TechEd

EdTech provides a novel framework that prioritizes privacy by analyzing student videos on their own devices and provides visual, easy to navigate, graphical summaries to educators. Most importantly, it enables a student support system to assist students with dire emotional states. Unlike other researchers, we hence are the first to make use of the data gathered by the model to achieve student wellbeing and efficiency.  Moreover, this project adds on to the plethora of research done in this field by providing the unique angle of analysis of comparing the effectiveness of ResNet-50, MobileNet, and EfficientNet in emotional engagement detection of online environments. It methodologically uses primary data to determine which is more lightweight and yields better results, along with the learning rate and epoch number at which it does so. The models will use high quality data, a feature of datasets that is rarely seen in research in this field. 

Our goal is to ensure students have more supportive environments in online classes. Through our model, we create a robust system that can easily and accurately classify student emotions in real time. It then compiles the data of all student emotions and generates an easy-to-read graph for educators. This can quickly be analyzed even between giving lectures and saved for later reference. Hence teachers have a stream of information that updates them about the environment of their online classes. From this, they can take required action. But even in the scenario where the teacher lacks the ability or motivation to help students manage their emotional health, we have constructed a framework that externally contacts students. After a certain metric of particular emotion expression (i.e. constant data of negative emotions of fear, anger and sadness over a prolonged period of time), the system automatically messages the student to ask if they require assistance. Full privacy and rights are provided so that students can easily decline. This is not detrimental to the theory of change since people can only be helped if they themselves wish to be helped. In the case that they accept, they will be taken through a self-explanative path of guidance to contact either a therapist, a school counsellor, or even close family and friends. Updates will be carried out to ensure the student has actually been helped. If negative emotions still persist, school counselling services are informed. Hence, due to the extensive layers of the process, there is no gap left and children's emotional states are thoroughly managed and taken care of. 

1. More than 60 percent reduction in the amount of data from students that shows negative emotions post implementation of EdTech models

2. Expanding reach to 500+ users/ downloads on the Zoom Marketplace

3. Verification of validity from 7+ mental health experts and professionals 

4. More than 50 percent decrease in negative mental health statistics collected by school counsellors where EdTech was implemented

The iterative building of this project was carried out on Google Collab software with T4 GPU, using the highly-acclaimed python libraries of Keras and Tensorflow. The dataset was uploaded to Google Drive, where file paths were used to reference the images and train the model on them. Initially, the employed system underwent training with the FER-2013 dataset, which contains 30,000+ images of people of different cultures and ages. However, due to low image quality and lack of color, the Facial Emotion Detection Dataset was chosen instead.  This dataset is a high-quality, coloured dataset consisting of 133 (2448 by 3264 pixels) images. It was taken from Kaggle, a public dataset publishing platform. To pre-process the data, multiple steps were taken. The labeled data was first sorted into its respective emotion class folders, and split into validation, training and testing data by a 10-80-10 split. Training and validation data was shuffled to ensure random selection. 

The employed CNN architectures were integral to this study. Both models were sequential models, which have a linear stack of layers in sequence. Transfer learning proved to be a crucial technique to increase the speed and accuracy of the model. We used pre-trained ResNet50, MobileNet, and EfficientNet models from Keras Applications. To construct the architecture, we removed the fully connected layers at the top of the pre-trained models to enable customization of layers. The shape of input images expected by the model were then specified as 351, 351, 3. 8 output classes were added, namely ‘Happy’, ‘Sad’, ‘Contempt’, ‘Surprised’, ‘Neutral’, ‘Fear’, and ‘Anger’. 

In terms of the dense layers in the models, the functional transfer learning layers were followed by alternating flatten and dense layers. These dense layers were composed of 1024 neurons. For activation, ReLu was used to prevent gradients from saturating and hence solve the issue of vanishing gradients. In the final layer, softmax was used, which helped training converge at a faster rate.  

Moreover, the model weights pre-trained on the standard ImageNet dataset were used. These weights were locked into the models to ensure learned representations are not lost. After the convolutional layers, global average pooling was used to reduce the amount of computation required while retaining important features. In terms of optimizers, we initially implemented Adam, which is a standard method to help the model converge faster. However, upon analysis, we deemed SGD to be better suited due to how well it converged to more optimal solutions. 

In this study, loss calculation was done through sparse categorical cross entropy. In comparison to other methods, it saves time in memory as well as computation. The key metric we used to measure the success of the model was training accuracy, which estimates the potential of a model.

