Khoa Pham-Dinh

Tampere University, Finland

• Major in Signal Processing & Machine Learning, minor in Applied Mathematics. GPA 5.00/5.
• Thesis (5/5), at the Signal Analysis and Machine Intelligence (SAMI) group: Optimizing the Rate-Quality-Task Performance in Learned Image Compression for Machine — recast joint compression–segmentation training as constrained optimization over the rate–quality–performance trade-off.

• MSc thesis

Tampere University, Finland

• Major in Applied Mathematics, minor in Statistical Data Analysis. GPA 4.61/5.
• Thesis (5/5): Representation Learning for Candlestick Time-Series Data: A Contrastive Learning Approach.

• BSc thesis

Trustworthy and Adversarial Computing Lab, Aalto University, Finland

• Adversarial examples for cyber-physical vision systems: attacks realized in the physical world (printed patches, modified objects, AR/VR) rather than purely digital perturbations.
• Gradient-based optimization of adversarial attacks on deep computer vision models under physical constraints.

Visual AI System Research team, Nokia Technologies & Aalto University, Finland

• Super-resolution as post filter for Versatile Video Coding.
• Low-rank adaptation for model finetuning.
• Quantization-aware training and mixed-precision quantization for model compression.

Nordic Inertial, Finland (part-time)

• Built machine learning models for zero-velocity detection on IMU sensor data.
• Implemented sensor-fusion-based headlight leveling control in automotive vehicles.

Video Coding for Machines project, Nokia Technologies & Tampere University, Finland

• Defined the semantic-segmentation evaluation methodology for Video Coding for Machines (Detectron2 and Mask2Former as task networks), adopted into the MPEG Common Test Conditions used to develop the standard (S1) and published as publication 2.
• Evaluated the VCM standard on object detection and tracking using the TVD and SFU-HW datasets.

GaitMaven project for gait analysis, Tampere University, Finland

• Performed sensor fusion across GPS-aided IMU, heart rate and spirometer streams from 16 participants, producing the synchronized dataset used for training.
• Benchmarked recurrent (RNN, LSTM, GRU), convolutional, ResNet, DenseNet and Xception architectures for oxygen-consumption estimation.