Knowledge Distillation: Enhancing Neural Network Compression with Integrated Gradients

Efficient deployment of deep neural networks on resource-constrained devices demands advanced compression techniques that preserve accuracy and interoperability. This paper proposes a machine learning framework that augments Knowledge Distillation (KD) with Integrated Gradients (IG), an attribution method, to optimise the compression of convolutional neural networks. We introduce a novel data augmentation strategy where IG maps, precomputed from a teacher model, are overlaid onto training images to guide a compact student model toward critical feature representations. This approach leverages the teacher's decision-making insights, enhancing the student's ability to replicate complex patterns with reduced parameters. Experiments on CIFAR-10 demonstrate the efficacy of our method: a student model, compressed 4.1-fold from the MobileNet-V2 teacher, achieves 92.5% classification accuracy, surpassing the baseline student's 91.4% and traditional KD approaches, while reducing inference latency from 140 ms to 13 ms--a tenfold speedup. We perform hyperparameter optimisation for efficient learning. Comprehensive ablation studies dissect the contributions of KD and IG, revealing synergistic effects that boost both performance and model explainability. Our method's emphasis on feature-level guidance via IG distinguishes it from conventional KD, offering a data-driven solution for mining transferable knowledge in neural architectures. This work contributes to machine learning by providing a scalable, interpretable compression technique, ideal for edge computing applications where efficiency and transparency are paramount.

Skin feature point tracking using deep feature encodings

Facial feature tracking is a key component of imaging ballistocardiography (BCG) where accurate quantification of the displacement of facial keypoints is needed for good heart rate estimation. Skin feature tracking enables video-based quantification of motor degradation in Parkinson's disease. Traditional computer vision algorithms include Scale Invariant Feature Transform (SIFT), Speeded-Up Robust Features (SURF), and Lucas-Kanade method (LK). These have long represented the state-of-the-art in efficiency and accuracy but fail when common deformations, like affine local transformations or illumination changes, are present. Over the past five years, deep convolutional neural networks have outperformed traditional methods for most computer vision tasks. We propose a pipeline for feature tracking, that applies a convolutional stacked autoencoder to identify the most similar crop in an image to a reference crop containing the feature of interest. The autoencoder learns to represent image crops into deep feature encodings specific to the object category it is trained on. We train the autoencoder on facial images and validate its ability to track skin features in general using manually labeled face and hand videos. The tracking errors of distinctive skin features (moles) are so small that we cannot exclude that they stem from the manual labelling based on a $\chi^2$-test. With a mean error of 0.6-4.2 pixels, our method outperformed the other methods in all but one scenario. More importantly, our method was the only one to not diverge. We conclude that our method creates better feature descriptors for feature tracking, feature matching, and image registration than the traditional algorithms.