Marco Fasoli

I am exploring white-box interpretability methods to understand how large language models arrive at their predictions.

I am comparing two approaches to identifying the painters of medieval manuscripts from cropped letters: classical visual descriptors paired with classifiers such as support vector machines (SVMs) and random forests, and learned representations from ResNet-50. The comparison examines predictive performance and interpretability, alongside how individual letters contribute to painter identification and the trade-off between their frequency and discriminative power. Grad-CAM visualizations help inspect the image regions contributing to the ResNet predictions. The work includes a collaboration with a humanities PhD researcher in France.

This research is ongoing and may lead to a publication. The source code and related research materials remain private while the work is in progress. Access may be granted on request; please describe your interest and intended use.

With my teammates, I developed a game engine and AI for the board game Hive. My main contribution was the AI: Monte Carlo tree search guided by a graph attention network trained through self-play reinforcement learning. We also implemented a second approach using minimax search with an efficiently updatable neural network (NNUE) to evaluate game states.

Hivegame.com profile

The source code for both versions is private at the request of Hive tournament players, to help prevent its use for cheating in competitive games. Access to the code and related materials is available on request; please mention which version you are interested in and how you plan to use it.

Written under the supervision of Prof. Valeria Cardellini, this survey reviews distributed training architectures, including parameter servers, all-reduce, and gossip learning. It covers data, model, and pipeline parallelism, then examines the challenges of training large language models, including communication overhead, memory constraints, and fault tolerance.

Inspired by Anthropic's Golden Gate Claude experiment, I used sparse autoencoder features from Gemma Scope to steer Gemma 2 2B IT and Gemma 3 1B IT toward Italian cultural associations. Adding selected feature directions to the models' internal activations sometimes brought Italian references into otherwise ordinary answers. I also explored the limits of this approach: stronger steering could produce fabricated claims or repetitive text.

I studied how random forests learn spurious color correlations in Colored MNIST (CMNIST). I compared models trained on raw pixels and neural network features, alongside balanced-data and grayscale baselines. Feature-importance maps and Gini impurity analyses helped investigate the models' reliance on color shortcuts. A follow-up article explores an intervention that uses color-dependence diagnostics to guide tree growth.

I compared three methods for isolating malaria parasites within infected cell images: color-based thresholding, watershed segmentation, and segmentation based on Canny edge detection. After contrast enhancement, I evaluated the resulting masks using the segmented-area ratio, a color accuracy index, and mean hue deviation. Visual comparisons helped assess each method's strengths and limitations on this dataset.