As the decade turns, we reflect on nearly thirty years of successfulmanipulation of the world's public equity markets. This reflection highlights afew of the key enabling ingredients and lessons... | Continue reading
We have carried out follow-up spectroscopy on three overdense regions of $g$-and $r$-dropout galaxies in the Canada-France-Hawaii Telescope Legacy SurveyDeep Fields, finding two new... | Continue reading
In response to Jaffe and Quinn [math.HO/9307227], the author discusses formsof progress in mathematics that are not captured by formal proofs of theorems,especially in his own work in the theory... | Continue reading
The style-based GAN architecture (StyleGAN) yields state-of-the-art resultsin data-driven unconditional generative image modeling. We expose and analyzeseveral of its characteristic artifacts,... | Continue reading
Objective: Preventing or minimising beer loss when opening a can of beer issocially and economically desirable. One theoretically grounded approach istapping the can prior to opening, although... | Continue reading
A new prior is proposed for learning representations of high-level conceptsof the kind we manipulate with language. This prior can be combined with otherpriors in order to help disentangling... | Continue reading
Diverse subfields of neuroscience have enriched artificial intelligence formany decades. With recent advances in machine learning and artificial neuralnetworks, many neuroscientists are... | Continue reading
This report focuses on the architecture and performance of the IntelligenceProcessing Unit (IPU), a novel, massively parallel platform recently introducedby Graphcore and aimed at Artificial... | Continue reading
The recently discovered binary system LB-1 has been reported to contain a$\sim 70\,M_{\odot}$ black hole (BH). The evidence for the unprecedentedly highmass of the unseen companion comes from... | Continue reading
Entropy and free-energy estimation are key in thermodynamic characterizationof simulated systems ranging from spin models through polymers, colloids,protein structure, and drug-design. Current... | Continue reading
Single-issue processor cores are very energy efficient but suffer from thevon Neumann bottleneck, in that they must explicitly fetch and issue theloads/storse necessary to feed their ALU/FPU.... | Continue reading
Human observers can learn to recognize new categories of images from ahandful of examples, yet doing so with machine perception remains an openchallenge. We hypothesize that data-efficient... | Continue reading
Financial time series forecasting is, without a doubt, the top choice ofcomputational intelligence for finance researchers from both academia andfinancial industry due to its broad... | Continue reading
Normalizing flows provide a general mechanism for defining expressiveprobability distributions, only requiring the specification of a (usuallysimple) base distribution and a series of bijective... | Continue reading
This article provides a very simple proof of the quadratic formula. Thederivation is computationally light and conceptually natural, and has thepotential to demystify the quadratic formula for... | Continue reading
Handwriting-based gender classification is a well-researched problem that hasbeen approached mainly by traditional machine learning techniques. In thispaper, we propose a novel deep... | Continue reading
Type Ia supernovae (SNe Ia) provided the first strong evidence that theexpansion of the universe is accelerating. With SN samples now more than tentimes larger than those used for the original... | Continue reading
Deep learning frameworks have often focused on either usability or speed, butnot both. PyTorch is a machine learning library that shows that these two goalsare in fact compatible: it provides an... | Continue reading
Deep Neural Network (DNN) workloads are quickly moving from datacenters ontoedge devices, for latency, privacy, or energy reasons. While datacenternetworks can be protected using conventional... | Continue reading
Neural networks have a reputation for being better at solving statistical orapproximate problems than at performing calculations or working with symbolicdata. In this paper, we show that they... | Continue reading
Scanning and filtering over multi-dimensional tables are key operations inmodern analytical database engines. To optimize the performance of theseoperations, databases often create clustered... | Continue reading
Crystals are a state of matter characterised by periodic order. Yetcrystalline materials can harbour disorder in many guises, such asnon-repeating variations in composition, atom displacements,... | Continue reading
Training a neural network is synonymous with learning the values of theweights. In contrast, we demonstrate that randomly weighted neural networkscontain subnetworks which achieve impressive... | Continue reading