Our core team was formed on a purely volunteering basis with the goal of improving online education environments. Currently we have 10 people in the team, with me heading the operations. We are all high school and college students who take out time everyday to work on EdTech. Moreover, we have a tie-up with a research organization (Ndeavors), under which 5 mentors work with us for assistance in software development and marketing. Their subsidized fees are paid for by our core team. They work for approximately 2 hours per week. Our school teachers, Ms. Manasvi and Juanita also provide a lot of assistance in terms of ideation and technical help. 

This project has been running since 18 months, 11 of which were spent in the research and ideation phase. The projected timeline to complete the software and get our first beneficiaries through the Zoom app in particular, is within the next 4 months. 

While our core team members were mainly people in close contact with each other and had limited diversity regulation, we will now need to expand the team to manage the increased operational work. So, to recruit volunteers and other contractors, we will use blind selections. The only details we will collect from them will be those required for the job or work they are being selected for. With the help of aid from Solve, we will also be able to reduce financial restrictions on team members by sponsoring people of lower socioeconomic status to join.  Moreover, we have a work environment that promotes participation through weekly update meetings where everyone is given a turn to speak. This ensures that no perspective is missed out, which is reflected in the complex layers and angles of our ideation. Our project is also aimed to be inclusive in itself, with the training data being from multiple countries, ethnic populations, genders, ages, etc. This prevents bias in the model's classification of emotions. 

In essence, EdTech is a web app integrated with machine learning technology that efficiently analyses students’ emotional states and gives a comprehensive summary to educators. Its emotional support network and approach makes it ideal for the target market of educational institutions like tuition classes or schools. However, it is also a great product for individual educators.

According to Save Our Schools, there are approximately 5.5 million schools worldwide as of 2023, not including unregistered institutions like private tuition centres (30,000+ in India alone), which are the largest consumers. Moreover, online learning is on the rise-

The 21st century has seen great emphasis placed on development in education due to its importance in building up the global workforce, reducing crime, etc. According to experts in the field of education, analyzing learner environments is the next big step in revolutionizing learning. Hence, the question is not of whether emotional engagement should be analyzed but rather how should it be analyzed. Unlike the alternative of traditional methods, such as taking feedback forms after classes or having substitute teachers analyze class environments, our solution is less time and resource (human resources) heavy. Moreover, since it is automated, it doesn't have any teacher bias or demand characteristics.  

In terms of marketing, our company prioritizes digital marketing as a mode of brand promotion because of how digitally oriented our product and current marketing is. Our key benefit is ease of educators and bringing a paradigm shift to online learning, hence we will extensively advertise our web app on those grounds. Our market image will be that of innovation and professionalism, which are qualities that modern educational institutions strive to achieve. 

We not only advocate for change but also heavily implement it through our routine user feedback forms. Every new user is asked for suggestions, which brings in a plethora of useful ideas. For instance, the idea of integrating it as a Zoom app was first suggested by a user in December, 2023. 

As a pre-established company, EdTech has established  its social media presence and has developed its technological base model. Now, as soon as it is published onto the Zoom Marketplace, it will gain widespread credibility. From there on, we can market it through paid social media promotions and collaborations with educational institutions. Downloading it will be easily accessible through the Zoom Marketplace platform. 

Primarily, we want to follow and modify Spotify’s business plan, with it initially being kept free for users and then slowly shifting its features onto the premium version. This will ensure the technology spreads and is integrated into daily lifestyle, hence resulting in more revenue in the long term. This is also sustainable with the current operational costs since development and management is done on volunteering basis, and we only incur minimal marketing costs. We will also have a subscription system in our premium program for large institutions where they can buy the app in wholesale for multiple devices, hence attracting the target market.

Currently, i handle all technical and operational aspects of TechEd. The marketing team that works with me is employed on volunteering basis. Moreover, my mentor assists in the AI development, for which she was paid a one-time fee of approximately 60,000 INR. This was an initial funding that i provided. Due to the majority of operations being on volunteer-basis and digitally-handled, we currently have minimal costs ranging from 4-5000 INR per year. This is funded through a fundraiser we initially held, where we raised around 15,000 INR. This is sufficient to keep the company smoothly running for a minimum of 1-2 years, after which we expect our break-even point and then a steep rise in profits. 

Our revenue model works mostly on long-term gains, which is sustainable for the company due to its low costs. It will mirror the Spotify business model as was explained before. This will work well because, like how Spotify introduced one AI-based platform for all music, we are presenting novel technology that brings ease to users.