Integrated Calendar

Much of the thinking about neural population codes was motivated in the past decades by reports on neurons with highly stereotyped tuning functions. Indeed, neurons are often observed to exhibit smooth, unimodal tuning to an encoded variable, centered around preferred stimuli that vary across the neural population. Recent experiments, however, have uncovered neural response functions that are much less stereotyped and regular than observed previously. Some of the most striking examples have been observed in the hippocampus and its associated brain areas. The classical view has been that hippocampal place cells are active only in a compact region of space and exhibit a stereotyped tuning to position. In contrast to this expectation, however, place cells in large environments typically fire in multiple locations, and the multiple firing fields of individual cells, as well as those of the whole population, vary in size and shape. We recently discovered that a remarkably simple mathematical model, in which firing fields are generated by thresholding a realization of a random Gaussian process, accounts for the statistical properties of place fields in precise quantitative detail. The model captures the statistics of field sizes and positions, and generates new quantitative predictions on the statistics of field shapes and topologies. These predictions are quantitatively verified in multiple recent data sets from bats and rodents, in one, two, and three dimensions, in both small and large environments. Together, these results imply that common mechanisms underlie the diverse statistics observed in the different experiments. I will discuss a mechanistic model, which suggests that the random Gaussian process statistics arise due to random connectivity within the CA3-CA1 hippocampal circuit. If time permits I will present another recent work, in which we uncover simple principles underlying the spatial selectivity in a new class of neurons in the medial entorhinal cortex, with tuning curves that are significantly less regular than those of classical grid cells. 

The equivalence between vastly different complex systems allows physcisits to make predictions without analyzing their microscopic details. Conversely, by reducing systems to their minimal constituents, we can describe phenomena that seem inscrutable. I will apply these principles to the synthetic world of neural networks, taking grokking, or delayed generalization, as a case study.

While often attributed to complex representation learning, I will show that grokking arises in simple, analytically tractable settings: linear teacher-student models and logistic regression. By explicitly solving the gradient flow dynamics, we identify the ratio of input dimension to training samples (λ=d/N) as the control parameter of a phase transition from overfitting to generalization. I will present two distinct mechanisms for grokking driven by this criticality. In regression, it appears as a divergence in "grokking time" near the interpolation peak. In classification, it emerges at the phase transition between linearly separable and inseparable data, caused by the divergence of the separating hyperplane. Ultimately, I will illustrate how viewing training dynamics through the lens of statistical physics and critical phenomena provides a simple solution for the mystery of delayed generalization.

We will present a recent result with Joe Chen on the low temperature (2+1)D integer-valued Discrete Gaussian (ZGFF) model. The level lines were conjectured to have cube-root fluctuations near the sides of the box, mirroring the Solid-On-Solid picture. The new results resolve this and further recover the joint limit law of the top level-lines near the sides of the box, which is a product of Ferrari—Spohn diffusions.

Large Language Models (LLMs) have transformed what machines can do and how systems are designed to serve them. These models are both computationally and memory demanding, revealing the limits of traditional optimization methods that once sufficed for conventional systems. A central challenge in building LLM systems is improving system metrics while ensuring response quality.

This talk presents approaches for reducing latency in LLM systems to support interactive applications, from scheduling algorithm design to deployment. It introduces scheduling frameworks that use lightweight predictions of request behavior to make informed decisions about prioritization and memory management across two core settings: standalone LLM inference and API-augmented LLMs that interact with external tools. Across both settings, prediction-guided scheduling delivers substantial latency reductions while remaining practical for deployment.

In this talk, we will explore circuit analysis for interpreting neural network models. After some background on the paradigm and techniques of circuit analysis, I'll present two (and a half) research studies demonstrating the breadth of these interpretability methods.

We will explore how this paradigm can help gain scientific insights into how neural network models operate, exemplified in the first work ("Arithmetic without Algorithms", https://technion-cs-nlp.github.io/llm-arithmetic-heuristics) where we use circuit analysis to reveal how language models solve arithmetic prompts. We will also show that circuit analysis can reveal findings on neural network models and help fix existing problems in them --- specifically targeting the issue of poor performance of VLMs on visual tasks compared to equivalent textual tasks (done in the work "Same Task, Different Circuits", https://technion-cs-nlp.github.io/vlm-circuits-analysis). Lastly, if time permits, we will discuss some current directions for future and ongoing work, mainly on scaling circuit analysis to complex tasks.

Bio:

Yaniv Nikankin is a PhD student at the Technion, working with Yonatan Belinkov. His work focuses on interpretability of neural networks, with a recent focus on scaling to analysis of long-form complex tasks. He is particularly excited about cross-domain applications of interpretability in scientific fields, for goals such as better understanding of scientific foundation models such as pLMs. Yaniv is a recipient of the Israeli Higher Education (VATAT) fellowship.