Diagnosis, Feedback, Adaptation: A Human-in-the-Loop Framework for Test-Time Policy Adaptation

Policies often fail due to distribution shift -- changes in the state andreward that occur when a policy is deployed in new environments. Dataaugmentation can increase robustness by making the model invariant totask-irrelevant changes in the agent's observation. However, designers don'tknow which concepts are irrelevant a priori, especially when different endusers have different preferences about how the task is performed. We propose aninteractive framework to leverage feedback directly from the user to identifypersonalized task-irrelevant concepts. Our key idea is to generatecounterfactual demonstrations that allow users to quickly identify possibletask-relevant and irrelevant concepts. The knowledge of task-irrelevantconcepts is then used to perform data augmentation and thus obtain a policyadapted to personalized user objectives. We present experiments validating ourframework on discrete and continuous control tasks with real human users. Ourmethod (1) enables users to better understand agent failure, (2) reduces thenumber of demonstrations required for fine-tuning, and (3) aligns the agent toindividual user task preferences.

Gator: Customizable Channel Pruning of Neural Networks with Gating

The rise of neural network (NN) applications has prompted an increasedinterest in compression, with a particular focus on channel pruning, which doesnot require any additional hardware. Most pruning methods employ eithersingle-layer operations or global schemes to determine which channels to removefollowed by fine-tuning of the network. In this paper we present Gator, achannel-pruning method which temporarily adds learned gating mechanisms forpruning of individual channels, and which is trained with an additionalauxiliary loss, aimed at reducing the computational cost due to memory,(theoretical) speedup (in terms of FLOPs), and practical, hardware-specificspeedup. Gator introduces a new formulation of dependencies between NN layerswhich, in contrast to most previous methods, enables pruning of non-sequentialparts, such as layers on ResNet's highway, and even removing entire ResNetblocks. Gator's pruning for ResNet-50 trained on ImageNet producesstate-of-the-art (SOTA) results, such as 50% FLOPs reduction with only0.4%-drop in top-5 accuracy. Also, Gator outperforms previous pruning models,in terms of GPU latency by running 1.4 times faster. Furthermore, Gatorachieves improved top-5 accuracy results, compared to MobileNetV2 andSqueezeNet, for similar runtimes. The source code of this work is available at:https://github.com/EliPassov/gator.

The impact of war on the environment and health: implications for readiness, response, and recovery in Ukraine

A New Israeli Approach to Iran?

Israeli Prime Minister Naftali Bennett’s cordial relations with the Biden administration and relatively muted posture during nuclear negotiations in Vienna raise questions about whether his government is pursuing a different strategy towards Iran than did his predecessor, Benjamin Netanyahu. This article argues that it is not. Despite reinvigorated Israeli debates critical of Netanyahu’s policies and improved atmospherics with the American government, official Israeli policy remains essentially unchanged. The Israeli government is still wary of nuclear diplomacy, offers few alternatives to continued diplomatic and economic pressure, and views military options as viable even if they can only set back Iran’s nuclear programme temporarily. While Bennett wants to avoid open confrontation with Washington, Israel will not relax tensions with Iran, particularly in non-nuclear arenas like Syria. In the past, Israeli sabotage against Iran’s nuclear assets subsided in the run-up to and after the nuclear agreement; this time around, Israel may not feel so constrained.

Benjamin Netanyahu as a mobilizing symbol in ethno-class divisions among Jewish Israelis, 2009–2021

The public discourse over Israel’s unprecedented political crisis in 2019–2021 (four general elections in only two years) has focused on the personality and actions of one person: Prime Minister Benjamin Netanyahu. Relying on a series of public opinion polls during Netanyahu’s second term (2009–2021), we examine the triadic relationship between the following components: (1) sentiments toward Netanyahu, (2) affiliation with ethno-class Jewish status groups, and (3) political attitudes along the liberal-conservative continuum. We show that while there are real socio-political divisions behind the controversy over Netanyahu, the conflict around his public image reflects and shapes the boundaries between various Jewish ethno-class status groups and enables alignments along these boundaries. The centrality of Netanyahu’s image in Israeli politics, we argue, substitutes substantive political discussions and has stemmed from the failure of some political actors, and especially the Secular Ashkenazi group, to articulate a coherent political vision.

