arXiv subjects

Daniel Rothchild

Publications and source records attributed to Daniel Rothchild.

11 recordsLinked to original sources

arXiv2023

Investigating the Behavior of Diffusion Models for Accelerating Electronic Structure Calculations

We present an investigation into diffusion models for molecular generation, with the aim of better understanding how their predictions compare to the results of physics-based calculations. The investigation into these models is driven by their potential to significantly accelerate electronic structure calculations using machine learning, without requiring expensive first-principles datasets for training interatomic potentials. We find that the inference process of a popular diffusion model for de novo molecular generation is divided into an exploration phase, where the model chooses the atomic species, and a relaxation phase, where it adjusts the atomic coordinates to find a low-energy geometry. As training proceeds, we show that the model initially learns about the first-order structure of the potential energy surface, and then later learns about higher-order structure. We also find that the relaxation phase of the diffusion model can be re-purposed to sample the Boltzmann distribution over conformations and to carry out structure relaxations. For structure relaxations, the model finds geometries with ~10x lower energy than those produced by a classical force field for small organic molecules. Initializing a density functional theory (DFT) relaxation at the diffusion-produced structures yields a >2x speedup to the DFT relaxation when compared to initializing at structures relaxed with a classical force field.

physics.chem-ph↗

arXiv2022

The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink

Machine Learning (ML) workloads have rapidly grown in importance, but raised concerns about their carbon footprint. Four best practices can reduce ML training energy by up to 100x and CO2 emissions up to 1000x. By following best practices, overall ML energy use (across research, development, and production) held steady at <15% of Google's total energy use for the past three years. If the whole ML field were to adopt best practices, total carbon emissions from training would reduce. Hence, we recommend that ML papers include emissions explicitly to foster competition on more than just model quality. Estimates of emissions in papers that omitted them have been off 100x-100,000x, so publishing emissions has the added benefit of ensuring accurate accounting. Given the importance of climate change, we must get the numbers right to make certain that we work on its biggest challenges.

cs.LG↗

arXiv2021

C5T5: Controllable Generation of Organic Molecules with Transformers

Methods for designing organic materials with desired properties have high potential impact across fields such as medicine, renewable energy, petrochemical engineering, and agriculture. However, using generative modeling to design substances with desired properties is difficult because candidate compounds must satisfy multiple constraints, including synthetic accessibility and other metrics that are intuitive to domain experts but challenging to quantify. We propose C5T5, a novel self-supervised pretraining method that enables transformers to make zero-shot select-and-replace edits, altering organic substances towards desired property values. C5T5 operates on IUPAC names -- a standardized molecular representation that intuitively encodes rich structural information for organic chemists but that has been largely ignored by the ML community. Our technique requires no edited molecule pairs to train and only a rough estimate of molecular properties, and it has the potential to model long-range dependencies and symmetric molecular structures more easily than graph-based methods. C5T5 also provides a powerful interface to domain experts: it grants users fine-grained control over the generative process by selecting and replacing IUPAC name fragments, which enables experts to leverage their intuitions about structure-activity relationships. We demonstrate C5T5's effectiveness on four physical properties relevant for drug discovery, showing that it learns successful and chemically intuitive strategies for altering molecules towards desired property values.

cs.LG↗

arXiv2021

Carbon Emissions and Large Neural Network Training

The computation demand for machine learning (ML) has grown rapidly recently, which comes with a number of costs. Estimating the energy cost helps measure its environmental impact and finding greener strategies, yet it is challenging without detailed information. We calculate the energy use and carbon footprint of several recent large models-T5, Meena, GShard, Switch Transformer, and GPT-3-and refine earlier estimates for the neural architecture search that found Evolved Transformer. We highlight the following opportunities to improve energy efficiency and CO2 equivalent emissions (CO2e): Large but sparsely activated DNNs can consume <1/10th the energy of large, dense DNNs without sacrificing accuracy despite using as many or even more parameters. Geographic location matters for ML workload scheduling since the fraction of carbon-free energy and resulting CO2e vary ~5X-10X, even within the same country and the same organization. We are now optimizing where and when large models are trained. Specific datacenter infrastructure matters, as Cloud datacenters can be ~1.4-2X more energy efficient than typical datacenters, and the ML-oriented accelerators inside them can be ~2-5X more effective than off-the-shelf systems. Remarkably, the choice of DNN, datacenter, and processor can reduce the carbon footprint up to ~100-1000X. These large factors also make retroactive estimates of energy cost difficult. To avoid miscalculations, we believe ML papers requiring large computational resources should make energy consumption and CO2e explicit when practical. We are working to be more transparent about energy use and CO2e in our future research. To help reduce the carbon footprint of ML, we believe energy usage and CO2e should be a key metric in evaluating models, and we are collaborating with MLPerf developers to include energy usage during training and inference in this industry standard benchmark.

