Filippo Grazioli

I am a senior engineer at Cruise, working on 3D object detection, LiDAR/camera/radar fusion and deep learning for autonomous vehicles.

Email  /  CV  /  Google Scholar  /  Github  /  LinkedIn

My research work has been devoted to machine learning applied to computer vision, automated driving and biomedical informatics.

Can we detect autonomous vehicles from their behaviors in scenarios in which they share the road with human-driven ones? For this work, we created the NexusStreet dataset and made it publicly available.

The interaction between T cell receptors (TCRs) and peptide antigens is critical for human immune responses. Designing antigen-specific TCRs represents an important step in adoptive immunotherapy. We propose a disentangled autoencoder, which can act as generative model for optimized TCR sequences. Optimized TCR sequences present enhanced binding affinity to a given peptide and preserve the backbone of a template TCR.

A multi-sequence variational method for binding prediction between T-cell receptors (TCRs) and peptides presented on cells. We present a multi-sequence generalization of Variational Information Bottleneck (VIB) (Alemi et al., 2016) and call it Attentive Variational Information Bottleneck (AVIB). AVIB leverages multi-head self-attention (Vaswani et al., 2017) to implicitly approximate a posterior distribution over latent encodings conditioned on multiple input sequences.

Here, we investigate the generalization capabilities of current deep-learning-based TCR binding predictors, i.e. models that predict the interaction of T cell receptors (TCRs) and peptide-MHC complexes presented on cells. We create a dataset, named TChard, including samples from IEDB, VDJdb, McPAS-TCR, and the MIRA set, as well as negative samples from both randomization and 10X Genomics assays. Our results show that modern deep learning methods fail to generalize to unseen peptides.

A multimodal learning model derived from the theory of the information bottleneck. By looking at a patient's gut microbiome, we predict if they are affected by a certain disease.

A deep Siamese architecture for depth-based face verification. Leveraging privileged information learning, the model relies only on depth images, hallucinating the color information.

A conditional Generative Adversarial Network (GAN) for facial depth map estimation from monocular intensity images.

A model of control distribution between users and the automated systems. Objective driving data and eye-tracking parameters are used to estimate the model’s accuracy.

A simulator that combines the benefits of both high-level and low-level simulators to execute component and connector models. The simulator allows to include new sensors, actuators and control systems.

WO2023139031 (A1) - METHOD AND SYSTEM FOR PREDICTING TCR (T CELL RECEPTOR)-PEPTIDE INTERACTIONS

WO2023193935 (A1) - METHOD AND SYSTEM FOR PREDICTING GENE EXPRESSION PERTURBATIONS

WO2023138755 (A1) - METHODS OF VACCINE DESIGN

WO2023072421 (A1) - SYSTEM AND METHOD FOR INDUCTIVE LEARNING ON GRAPHS WITH KNOWLEDGE FROM LANGUAGE MODELS

WO2022242886 (A1) - A METHOD AND SYSTEM FOR PREDICTING A PHENOTYPIC FEATURE OF A HOST BASED ON A MICROBIOME OF THE HOST

IT201800008237 (A1) - Sistema e metodo di autenticazione di persone in ambienti a limitata visibilità

Website template by Jon Barron.