Jack Taylor

Finally, even though it is more computationally constrained and interpretable, Wasserstein distance performs similarly well to features extracted from the best-performing layers of artificial neural networks!

We also show that the optimal transport framework can accommodate geometric invariance. For example, Gromov-Wasserstein distance can be used to capture representations invariant to differences in position and size

New preprint! 📰 We argue that optimal transport provides a useful framework for describing early orthographic processing/representations. In particular, we focus on letter shape similarity... biorxiv.org/content/10.110…

🚀 Exciting News! I'm recruiting a PhD student to join my lab at @shefcompsci and @ShefRobotics, where you'll help shape the future of Human-Centric AI for Social Robots! 📬Project & funding details: tinyurl.com/msr46z48 We’d love your help with a repost! #PhD #AI #Robotics

@stephenburanyi Update: I have a copy - get in touch if you fancy a gander :)

tagging @stephenburanyi in case he happens to remember

I often see this quoted from some 2005(?) Deutsche Bank report, but I've never been able to find the original. Does anyone happen to know where I could find a copy?

So happy to see this EEG study out! 🧠⚡️ We were surprised to see that, in biasing contexts, the word N1 showed a pattern seemingly inconsistent with a simple predictive coding model

he looks so cool for 1865

Fonts hanging out

@wmvanvliet @robustgar (5)...(if we treat complexity as a kind of prediction error). This puts me in mind of some of the work from @GaglBenjamin et al. on orthographic prediction error where they calculate pixel differences from an orthographic prior.

@wmvanvliet @robustgar (4) Looking at whether the effect differed across 4 bins of predictability... maybe? It looks like the complexity-amplitude correlation depends on predictability. Again though, it looks like the direction is inconsistent with predictive coding...

New EEG preprint! 🧠⚡️👲 With @robustgar and Sara Sereno, we collected EEG data from 68 participants to test a simple predictive coding account of the word N1 (N170). Surprisingly, we found a pattern of effects in the opposite direction to that expected... doi.org/10.1101/2023.0…

@wmvanvliet @robustgar No actually! If using the canny edge detector, would you just sum the pixels in the edges as an estimate of complexity?

@JackEdTaylor @robustgar Did you compute visual complexity of the stimuli? In my experience N1 is super sensitive to it. You controlled word length, but even differences in visual complexity between letters can affect N1. Canny edge detection is a very effective N1 predictor.