Israel’s relations with Central and Eastern Europe and the Jewish-Israeli Dilemma, 2009-2021

This article examines the role of Jewish Diaspora considerations in Prime Minister Benjamin Netanyahu’s policy towards the states of Central and Eastern Europe. Israel has traditionally sought to attain both its state-centred national goals and those of the Jewish Diaspora. Under the Netanyahu governments (2009–21) a major change took place whereby the Diaspora’s existence was viewed as dependent on Israel’s continued survival and success; hence only a strong Israel can help the Diaspora. The security and wellbeing of the Jewish Diaspora thus remains a central Israeli goal, but it is to be pursued via Israel’s strengthening on the one hand, and the deepening of Jewish identity and awareness among Diaspora Jews, on the other.

The effect of news media political bias on pre-elections poll results: evidence from the 2019–20 Israeli elections

Research has had very little to say about whether polling predictions of elections outcomes are biased in line with the political bias of the news outlets that commission the polls. This article examines the relationship between news media political bias and bias in the published results of media-sponsored pre-elections polls in the three Israeli elections that took place in 2019–20. Given that these elections were largely referenda on Benjamin Netanyahu’s corruption charges – dividing the political system, the media, and the Israeli public into pro-Netanyahu and anti-Netanyahu camps – media political biases are narrowly defined in accordance with news outlets’ general attitudes to the charges. Thus, polling bias is defined as a systematic overestimate or underestimate of the number of parliamentary seats that the bloc of pro-Netanyahu parties will actually receive. It is found that, on average, polls commissioned by anti-Netanyahu media consistently underestimated the number of seats that the pro-Netanyahu bloc would win.

Israel: Political Developments and Data in 2021

The year 2021 can be described in Israeli politics as one of change. After 12 consecutive years in power as Prime Minister, Benjamin Netanyahu left office. After four election campaigns in two years, the ‘Anti‐Netanyahu bloc’ managed to form a minimum winning coalition of 61 Knesset members (one Member of the Knesset did not vote with the coalition). A rotation government was agreed upon, with the post of Prime Minister switching between Naftali Bennet, Yemina (right‐wing party) leader and Yair Lapid, Yesh Atid leader (the largest party in the coalition). Bennet served as Prime Minister first, even though his party only had six seats in the coalition. During 2021, Israel experienced terrorist incidents and a major military operation, as well as severe riots in mixed Jewish–Arab cities. Economic performance improved as the COVID‐19 restrictions were lifted, with economic indicators being better than for other OECD countries.

DNN Intellectual Property Extraction Using Composite Data

As state-of-the-art deep neural networks are being deployed at the core level of increasingly large numbers of AI-based products and services, the incentive for “copying them” (i.e., their intellectual property, manifested through the knowledge that is encapsulated in them) either by adversaries or commercial competitors is expected to considerably increase over time. The most efficient way to extract or steal knowledge from such networks is by querying them using a large dataset of random samples and recording their output, which is followed by the training of a student network, aiming to eventually mimic these outputs, without making any assumption about the original networks. The most effective way to protect against such a mimicking attack is to answer queries with the classification result only, omitting confidence values associated with the softmax layer. In this paper, we present a novel method for generating composite images for attacking a mentor neural network using a student model. Our method assumes no information regarding the mentor’s training dataset, architecture, or weights. Furthermore, assuming no information regarding the mentor’s softmax output values, our method successfully mimics the given neural network and is capable of stealing large portions (and sometimes all) of its encapsulated knowledge. Our student model achieved 99% relative accuracy to the protected mentor model on the Cifar-10 test set. In addition, we demonstrate that our student network (which copies the mentor) is impervious to watermarking protection methods and thus would evade being detected as a stolen model by existing dedicated techniques. Our results imply that all current neural networks are vulnerable to mimicking attacks, even if they do not divulge anything but the most basic required output, and that the student model that mimics them cannot be easily detected using currently available techniques.

Sustainability of Interreligious Dialogue in Indonesia under the Phenomenon of Intolerance by Islamic Populists