cs.LG↗

arXiv2021

The Impact of Observing Strategy on Cosmological Constraints with LSST

The generation-defining Vera C. Rubin Observatory will make state-of-the-art measurements of both the static and transient universe through its Legacy Survey for Space and Time (LSST). With such capabilities, it is immensely challenging to optimize the LSST observing strategy across the survey's wide range of science drivers. Many aspects of the LSST observing strategy relevant to the LSST Dark Energy Science Collaboration, such as survey footprint definition, single visit exposure time and the cadence of repeat visits in different filters, are yet to be finalized. Here, we present metrics used to assess the impact of observing strategy on the cosmological probes considered most sensitive to survey design; these are large-scale structure, weak lensing, type Ia supernovae, kilonovae and strong lens systems (as well as photometric redshifts, which enable many of these probes). We evaluate these metrics for over 100 different simulated potential survey designs. Our results show that multiple observing strategy decisions can profoundly impact cosmological constraints with LSST; these include adjusting the survey footprint, ensuring repeat nightly visits are taken in different filters and enforcing regular cadence. We provide public code for our metrics, which makes them readily available for evaluating further modifications to the survey design. We conclude with a set of recommendations and highlight observing strategy factors that require further research.

astro-ph.CO↗

arXiv2020

FetchSGD: Communication-Efficient Federated Learning with Sketching

Existing approaches to federated learning suffer from a communication bottleneck as well as convergence issues due to sparse client participation. In this paper we introduce a novel algorithm, called FetchSGD, to overcome these challenges. FetchSGD compresses model updates using a Count Sketch, and then takes advantage of the mergeability of sketches to combine model updates from many workers. A key insight in the design of FetchSGD is that, because the Count Sketch is linear, momentum and error accumulation can both be carried out within the sketch. This allows the algorithm to move momentum and error accumulation from clients to the central aggregator, overcoming the challenges of sparse client participation while still achieving high compression rates and good convergence. We prove that FetchSGD has favorable convergence guarantees, and we demonstrate its empirical effectiveness by training two residual networks and a transformer model.

cs.LG↗

arXiv2020

SqueezeWave: Extremely Lightweight Vocoders for On-device Speech Synthesis

Automatic speech synthesis is a challenging task that is becoming increasingly important as edge devices begin to interact with users through speech. Typical text-to-speech pipelines include a vocoder, which translates intermediate audio representations into an audio waveform. Most existing vocoders are difficult to parallelize since each generated sample is conditioned on previous samples. WaveGlow is a flow-based feed-forward alternative to these auto-regressive models (Prenger et al., 2019). However, while WaveGlow can be easily parallelized, the model is too expensive for real-time speech synthesis on the edge. This paper presents SqueezeWave, a family of lightweight vocoders based on WaveGlow that can generate audio of similar quality to WaveGlow with 61x - 214x fewer MACs. Code, trained models, and generated audio are publicly available at https://github.com/tianrengao/SqueezeWave.

cs.SD↗

arXiv2019

Communication-efficient distributed SGD with Sketching

Large-scale distributed training of neural networks is often limited by network bandwidth, wherein the communication time overwhelms the local computation time. Motivated by the success of sketching methods in sub-linear/streaming algorithms, we introduce Sketched SGD, an algorithm for carrying out distributed SGD by communicating sketches instead of full gradients. We show that Sketched SGD has favorable convergence rates on several classes of functions. When considering all communication -- both of gradients and of updated model weights -- Sketched SGD reduces the amount of communication required compared to other gradient compression methods from $\mathcal{O}(d)$ or $\mathcal{O}(W)$ to $\mathcal{O}(\log d)$, where $d$ is the number of model parameters and $W$ is the number of workers participating in training. We run experiments on a transformer model, an LSTM, and a residual network, demonstrating up to a 40x reduction in total communication cost with no loss in final model performance. We also show experimentally that Sketched SGD scales to at least 256 workers without increasing communication cost or degrading model performance.

