All Articles

Hot Papers 2020-07-16

1. Anatomy of Catastrophic Forgetting: Hidden Representations and Task Semantics

Vinay V. Ramasesh, Ethan Dyer, Maithra Raghu

A central challenge in developing versatile machine learning systems is catastrophic forgetting: a model trained on tasks in sequence will suffer significant performance drops on earlier tasks. Despite the ubiquity of catastrophic forgetting, there is limited understanding of the underlying process and its causes. In this paper, we address this important knowledge gap, investigating how forgetting affects representations in neural network models. Through representational analysis techniques, we find that deeper layers are disproportionately the source of forgetting. Supporting this, a study of methods to mitigate forgetting illustrates that they act to stabilize deeper layers. These insights enable the development of an analytic argument and empirical picture relating the degree of forgetting to representational similarity between tasks. Consistent with this picture, we observe maximal forgetting occurs for task sequences with intermediate similarity. We perform empirical studies on the standard split CIFAR-10 setup and also introduce a novel CIFAR-100 based task approximating realistic input distribution shift.

2. Motifs for processes on networks

Alice C. Schwarze, Mason A. Porter

The study of motifs in networks can help researchers uncover links between structure and function of networks in biology, the sociology, economics, and many other areas. Empirical studies of networks have identified feedback loops, feedforward loops, and several other small structures as “motifs” that occur frequently in real-world networks and may contribute by various mechanisms to important functions these systems. However, the mechanisms are unknown for many of these mechanisms. We propose to distinguish between “structure motifs” (i.e., graphlets) in networks and “process motifs” (which we define as structured sets of walks) on networks and consider process motifs as building blocks of processes on networks. Using the covariances and correlations in a multivariate Ornstein—Uhlenbeck process on a network as examples, we demonstrate that the distinction between structure motifs and process motifs makes it possible to gain quantitative insights into mechanisms that contribute to important functions of dynamical systems on networks.

3. Green Algorithms: Quantifying the carbon emissions of computation

Loïc Lannelongue, Jason Grealey, Michael Inouye

  • retweets: 33, favorites: 55 (07/17/2020 09:11:44)
  • links: abs | pdf
  • cs.CY | cs.DC

Climate change is profoundly affecting nearly all aspects of life on earth, including human societies, economies and health. Various human activities are responsible for significant greenhouse gas emissions, including data centres and other sources of large-scale computation. Although many important scientific milestones have been achieved thanks to the development of high-performance computing, the resultant carbon impact has been underappreciated. In this paper, we present a methodological framework to estimate the carbon impact (CO2 equivalent) of any computational task in a standardised and reliable way, based on the running time, type of computing core, memory used and the efficiency and location of the computing facility. Metrics to interpret and contextualise carbon impact are defined, including the equivalent distance travelled by car or plane as well as the number of tree-months necessary for carbon sequestration. We develop a freely available online tool, Green Algorithms (www.green-algorithms.org), which enables a user to estimate and report the environmental impact of their computation. The Green Algorithms tool easily integrates with computational processes as it requires minimal information and does not interfere with existing code, while also accounting for a broad range of CPUs, GPUs, cloud computing, local servers and desktop computers. Finally, by applying Green Algorithms, we quantify the environmental impact of algorithms used for particle physics simulations, weather forecasts and natural language processing. Taken together, this study develops a simple generalisable framework and freely available tool to quantify the carbon impact of nearly any computation. Combined with a series of recommendations to minimise unnecessary CO2 emissions, we hope to raise awareness and facilitate greener computation.

4. Visualizing Transfer Learning

Róbert Szabó, Dániel Katona, Márton Csillag, Adrián Csiszárik, Dániel Varga

  • retweets: 10, favorites: 64 (07/17/2020 09:11:44)
  • links: abs | pdf
  • cs.CV | cs.LG

We provide visualizations of individual neurons of a deep image recognition network during the temporal process of transfer learning. These visualizations qualitatively demonstrate various novel properties of the transfer learning process regarding the speed and characteristics of adaptation, neuron reuse, spatial scale of the represented image features, and behavior of transfer learning to small data. We publish the large-scale dataset that we have created for the purposes of this analysis.

5. A Survey of Privacy Attacks in Machine Learning

Maria Rigaki, Sebastian Garcia

  • retweets: 15, favorites: 52 (07/17/2020 09:11:44)
  • links: abs | pdf
  • cs.CR | cs.LG

As machine learning becomes more widely used, the need to study its implications in security and privacy becomes more urgent. Research on the security aspects of machine learning, such as adversarial attacks, has received a lot of focus and publicity, but privacy related attacks have received less attention from the research community. Although there is a growing body of work in the area, there is yet no extensive analysis of privacy related attacks. To contribute into this research line we analyzed more than 40 papers related to privacy attacks against machine learning that have been published during the past seven years. Based on this analysis, an attack taxonomy is proposed together with a threat model that allows the categorization of the different attacks based on the adversarial knowledge and the assets under attack. In addition, a detailed analysis of the different attacks is presented, including the models under attack and the datasets used, as well as the common elements and main differences between the approaches under the defined threat model. Finally, we explore the potential reasons for privacy leaks and present an overview of the most common proposed defenses.

Published 17 Jul 2020

ML Lead at Beatrust. (https://beatrust.com)Tatsuya Shirakawa on Twitter