Pinned
The PSI blog will be a good place to look for world model updates: neuroailab.github.io/psi-website/bl…
Daniel Yamins
675 posts
Daniel Yamins
@dyamins
CS, psych, and neuro prof @ Stanford. NeuroAI and "regular AI". Also harpsichords and bonsai. danyamins.substack.com
- 1/ I'm often confronted with skepticism that neural network models of the brain are intelligible, or that they're even proper models at all, considering how "different they look" from real brains.
- Counterfactual World Modeling (CWM) is a new project from my group. Our ultimate goal is to build a single unified model that could solve a wide range of human visual tasks in a zero-shot manner -- a kind of pure-vision foundation model. arxiv.org/abs/2306.01828
- 1/ Our work on unified principles for Topographic Deep Artificial Neural Networks is finally out in Neuron! 7 years in the making.
- 1/7 A big problem with deepnet models of the brain is that they require training on huge supervised datasets. So even if they are approximations of neural responses in the "adult animal", the training process is a totally implausible model of learning in real visual development.
- We just finished up Winter quarter CS375: Large-Scale Neural Network Models for Neuroscience. Check out the publicly available Syllabus and lecture notes cs375.stanford.edu/course-calenda…
- If you're interested in computational cognitive neuroscience (CCN) PhD programs, definitely apply to our new track CCN at Stanford CS. cs.stanford.edu/people/faculty… Deadline Dec 5.
- 1/ Do unsupervised learning algorithms match the details of human learning? In our new NeurIPS paper, @ChengxuZhuang and team evaluated this at both real-time and lifelong timescales. Link:
- Cool news about the Stanford CS PhD program: We've just added two new interest areas to the application: Computational Cognition & Neuroscience, and Human-Centered AI. Really excited to see this happen - consider applying! cs.stanford.edu/admissions/phd… R . .
- New paper on self-supervised optical flow and occlusion estimation from video foundation models. @sstj389 @jiajunwu_cs @SeKim1112 @Rahul_Venkatesh tinyurl.com/dpa3auzd @
00:00 - 1/ Excited to announce EISEN, our new work on self-supervised, category-agnostic instance segmentation: EISEN learns to segment real-world objects from a single image by observing how they move in training videos. neuroailab.github.io/eisen/.
- 1/ I've long been interested in intrinsic motivation and theory of mind. with @nickhaber and associates, we did a bunch of work over the years. [like arxiv.org/abs/2305.13452 & bit.ly/3W2feXp & arxiv.org/abs/2305.13396] @JulianDeFreitas @_kunoai
- 1/ Def'n: a "learning rule" is a functional that converts error signals (for some given objective function) to changes in system parameters (e.g. synaptic strengths) such that error decreases after iterated application.
- Replying to @dyamins13/ We thus argue for a "contravariance" principle: the harder the constraint, the smaller the set of mechanisms that can solve the constraint, and thus the more likely any two solving mechanisms (whether biological or artificial) are to be similar in key ways.
