talks

I’ll try to occasionally post some of the slides of my talks, in a way that there’s not too much overlap.

Birmingham 2024

At Univ. Birmingham Synergising the Human Brain and Arficial Neural Networks Workshop

• Why care about SSL, what is SSL
• TimeTuning: further tuning a DINO model with videos
• 1-video pretraining: better models than ImageNet from a single video
• 1-image distillation: achieving 69% top-1 on IN-1k with only 1-image as training inputs
• Find the slides here

Barcelona 2024

At Apple Research

• PIN: PEFT method for vision-language models for object localisation
• Brief overview of PEFT methods
• VeRA: LoRA but 10-100x less parameters
• Find the slides here

New Orleans 2023

At NeurIPS: Self-supervised Learning: Theory & Practice workshop

• Why care about SSL when we have CLIP?
• Why video SSL is sooo back
• TimeTuning: further tuning a DINO model with videos
• 1-video pretraining: better models than ImageNet from a single video
• Mimicking video-caption data with non-sense text and clustered images
• Find the slides here

Accra 2023

At Google Ghana & University of Ghana

• Why care about SSL
• SeLa: SSL from Sinkhorn-Knopp optimal transport
• Leopart: Spatially-dense SSL for unsupervised object segmentation
• SeLaVi: video-audio multimodal learning
• TimeTuning: further tuning a DINO model with videos
• Guiding diffusion models with via a SSL signal
• 1-image distillation: achieving 69% top-1 on IN-1k with only 1-image as training inputs
• Find the slides here

Tel-Aviv 2022

At Amazon Research

• Video-text learning with support-set bottlenecks
• Self-supervised object detection via audio-video
• 1-image distillation: achieving 69% top-1 on IN-1k with only 1-image as training inputs
• Find the slides here