uneven modernisation and inequalities in African countries, unlike the developed context, the increasing spread of networks, sensors and artificial intelligence are driving a revolution to an unknown destination across the African continent. The proliferation of emerging technologies, such as AI and ML applications in Africa, specifically Nigeria, suggest that it is imperative to develop critical mass of AI skills to contribute and promote local innovations. In order to achieve this, youth should be equipped to be able to interact with AI systems seamlessly as well as trained to be creators and shapers of future AI-related technologies. Beyond informal trainings, the AI/ML competencies of citizens can be developed by imbedding AI/ML lessons in the school curriculum. This could also be an essential part of the country’s national AI strategy implementation as has occurred in several other countries (Han et al., 2018). Learning ML in Nigerian schools is as important as it is in the developed world. The children in this context also grow up with intelligent and smart devices, even though there is unequal access to new technology due to social inequalities. Regardless of the disadvantaged access, youth must be aware, exposed and equipped to contribute to the future AI and technology revolution.

Limitations

Since this paper builds on Vartiainen, Toivonen, et al. (2020) results and used their research design to allow for a continuation and comparison, the limitations of their study are also a potential threat to this study and may affect our findings. Given that we could not change the research design because of the nature of our study (continuation and comparison), we add the following concerns regarding research design to this study’s limitations. First is the issue of the appropriateness of the comparison group. Several characteristics such as family background differ between the groups thus making comparability difficult. Previous knowledge of the children’s technology use also differs. Finally, data was collected through self-reports and observation because the familiar adults and parents did not provide an alternative source of information for the children. Thus, the perspective of those involved in the learning process would be valuable in evaluating the effect of the intervention.

Future directions of machine learning in K–12 education

We highlight the possible future directions and research potential to improve the state of the art in the field of teaching ML in K–12. First, there is a need to report rich and complete empirical contexts in future studies. Although the field is maturing, empirical characteristics and contextual information about the cases and their distribution scenarios are scarce in the literature. Therefore, future research should report contextual information based on the highlighted gap. Second, more research is needed covering specific knowledge areas rather than the whole ML process. Studies on specific knowledge areas of the ML process, such as, for example, requirements engineering, testing and design, may also produce more detailed descriptions of how ML processes can be unravelled. Third, researchers should report the shortcomings or failure cases encountered in the machine learning exposition process (Sanusi, Oyelere, et al., 2022). To increase knowledge on machine learning for novices or young learners, it is essential to analyse and understand the circumstances under which tools or cases have failed. Finally, comparative studies that consider testing or experimenting with two or more ML tools, such as ML4Kids or LearningML (Sanusi et al., 2021), are encouraged in future studies. Such studies will expose which tools work, and which do not, and their effectiveness in introducing the ML concepts. The findings might also inspire a new design model. Co-designing and experimentation need to be explored to deepen children’s understanding of how machine learning works and has been developed.

In this study, we introduced how GTM operates to parents/guardians, who later engaged their children/wards in exploring the learning tool. This approach opens up interesting possibilities for family learning in relation to ML, which could be a powerful way to explore such novel topics in informal contexts. Understanding how researchers are partnering with the community to promote the learning of ML among youths is important. An additional important direction for future research is methodological. While the narratives by the children and the approaches adopted in this study provide rich descriptions of the interactions, including learning about ML, the lack of more structured tasks or pre-post testing makes it difficult to conclude that the ML concept was learned in a robust way. The use of more instruments and data that relates directly to learning developed over several hours or weeks such as pre-post tests, follow up tasks and task-based post interviews should be considered. By doing so, the data would be triangulated with the notes and narratives from the sessions with the children for a more comprehensive portrait of what occurred. Future studies are expected to provide more insights into the questions that this study could not answer. This paper demonstrates how creating a positive and rich exposure to interfacing with a teachable machine through play in an informal setting can be instrumental in preparing children for ML concept learning.

Acknowledgments

We would like to thank all the children, their families and facilitators who made this study possible.

Disclosure statement

No potential conflict of interest was reported by the authors.