FriSem

Lightning Talks

Eline Kupers, Postdoc with Professor Kalanit Grill-Spector in the Vision and Perception Neuroscience Lab, Department of Psychology, Stanford University

Title: Compressive spatiotemporal summation predicts simultaneous suppression in human extrastriate visual cortex

Abstract: When multiple visual stimuli are presented simultaneously in receptive fields of neurons in extrastriate visual areas, their responses are surprisingly lower than when the identical stimuli are presented sequentially. However, what neural computations underlie simultaneous sensory suppression remains elusive. Leveraging fMRI measurements and population receptive field (pRF) modeling, we examined if linear, compressive spatial (CSS), or compressive spatiotemporal (CST) summation within pRFs in higher visual areas generates simultaneous suppression. To disentangle sub-additive spatial summation reducing responses within the pRF from nonlinear temporal summation enhancing transient responses within the pRF, we varied the size and number of transients of both simultaneous and sequential stimuli. In V1, we found that most voxels showed no suppression. However, we did observe some simultaneous suppression for low, but not high, transient conditions. In higher visual areas, voxel responses showed substantial simultaneous suppression in all stimulus conditions. Suppression was largest in voxels of ventral visual areas, followed by dorsal and lateral visual areas. While linear pRFs predicted the lack of suppression in most V1 voxels, CST—but not CSS—pRFs best predicted responses of voxels outside of V1. Indeed, the CST model not only predicted simultaneous suppression in pRFs overlapping multiple stimuli, but also enhanced responses to shorter, transient stimuli, as well as the modest increase in response to larger stimuli. Our results not only provide a new computational framework to understand simultaneous sensory suppression, but also provide a foundation for understanding processing of dynamic visual stimuli in human visual cortex more broadly.

______________________

Effie Li, Postdoc with Professor James McClelland in the McClelland Lab, Department of Psychology, Stanford University

Title: Systematic Generalization and Emergent Structures in Transformers Trained on Structured Tasks

Abstract: Transformer networks have seen great success in natural language processing and machine vision, where task objectives such as next word prediction and image classification benefit from nuanced context sensitivity across high-dimensional inputs. However, there is an ongoing debate about how and when transformers can acquire highly structured behavior and achieve systematic generalization. Here, we explore how well a causal transformer can perform a set of algorithmic tasks, including copying, sorting, and hierarchical compositions of these operations. We demonstrate strong generalization to sequences longer than those used in training by replacing the standard positional encoding typically used in transformers with labels arbitrarily paired with items in the sequence. By finding the layer and head configuration sufficient to solve the task, then performing ablation experiments and representation analysis, we show that two-layer transformers learn generalizable solutions to multi-level problems and develop signs of systematic task decomposition. They also exploit shared computation across related tasks. These results provide key insights into how transformer models may be capable of decomposing complex decisions into reusable, multi-level policies in tasks requiring structured behavior.

______________________

Gabriel Poesia,  PhD student with Noah Goodman in the Computation and Cognition Lab, Department of Psychology, Stanford University

Title: Learning Mathematical Reasoning Across Generations

Abstract: General mathematical reasoning is computationally undecidable. Nevertheless, humans routinely succeed at solving mathematical problems, as well as at teaching the next generations, who can explore much further. What structure of human mathematics would allow this to be possible? We suggest that pieces of this puzzle lie in the hierarchy of abstractions that we construct and use to express problems and solutions. Given the right abstractions, even extremely complex problems that humans find interesting admit relatively short solutions, making them tractable. We show a case study of formalizing and solving problems from the Khan Academy platform in a theorem-proving language. Starting with a base set of axioms, few problems are solvable with naïve search. A reinforcement learning agent manages to push that further but fails to make progress in the harder problems. When combined with an abstraction learning step, where previously found solutions are revisited and rewritten in terms of invented higher-level actions, the agent learns to solve all problems, finding solutions that would be utterly unlikely given just the starting axioms. Furthermore, the learned abstractions let us construct a post-hoc ordering on the solved problems. We find that that order has significant agreement with the Khan Academy curriculum itself, unseen during training. Finally, this reconstructed curriculum accelerates learning of a second-generation agent, illustrating that curricula are especially effective for teaching when learners are inducing abstractions.

_______________________

Martin Noergaard, Postdoc with Professor Russell Poldrack in the Poldrack Lab, Department of Psychology, Stanford University

Title: Reviving Historical Neuroimaging Data to Understand the Role of Serotonin in Depression

Abstract: In recent years, the importance of data sharing has increasingly been recognized by the neuroimaging community, as a solution to leverage optimal and maximally powered research. Good reasons to share data include the lack of replication of neuroimaging findings and the greater scientific impact of multilateral collaborations. In addition, funding bodies and scientific journals increasingly require that data be shared.

One of the most debated research findings in psychiatry concerns the serotonin theory of depression. The serotonin theory of depression suggests that levels of serotonin are lower in the depressed brain and supports why antidepressant medication such as selective serotonin reuptake inhibitors (SSRI’s), blocking the serotonin transporter to increase levels of serotonin, has been used extensively as a first-choice treatment of depression. However, many individual studies investigating the theory of depression have been inconsistent and underpowered (N=10-20 subjects in each group per study), but meta-analyses indicate a trend towards reduced serotonin transporter availability in patients with depression.  

In this work, we seek to revive up to 20 years old historical raw neuroimaging data from institutions all over the world (N=502), all investigating the role of the serotonin transporter in depression. After the data has been shared from the different institutions, the data will go through a state-of-the-art preprocessing pipeline with motion correction, registration, and automatic segmentation of brain regions that was not available at the time when the original data were acquired. Next, we will merge all the data into a mega-analysis and will reveal with sufficient statistical power, whether there is a difference in serotonin transporter availability between patients with depression and healthy controls.

_______________________

Lara Kirfel, Postdoc with Assistant Professor Tobias Gerstenberg, in the Causality in Cognition Lab, Department of Psychology, Stanford University

Title: The communicative function of explanation

Abstract: In this talk, I investigate the communicative dimensions of explanation, revealing some of the rich and subtle inferences people draw from them. I find that people are able to infer additional information from a causal explanation beyond what was explicitly communicated, such as causal structure and normality of the causes. My studies show that people make these inferences in part by appeal to what they themselves would judge reasonable to say across different possible scenarios. The overall pattern of judgments and inferences brings us closer to a full understanding of how causal explanations function in human discourse and behavior, while also raising new questions concerning the prominent role of norms in causal judgment and the function of causal explanation more broadly.