Integrated Calendar

Developing better Language Models would benefit a myriad of communities. However, it is prohibitively costly. The talk would describe collaborative approaches to pretraining, such as model merging, which allows the combining of several specialized models into one. Then, it would introduce efficient evaluation to reduce overheads and touch on other accessible and collaborative aspects that best harness the expertise and diversity in Academia.

The rapid growth of generative models allows an ever-increasing variety of capabilities. Yet, these models may also produce undesired content such as unsafe images, private information, or copyrighted material.

In this talk, I will discuss practical methods to prevent undesired generations. First, I will show how the challenge of avoiding undesired generations manifested itself in a simple Capture-the-Flag LLM setting, where even our top defense strategy was breached. Next, I will demonstrate a similar vulnerability in state-of-the-art concept erasure methods for Text-to-Image models. Finally, I will describe the notion of ‘Unconditional Concept Erasure’ aiming to mitigate such vulnerabilities. I will show that Task Vectors can achieve Unconditional Concept Erasure, and discuss the challenge of applying Task Vectors in practice.

Bio: Niv is a postdoctoral researcher at New York University hosted by Prof. Chinmay Hegde. He received a BSc in mathematics with physics as part of the Technion Excellence Program. He received his PhD in computer science from the Hebrew University of Jerusalem, advised by Prof. Yedid Hoshen. Niv was awarded the Israeli data science scholarship for outstanding postdoctoral fellows (VATAT). He is interested in anomaly detection, model personalization, and AI safety for Vision

Plasmodium falciparum (Pf) parasite is the major cause of malaria disease, resulting in more than 600,000 deaths annually. Patients with cerebral malaria, the most severe form of malaria, show elevated plasma L-pipecolic acid (PA) concentrations in their blood compared to those with mild malaria. However, the origin and function of PA in Pf infection remain mostly elusive. Here, using LC/MS targeted metabolomics we found that the malaria parasite, while growing inside its host human Red Blood Cell (RBC), secretes PA during a specific life stage, the trophozoite. We then demonstrated that pretreatment of the host naïve human RBCs with PA significantly enhances parasitic growth. To further investigate the effect of PA on its primary host, RBCs, we measured the biophysical alterations in the pretreated naïve RBCs using atomic force microscopy combined with machine learning.  Surprisingly, we found that PA modifies the mechanical properties of the host cell’s membrane, turning it significantly softer. Electron paramagnetic resonance data on liposomes suggest that PA’s mechanism may involve altering the lipid mobility. Overall, our findings reveal that the parasite secretes PA to prime its host RBCs for invasion by inducing mechanical changes in the stiffness of the host membrane. These results indicate that PA functions as an active secreted metabolite, facilitating Pf growth within its host cell.

How do animals make perceptual decisions about sensory stimuli to guide motor actions? One hypothesis is that dedicated "perceptual decision" cells process sensory information and drive the appropriate action. Alternatively, perceptual decisions result from competition among cells driving different actions, making decisions inherently coupled to actions. To distinguish between these hypotheses, we designed a vibrotactile detection task in which mice flexibly switched between standard and reversed contingency blocks, respectively requiring them to lick after stimulus presence or absence. Optogenetic inactivation of somatosensory and secondary motor cortices reduced stimulus sensitivity without impairing the ability to lick. However, widefield and two-photon imaging found that differences in cortical activity across perceptual decisions were almost exclusively action-coupled. In addition, we identified a subset of cells that encoded the current contingency block in a gated manner, enabling mice to flexibly make decisions without relying on action-independent decision coding.

Differentially private (DP) algorithms typically exhibit a significant dependence on the dimensionality of their input, as their error or sample complexity tends to grow polynomially with the dimension. This cost of dimensionality is inherent in many problems, as Bun, Ullman, and Vadhan (STOC 2014) showed that any method that achieves lower error rates is vulnerable to tracing attacks (also known as membership inference attacks). Unfortunately, such costs are often too high in many real-world scenarios, such as training large neural networks, where the number of parameters (the ambient dimension) is very high.

On the positive side, the lower bounds do not rule out the possibility of reducing error rates for "easy" inputs. But what constitutes "easy" inputs? And how likely is it to encounter such inputs in real-world scenarios?
In this talk, I will present a few ways to quantify "input easiness" for the fundamental task of private averaging and support them with upper and lower bounds. In particular, I will show types of properties that are both sufficient and necessary for eliminating the polynomial dependency on the dimension.

I will conclude by outlining future research directions and providing a broader perspective on my work.

The talk is mainly based on the following three papers:

(1) FriendlyCore https://arxiv.org/abs/2110.10132 (joint with Edith Cohen, Haim Kaplan, Yishay Mansour, and Uri Stemmer, ICML 2022),
(2) https://arxiv.org/abs/2307.07604 (joint with Naty Peter and Jonathan Ullman, COLT 2024),
(3) https://arxiv.org/abs/2402.06465 (NeurIPS 2024)

How does the brain regulate innate behaviors and emotional states? My researchis driven by a vision to decode evolutionarily conserved neural circuits that regulateaffective states like aggression and anxiety. In my work, I combine deep-brain 2-photoncalcium imaging and holographic optogenetics with theoretical neuroscience approachesto unravel latent manifolds of neural activity and their dynamics. One such dynamic, lineattractors, is hypothesized to encode continuous variables such as eye position, workingmemory, and internal states. However, direct evidence of neural implementation of a lineattractor in mammals has been hindered by the challenge of targeting perturbations tospecific neurons within ensembles. In this talk, I will present our recent breakthroughsdemonstrating causal evidence for line attractor dynamics in neurons encoding anaggressive state and highlight functional connectivity within specific neuronalensembles. This work effectively bridges circuit and manifold levels, providing strongevidence of intrinsic continuous attractor dynamics in a behaviorally relevant mammaliansystem.