We study the phenomenon that some modules of deep neural networks (DNNs) aremore \emph{critical} than others. Meaning that rewinding their parameter valuesback to initialization, while keeping... | Continue reading
Learned world models summarize an agent's experience to facilitate learningcomplex behaviors. While learning world models from high-dimensional sensoryinputs is becoming feasible through deep... | Continue reading
If $A$ is an $n \times n$ Hermitian matrix with eigenvalues$λ_1(A),\dots,λ_n(A)$ and $i,j = 1,\dots,n$, then the$j^{\mathrm{th}}$ component $v_{i,j}$ of a unit eigenvector $v_i$... | Continue reading
Constant-factor differences are frequently ignored when analyzing thecomplexity of algorithms and implementations, as they appear to beinsignificant in practice In this paper, we demonstrate... | Continue reading
Training generative adversarial networks requires balancing of delicateadversarial dynamics. Even with careful tuning, training may diverge or end upin a bad equilibrium with dropped modes. In... | Continue reading
Markov Decision Processes (MDPs), the mathematical framework underlying mostalgorithms in Reinforcement Learning (RL), are often used in a way thatwrongfully assumes that the state of an agent's... | Continue reading
There are massive amounts of textual data residing in databases, valuable formany machine learning (ML) tasks. Since ML techniques depend on numerical inputrepresentations, word embeddings are... | Continue reading
With new advances in machine learning and in particular powerful learninglibraries, we illustrate some of the new possibilities they enable in terms ofnonlinear system identification. For a... | Continue reading
Weird machines---the computational models accessible by exploiting securityvulnerabilities---arise from the difference between the model a programmer hasin her head of how her program should run... | Continue reading
Computer vision models learn to perform a task by capturing relevantstatistics from training data. It has been shown that models learn spuriousage, gender, and race correlations when trained for... | Continue reading
Wide neural networks with random weights and biases are Gaussian processes,as observed by Neal (1995) for shallow networks, and more recently by Lee etal. (2018) and Matthews et al. (2018) for... | Continue reading
We propose a novel system SwarmCloak for landing of a fleet of four flyingrobots on the human arms using light-sensitive landing pads with vibrotactilefeedback. We developed two types of... | Continue reading
In open agent systems, the set of agents that are cooperating or competingchanges over time and in ways that are nontrivial to predict. For example, ifcollaborative robots were tasked with... | Continue reading
The power consumption of a microprocessor is a huge channel for informationleakage. While the most popular exploitation of this channel is to recovercryptographic keys from embedded devices,... | Continue reading
The leading approaches in language modeling are all obsessed with TV shows ofmy youth - namely Transformers and Sesame Street. Transformers this,Transformers that, and over here a bonfire worth... | Continue reading
Finding appropriate inductive loop invariants for a program is a keychallenge in verifying its functional properties. Although the problem isundecidable in general, several heuristics have been... | Continue reading
Today, at 15:35 (UTC+1) on 25 November 2019, we made our final /22 IPv4 allocation from the last remaining addresses in our available pool. We have now run out of IPv4 addresses. | Continue reading
As a natural consequence of the elementary processes of dust growth, wediscovered that a new class of planets can be formed around supermassive blackholes (SMBHs). We investigated a growth path... | Continue reading
Historically, the pursuit of efficient inference has been one of the drivingforces behind research into new deep learning architectures and buildingblocks. Some recent examples include: the... | Continue reading
Deep neural networks have achieved impressive performance and become de-factostandard in many tasks. However, phenomena such as adversarial examples andfooling examples hint that the... | Continue reading
With the long-term goal of studying quantum gravity in the lab, we proposeholographic teleportation protocols that can be readily executed in table-topexperiments. These protocols exhibit... | Continue reading
Lightning Network (LN) is designed to amend the scalability and privacyissues of Bitcoin. It is a payment channel network where Bitcoin transactionsare issued off the blockchain and onion routed... | Continue reading
Human settlements are the cause and consequence of most environmental andsocietal changes on Earth; however, their location and extent is still underdebate. We provide here a new 10m resolution... | Continue reading
We identify a phenomenon, which we refer to as multi-model forgetting, thatoccurs when sequentially training multiple deep networks with partially-sharedparameters; the performance of... | Continue reading