CIFRE Ph.D. contract

• Demonstrated ability to collaborate and innovate within international R&D teams through patents, and papers in international conferences
• Delivered two prototypes with live demos for audio, and gesture recognition running a tiny neural network
• Researched a novel algorithm for flexible parameter precisions down to 1-bit, reducing model size by 50% with acceptable loss
• Created and developed an industry-standard software to train and deploy neural networks on the most constrained hardware (< 8KB)

Advisors: Etienne De Foras @TDK InvenSense and Julyan Arbel, Statify team @INRIA Grenoble Rhône-Alpes.

Patents:

• Lê, M. T. & de Foras, E. (2022). One bit quantization for embedded systems. U.S. patent, filed.
• Ponçot, R., Lê, M. T., de Foras, E., Ataya, A., & Hartwell, P. G. (2022). Method for improved keyword spotting. U.S. patent application, pending.

Publications:

• Lê, M. T. et al. (2023). Efficient Neural Networks for Tiny Machine Learning: A Comprehensive Review. arXiv preprint.
• Lê, M. T. et al. (2023). Regularization for Hybrid N-Bit Weight Quantization of Neural Networks on Ultra-Low Power Microcontrollers. In ICANN 2023, Heraklion, Greece.
• Lê, M. T. et al. (2023). TinyMLOps for real-time ultra-low power MCUs applied to frame-based event classification. In EuroMLSys 2023, Roma, Italy.
• Lê, M.T. et al. (2022). Towards Universal 1-bit Weight Quantization of Neural Networks with end-to-end deployment on ultra-low power sensors. In TinyML EMEA 2022, Limassol, Cyprus.

→ Ph.D. thesis
→ Thesis summary

Keywords: Neural Networks, TinyML, Model Compression, Quantization, Edge Inference, Microcontrollers, Sensors, R&D, Innovation, Prototyping, TensorFlow

Objectives included:

• Outperforming current classic algorithms inaccuracy, power consumption & efficiency of the design cycle
• First application to fingerprints: classification of dry vs wet finger and detection of finger push or pull events

Keywords: Constrained/embedded neural networks, autoML, quantization, TensorFlow

Development of an application for a cloud connected keyboard.
Objectives included:

• Improved real-time communication latency with socket connection to receive signals from the cloud to the application in≈30ms
• Implemented and suggested new features: Dashboards for socket connections and task scheduler (with Angular 5)
• Manage data coming in and out from the cloud to the application.
• Collaborated in SCRUM & Kanban agile framework

Keywords: Front end development, Back end development, Angular 4/5, JavaScript, TypeScript, HTML5, CSS3, Bootstrap, Nodejs, Electron, REST API, WebSocket

Preliminary research study for a CIFRE Ph.D. thesis supervised by Thibaut Arribe, Ph.D. @Kelis and Stéphane Crozat, research professor @Université de Technologie de Compiègne.
The goal was to propose and research what states (e.g. test average, variance, behaviors, hesitations…) could be recover from a large database of singular events, to design a learning dashboard for professors to help identifying first-year students difficulties in STEM for Faq2sciences website in a Scenari context. An event can be switching an answer, validate, coming back to previous questions… and contains a set of JSON attributes (e.g. timestamp, event nature, clicks, quizz values…) and may be noisy.
Objectives included:

• Analyzed data (test results, answering time) using ElasticSearch to understand student progress and behavior (≈75.000 quiz samples)
• Designed a learning dashboard from the analysis with Kibana and then Plotly, intended for an integration to Faq2sciences website.

Keywords: Research, Aggregation, Event-State recovery, Data analysis, Data visualization, Learning Dashboards, NoSQL