in/gabriel-lucchesi | https://ghubnerr.github.io
I'm a recent Computer Science graduate from Florida International University and passionate about Applied Deep Learning research and Software Engineering. I love to make an impact in my community and inspire curious people to do the same. Trying to make the world more sci-fi :) I've previously worked at both Google and NVIDIA as an intern, and now partake in fun research projects related to LLMs, Multimodality and Robotics.
- 🌱 I’m currently learning: CUDA, the XLA API, and Diffusion Language Models
- 🍓 Working on a StyleGAN project for real-time video generation and native image outputs with LLMs + diffusion 🍌.
- GPT-2 (124M) From Scratch - Trained on 10% of OpenWebText sharded on 8 A100s using JAX. Optimized for KV-Caching, achieving linear token/s performance. (+ Bonus: Applied Pallas attention kernels to boost performance)
- Variational Auto-Encoders for MNIST → See my lecture! - Implemented in Flax & Optax to reconstruct smooth interpolations of a latent space trained on MNIST representations.
- Reinforcement Learning - RL Agents all the way from Q-Learning to Actor-Critic methods, in Flax
- NNs in JAX - RNNs, CNNs, LSTMs, and MLPs implemented in Flax.
- Tensor Autograd - Auto-differentiation engine in Python for PyTorch-like Tensor API for SGD optimization in computational graphs and neural networks. Includes backward methods for backpropagation in reverse mode AD.
- Vector-Jacobian / Jacobian-Vector Products - A thorough walkthrough of reverse and forward-mode AD with JVPs and VJPs. Visualized on a simple linear regression (very nice charts :D)
+ All gists here
My main research interests are in:
- Reinforcement Learning
- Large Language Models and Reasoning
- Model Interpretability
- Vision-Language Action Models
- Explainability in AI for Robustness in Autonomous Control Systems (NCUR 2024), Advised by Dr. Sumit Kumar Jha
See License and Copyright details for this README here