Towards an Audio-based CNN for Classroom Observation on a Smartwatch

Zualkernan, Imran A.; Khan, Muhammed S.

doi:10.1109/ai4g50087.2020.9311083

Cited by 2 publications

(2 citation statements)

References 28 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The original CNN proposed in (Salamon & Bello, 2017) had three convolutional layers with 24, 48 and 48 layers respectively and used an input size of 128x128. However, (Zualkernan & Khan, 2020) found that using 5 convolutional layers with 128,96,96,32 and 64 filters respectively and an input shape of 8x16x1 gave better results on the classroom audio data. Batch normalization was applied after each convolutional layer and dropout is applied after the 1st, 3rd and 5th convolutional layer.…”

Section: 𝑙 = 𝑙S + 𝜆U 𝑙Umentioning

confidence: 94%

See 1 more Smart Citation

Exploring Semi-Supervised Learning for Audio-Based Automated Classroom Observations

2022

Proceeedings of the 19th International Conference on Cognition and Exploratory Learning in the Digital Age (CELDA 2022)

View full text Add to dashboard Cite

Systematic classroom observation is often used in evaluating and enhancing the quality of classroom instruction. However, classroom observation can potentially suffer from human bias. In addition, the traditional classroom observation is too expensive for resource-constrained environments (e.g., Sub-Saharan Africa, South and Central Asia). A cost-effective automation of classroom observation could potentially enhance both quality and resolution of feedback to the teacher, and hence potentially result in enhancing quality of instruction. Audio-based automatic classroom observation using supervised deep learning techniques has yielded good results in limited contexts. However, one challenge when using supervised techniques is the high cost of collecting and labelling the classroom audio data. One solution for such data-starved scenarios is to use semi-supervised learning (SSL) which requires significantly lesser data and labels. This paper explores an audio-adaptation of the state-of-the-art SSL FixMatch algorithm to automate classroom observation. An adaptation of the FixMatch algorithm was proposed to automate the coding for the Stallings class observation system. The proposed system was trained on classroom audio data collected in the wild. The supervised approach had an F1-score of 0.83 on 100% labeled data. The proposed FixMatch adaptation achieved an impressive F1-score of 0.81 on 20% labeled data, 0.79 on 15% labeled data, 0.76 on 10% labeled data, and 0.72 using only 5% of labeled data. This suggests that algorithms like FixMatch that use consistency regularization and pseudo-labeling have a great potential for being used to automate classroom observation using a small labelled set of audio snippets.

show abstract

Section: 𝑙 = 𝑙S + 𝜆U 𝑙Umentioning

confidence: 94%

“…The Initial Learning rates for the model was set to 1.6e-5 for each of the different experiments, informed by the experiments performed in (Zualkernan & Khan, 2020). For a learning rate schedule, a cosine learning rate decay (Loshchilov & Hutter, 2016) was used which sets the learning rate to (6).…”

Section: Training and Testingmentioning

confidence: 99%