Daniel Anthes

This is the result of a fantastic collaboration with Sushrut Thorat, Anna Mitola, Paolo Papale, @konigpeter & @TimKietzmann

All our results point into a new direction: even the earliest phase of visual processing is governed by spatiotemporal trajectories, not a spatial feedforward transfer of information. This has implications for our core understanding of primate vision, modelling, and NeuroAI.

Excited about our new preprint: “The illusory simplicity of the feedforward pass: evidence for the dynamical nature of stimulus encoding along the primate ventral stream” arxiv.org/abs/2604.12825 Work with Sushrut Thorat, Anna Mitola, Paolo Papale, Peter König & Tim Kietzmann

🧵Join us for a short tour of how a large-scale MEG dataset collected during active vision made us question it all. New work with @carmen_amme @PhilipSulewski @EelkeSpaak @martin_hebart @konigpeter biorxiv.org/content/10.110…

Happening today! If you are at #CCN2024 and interested in continual learning / representational drift, come have a chat! (Poster C92)

While I cannot be present at #CCN2024 , the lab of course is - make sure to check out the projects: 🧵

🚨 CogSci paper w/ @martisamuser @adriendoerig @TimKietzmann - escholarship.org/uc/item/1516x0… How is attention controlled in the brain? Attention Schema Theory suggests there is a model controlling the attention (cf. body schema). Does a RL agent need an attention schema? 🧵1/10

1/5 Excited to finally share our new paper (led by @lndriscoll, now a group leader at the Allen!) in @NatureNeuro on modular computation in neural networks! We've explored how artificial recurrent networks handle multiple tasks, offering insights into flexible computation. #tweeprint nature.com/articles/s4159…

1/13 Heavily updated preprint! arxiv.org/abs/2209.11737 We show that the contextual information encoded in Large Language Models (LLMs) is beneficial for modelling the complex visual information extracted by the brain from natural scenes. 🧵

Reconstructing an 𝘂𝗻𝗯𝗼𝘂𝗻𝗱𝗲𝗱 𝘀𝗰𝗲𝗻𝗲 𝗳𝗿𝗼𝗺 𝗮 𝗳𝗲𝘄 𝗼𝘂𝘁𝘄𝗮𝗿𝗱-𝗳𝗮𝗰𝗶𝗻𝗴 𝗶𝗺𝗮𝗴𝗲𝘀 is tough. We present 6Img-to-3D, an efficient, scalable single-shot image to 3D reconstruction method.

What defines the computational contribution of a brain or neural model representation? Should we characterize representations by the geometry of the points corresponding to pieces of content in the high-dimensional response space? Or should we care about the topology? 1/10

@daphne_cor @Timothy0Leary Thank you Daphne!

@AnthesDaniel @Timothy0Leary Very cool work!

It’s time for my first tweet print! Presented at CCN2023 w/ @martisamuser @konigpeter @TimKietzmann “Diagnosing Catastrophe: Large parts of accuracy loss in continual learning can be accounted for by readout misalignment” arxiv.org/abs/2310.05644 🧵1/12

Representations have a tendency to drift during learning, but drift ⇏ forgetting! Stay tuned for future work where we analyse drift as a consequence of learning in more detail (cf. Micou & @Timothy0Leary, CON 2023). Link to our CCN paper: arxiv.org/abs/2310.05644 🧵12/12

We also analysed the dependence of the above phenomena on the dimensionality of the latent space. Smaller networks actually forget more, larger networks are more stable. In all networks, we see signatures of representational drift. 🧵11/12