Talks

The performance of neural networks like large language models (LLMs) is governed by "scaling laws": the error of the network, averaged across the whole dataset, drops as a power law in the number of network parameters and the amount of data the network was trained on. While the mean error drops smoothly and predictably, scaled up LLMs seem to have qualitatively different (emergent) capabilities than smaller versions when one evaluates them at specific tasks. So how does scaling change what neural networks learn? We propose the "quantization model" of neural scaling, where smooth power laws in mean loss are understood as averaging over many small discrete jumps in network performance. Inspired by Max Planck's assumption in 1900 that energy is quantized, we make the assumption that the knowledge or skills that networks must learn are quantized, coming in discrete chunks which we call "quanta". In our model, neural networks can be understand as being implicitly a large number of modules, and scaling simply adds modules to the network. In this talk, I will discuss evidence for and against this hypothesis, its implications for interpretability and for further scaling, and how it fits in with a broader vision for a "science of deep learning".

---

Zoom link https://pitp.zoom.us/j/93886741739?pwd=NzJrcTBNS2xEUUhXajgyak94LzVvdz09

The rich EM phenomenology in the seconds, minutes, and hours just before, during, and after a compact object merger encodes the magnetization of the binary components, the nature of the post-merger remnant, the neutron star equation of state, the free neutron abundance, and a wide array of other compelling physics. Unfortunately, the requirement to search, find, and classify an electromagnetic counterpart within the large GW localization regions before targeted follow-up with sensitive instruments can begin, excludes access to these earliest times, even for the most well localized GW sources. The ability to promptly localize a GW source to within the field-of-view of a narrow-field sensitive facility, would enable extraordinary science. I will discuss the science cases that require extremely early time observations, and the coordination, instruments, and analyses necessary to achieve it. These include gamma-ray imaging, novel data analysis techniques, pre-merger GW detection, faster space telescopes, and new experiments.

---

Zoom link https://pitp.zoom.us/j/96186614241?pwd=R0xpT0dDZVZzek5RT0x4Q1c3Z1RuUT09

Motivated by the static patch of de Sitter space, we discuss timelike surfaces in general relativity and the initial boundary value problem. We consider the Dirichlet problem and a conformal version thereof. Time permitting we discuss some ideas in lower dimensions.

---

Zoom link https://pitp.zoom.us/j/98357788662?pwd=b3lHS3Q3T2xkNm53SkR3QVVvTGJMdz09

In general relativity, the remnant black hole of a binary black hole merger emits “ringdown” radiation: a superposition of quasinormal modes with characteristic complex frequencies. The detection of more than one of these frequencies would serve as smoking-gun evidence of a black hole and would enable a test of the no-hair theorem. In this talk, I will discuss strategies for identifying quasinormal modes within simulated ringdown waveforms both in linear perturbation theory and full general relativity. I will show that a rich spectrum of modes could exist in the ringdown, including overtones, retrograde modes and nonlinear modes, and that interesting quasi-universal relationships exist between some of their amplitudes. I will also discuss the spectral instability of quasinormal modes and its potential imprints on the ringdown waveform. These results could guide the analysis of real ringdown signals detected by gravitational-wave detectors.

---

Zoom link https://pitp.zoom.us/j/92002625163?pwd=bHNDVVZKYnVNTENLNXJBLzVNKzVSQT09