Electrical and Computer Engineering student at the University of Washington working at the intersection of machine learning research and embedded systems.

Per-Neuron Gradient Dynamics as an Overfitting Signal

Investigating whether overfitting leaves a detectable signature in per-neuron gradient behavior during training — independent of the standard train/val loss gap. Core hypothesis: neurons in overfitted models show higher gradient instability over time than neurons in regularized, generalizing models, even at comparable training loss.

• Paired experiments with identical weight initialization, comparing an unregularized model against a regularized counterpart
• Developed directional cosine similarity (gradient agreement epoch-to-epoch, averaged over a post-convergence window) as the primary metric, after raw gradient magnitude variance proved too noisy and scale-sensitive
• Complementary log-ratio magnitude volatility metric, robust to scale differences between models
• Cross-validating across five dataset types — synthetic classification, image, tabular, and real-world regression — to separate a genuine, generalizable effect from a single-setup artifact
• Follow-on direction in progress: using these per-neuron signals as a live training-time trigger for selective neuron freezing — locking in neurons with stable, generalizing gradient behavior while the rest of the network keeps training

→ Repository

A spectral filtering approach that improves a trained DNN's test-set predictions without retraining — by treating the model's output curve as a signal and removing high-frequency noise via a low-pass FFT filter.

• Core insight: an overfit DNN's predictions contain both the true underlying pattern and high-frequency memorisation artefacts; these are spectrally separable, so a hard LPF cutoff in the frequency domain can strip the artefacts without touching the signal
• The DNN is used as an interpolating device — queried at a regular time grid to produce a uniformly sampled 1-D curve suitable for FFT, rather than using its predictions on the (irregularly attended) test set directly
• Cutoff frequency selected via a sweep that maximises R² on a held-out window; no labels required during the sweep — only the model's own output
• Validated on four synthetic time-series with increasing complexity (pure lag-1 → staggered multi-lag → AR(1) with coefficient 10 → five-term multi-lag AR at 40% noise) before being applied to a real-world electricity demand forecasting dataset (2020–2024, hourly), where it reduced RMSE by 20 MW and improved R² by 0.007 over the raw DNN
• Identified key limitation: the method requires a 1-D (time-indexed) output curve and cannot extend to higher-dimensional input spaces — sampling a DNN defined over 6+ features at a dense regular grid is computationally intractable (curse of dimensionality), and FFT has no meaningful interpretation on an irregular high-dimensional sample

→ curve-optimizer

Solo embedded systems project: a 300-LED WS2812B matrix (15×20, serpentine wiring) reacting to live audio in real time — hardware, firmware, and signal processing built end-to-end.

• arduinoFFT (evaluated against esp-dsp) for frequency analysis from a MAX9814 mic
• EMA-based beat detection with min/max normalization, tuned against false-triggering on noise-floor drift
• Three render modes: exponential-decay bass flash, HSV radial frequency map, button-driven lyric display
• Dual-core architecture — audio sampling/FFT on one core, LED rendering/UI input on the other — synchronized via a shared volatile struct rather than FreeRTOS queues, prioritizing low-latency access over queue guarantees given the read/write pattern

→ Embedded-Audio-Visualization-System

CNN pipeline classifying fencing strikes from accelerometer/gyroscope data, built for on-device TinyML deployment. Includes class-imbalance-aware augmentation, z-score normalization, stratified train/val/test splitting, and a separate multi-branch ensemble DNN for activity recognition on the Arduino Nano 33 BLE Sense.

→ sword-strike-classifier

INT8 / float16 / dynamic-range quantization, magnitude-based pruning, and knowledge distillation workflows deployed via Edge Impulse to the Arduino Nano 33 BLE Sense — including a UCI HAR activity-recognition pipeline and a fan-monitoring anomaly detection project.

→ tinyml-quantization

ML / Research — PyTorch · TensorFlow/Keras · scikit-learn · pandas/NumPy Embedded — ESP32 / ESP32-S3 · Arduino · TFLite Micro · Edge Impulse · RTOS concepts Systems — C/C++ · GDB · Valgrind

University of Washington — Electrical & Computer Engineering Coursework spanning TinyML/DNN quantization, embedded systems (ESP32 firmware, sensor interfacing), and systems programming in C.

• Email: bhagwani@uw.edu
• LinkedIn: linkedin.com/in/adityabhagwani
• Portfolio: adityabhagwani.com