cs.LG↗

arXiv2019

ALTSched: Improved Scheduling for Time-Domain Science with LSST

Telescope scheduling is the task of determining the best sequence of observations (pointings and filter choices) for a survey system. Because it is computationally intractable to optimize over all possible multi-year sequences of observations, schedulers use heuristics to pick the best observation at a given time. A greedy scheduler selects the next observation by choosing whichever one maximizes a scalar merit function, which serves as a proxy for the scientific goals of the telescope. This sort of bottom-up approach for scheduling is not guaranteed to produce a schedule for which the sum of merit over all observations is maximized. As an alternative to greedy schedulers, we introduce ALTSched, which takes a top-down approach to scheduling. Instead of considering only the next observation, ALTSched makes global decisions about which area of sky and which filter to observe in, and then refines these decisions into a sequence of observations taken along the meridian to maximize SNR. We implement ALTSched for the Large Synoptic Survey Telescope (LSST), and show that it equals or outperforms the baseline greedy scheduler in essentially all quantitative performance metrics. Due to its simplicity, our implementation is considerably faster than OpSim, the simulated greedy scheduler currently used by the LSST Project: a full ten year survey can be simulated in 4 minutes, as opposed to tens of hours for OpSim. LSST's hardware is fixed, so improving the scheduling algorithm is one of the only remaining ways to optimize LSST's performance. We see ALTSched as a prototype scheduler that gives a lower bound on the performance achievable by LSST.

astro-ph.IM↗

arXiv2018

Optimizing the LSST Observing Strategy for Dark Energy Science: DESC Recommendations for the Wide-Fast-Deep Survey

Cosmology is one of the four science pillars of LSST, which promises to be transformative for our understanding of dark energy and dark matter. The LSST Dark Energy Science Collaboration (DESC) has been tasked with deriving constraints on cosmological parameters from LSST data. Each of the cosmological probes for LSST is heavily impacted by the choice of observing strategy. This white paper is written by the LSST DESC Observing Strategy Task Force (OSTF), which represents the entire collaboration, and aims to make recommendations on observing strategy that will benefit all cosmological analyses with LSST. It is accompanied by the DESC DDF (Deep Drilling Fields) white paper (Scolnic et al.). We use a variety of metrics to understand the effects of the observing strategy on measurements of weak lensing, large-scale structure, clusters, photometric redshifts, supernovae, strong lensing and kilonovae. In order to reduce systematic uncertainties, we conclude that the current baseline observing strategy needs to be significantly modified to result in the best possible cosmological constraints. We provide some key recommendations: moving the WFD (Wide-Fast-Deep) footprint to avoid regions of high extinction, taking visit pairs in different filters, changing the 2x15s snaps to a single exposure to improve efficiency, focusing on strategies that reduce long gaps (>15 days) between observations, and prioritizing spatial uniformity at several intervals during the 10-year survey.

astro-ph.IM↗

arXiv2018

Optimizing the LSST Observing Strategy for Dark Energy Science: DESC Recommendations for the Deep Drilling Fields and other Special Programs

We review the measurements of dark energy enabled by observations of the Deep Drilling Fields and the optimization of survey design for cosmological measurements. This white paper is the result of efforts by the LSST DESC Observing Strategy Task Force (OSTF), which represents the entire collaboration, and aims to make recommendations on observing strategy for the DDFs that will benefit all cosmological analyses with LSST. It is accompanied by the DESC-WFD white paper (Lochner et al.). We argue for altering the nominal deep drilling plan to have $>6$ month seasons, interweaving $gri$ and $zy$ observations every 3 days with 2, 4, 8, 25, 4 visits in $grizy$, respectively. These recommendations are guided by metrics optimizing constraints on dark energy and mitigation of systematic uncertainties, including specific requirements on total number of visits after Y1 and Y10 for photometric redshifts (photo-$z$) and weak lensing systematics. We specify the precise locations for the previously-chosen LSST deep fields (ELAIS-S1, XMM-LSS, CDF-S, and COSMOS) and recommend Akari Deep Field South as the planned fifth deep field in order to synergize with Euclid and WFIRST. Our recommended DDF strategy uses $6.2\%$ of the LSST survey time. We briefly discuss synergy with white papers from other collaborations, as well as additional mini-surveys and Target-of-Opportunity programs that lead to better measurements of dark energy.

astro-ph.IM↗