Residual adapters and knowledge distillation are usually combined — but they capture the same improvement. A framework for choosing between them when compressing biological foundation models to the edge.
A deployable adapter-mixture for compact vision transformers that keeps static execution graphs for edge accelerators — and asks whether expert-style mixtures actually help.
When does geography actually improve species recognition — and when does it just re-express what the image already shows?
Neuroscience-informed heuristics for searching ultra energy-efficient, performant edge vision transformers via data-dependent subnetwork computation.
Battery-free detection and classification of insects on solar-powered milli-robots, with tiny ML for microcontrollers.
A field study across Pacific Northwest & Namibia conservancies, paired with an on-device transcription→LLM prototype built on EarthRanger.
Putting adaptive federated learning in a 2G context.
Zynq Parrot core profiling for matrix multiplication.
Published in IEEE Sensors, 2018
We use a deep autoencoder approach to learn representations of multivariate physiological signals that can be hashed and used to compute similarities between patients to assist in predicting critical events.
Recommended citation: Dhamala, J., Azuh, E., Al-Dujaili, A., Rubin, J., & O’Reilly, U. M. (2018). Multivariate time-series similarity assessment via unsupervised representation learning and stratified locality sensitive hashing: Application to early acute hypotensive episode detection. IEEE Sensors Letters, 3(1), 1-4. https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8506445
Published in Interspeech, 2019
This work presented a multimodal approach to learn bilingual lexicon directly from speech signals in two languages without the need for text by using vision as an interlingua. The approach starts a line of inquiry that can build word level translation between a pair of languages using say youtube videos that have similar objects but with speech in the two languages.
Recommended citation: Azuh, Emmanuel, David Harwath, and James R. Glass. "Towards Bilingual Lexicon Discovery From Visually Grounded Speech Audio." INTERSPEECH. 2019. http://groups.csail.mit.edu/sls/publications/2019/EmmanuelAzuh_Interspeech-2019.PDF
Published in Computing Within Limits, 2021
In the face of ecological and political limits, we propose a computational approach to opening borders to migrants in an automated way while ensuring safety for the host nation.
Recommended citation: Mensah, E. A., Singanamalla, S., Anderson, R., & Heimerl, K. (2021). When Borders Blur-Overcoming Political Limits with Computing in Truly Global Societies. https://computingwithinlimits.org/2021/papers/limits21-mensah.pdf
Published in To be submitted, NeurIPS 2025, 2025
A learning algorithm for single-tower, on-device edge mixture-of-experts models over hybrid vision transformers (convolution/transformer), toward efficient temporally-aligned multimodal (audio + visual) sensor fusion in off-grid settings.
Recommended citation: Mensah, E. A., et al. Towards Mixture of Audio-Visual Modalities for Wildlife Monitoring on the Edge. (To be submitted, NeurIPS 2025). Forthcoming
Published in ACM Transactions on Computer-Human Interaction (TOCHI), 2026
Interviews with wildfire and forest-management practitioners reveal socio-technical challenges in adopting geospatial technologies — fragmented data, knowledge-sharing barriers, and model-bias concerns — and where HCI can help.
Recommended citation: Migineishvili, N., Grunde-McLaughlin, M., Azuh, E., Wood, S., Just, R., & Reinecke, K. (2026). Wildfire and Forest Management: Opportunities for HCI Research. ACM Transactions on Computer-Human Interaction. https://dl.acm.org/doi/10.1145/3765288
Published in CHI Conference on Human Factors in Computing Systems (CHI 2026), 2026
A field study across Pacific Northwest & Namibia conservancies with an on-device transcription→LLM prototype on EarthRanger, analyzing when on-device AI actually helps conservation field staff.
Recommended citation: Dong, C., Azuh Mensah, E., Ranganathan, V., & Heimerl, K. (2026). The Promise and Peril of On-Device AI for Conservation Work. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. https://dl.acm.org/doi/10.1145/3772318.3791359
Undergraduate Teaching Assistant, MIT, 2018
Created the computer vision component of a practical deep learning class launched in spring 2018, led recitations and mentored student teams in their end of semester projects.
Graduate Teaching Assistant, University of Washington, Computer Science Department, 2021
Prepared teaching material and assisted learning for graduate students taking the systems-for-all breadth course.
Graduate Teaching Assistant, University of Washington, 2022
Taught professional master’s students computer networking concepts, including machine learning approaches for networking and networking systems for machine learning.
Research Mentorship, University of Washington, 2023
Mentored three UW undergraduates in a guided Undergraduate Research Program course, as part of my research on low-resource machine learning for ecology.
Graduate Teaching Assistant, University of Washington, 2023
Created assignments, worksheets, and exam questions introducing non-computer-science majors to important concepts in AI and machine learning.
Graduate Teaching Assistant, University of Washington, 2024
Led office hours introducing undergraduates to major concepts in artificial intelligence, such as Markov models and reinforcement learning.
Graduate Teaching Assistant, University of Washington, 2025
Teaching assistant across multiple offerings of UW’s undergraduate/graduate machine learning course (Spring 2023; Spring 2024 – present). Led instruction sections and office hours on core ML concepts, and advised students on their end-of-quarter projects.