Author name cluster

Arnon Netzer

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

4 papers

2 author rows

ICML Conference 2025 Conference Paper

Inverse Problem Sampling in Latent Space Using Sequential Monte Carlo

Idan Achituve
Hai Victor Habi
Amir Rosenfeld
Arnon Netzer
Idit Diamant
Ethan Fetaya

In image processing, solving inverse problems is the task of finding plausible reconstructions of an image that was corrupted by some (usually known) degradation operator. Commonly, this process is done using a generative image model that can guide the reconstruction towards solutions that appear natural. The success of diffusion models over the last few years has made them a leading candidate for this task. However, the sequential nature of diffusion models makes this conditional sampling process challenging. Furthermore, since diffusion models are often defined in the latent space of an autoencoder, the encoder-decoder transformations introduce additional difficulties. To address these challenges, we suggest a novel sampling method based on sequential Monte Carlo (SMC) in the latent space of diffusion models. We name our method LD-SMC. We define a generative model for the data using additional auxiliary observations and perform posterior inference with SMC sampling based on a backward diffusion process. Empirical evaluations on ImageNet and FFHQ show the benefits of LD-SMC over competing methods in various inverse problem tasks and especially in challenging inpainting tasks.

Details

ICML Conference 2024 Conference Paper

Bayesian Uncertainty for Gradient Aggregation in Multi-Task Learning

Idan Achituve
Idit Diamant
Arnon Netzer
Gal Chechik
Ethan Fetaya

As machine learning becomes more prominent there is a growing demand to perform several inference tasks in parallel. Multi-task learning (MTL) addresses this challenge by learning a single model that solves several tasks simultaneously and efficiently. Often optimizing MTL models entails first computing the gradient of the loss for each task, and then aggregating all the gradients to obtain a combined update direction. However, common methods following this approach do not consider an important aspect, the sensitivity in the dimensions of the gradients. Some dimensions may be more lenient for changes while others may be more restrictive. Here, we introduce a novel gradient aggregation procedure using Bayesian inference. We place a probability distribution over the task-specific parameters, which in turn induce a distribution over the gradients of the tasks. This valuable information allows us to quantify the uncertainty associated with each of the gradients’ dimensions which is factored in when aggregating them. We empirically demonstrate the benefits of our approach in a variety of datasets, achieving state-of-the-art performance.

Details

JAAMAS Journal 2015 Journal Article

Distributed envy minimization for resource allocation

Arnon Netzer
Amnon Meisels
Roie Zivan

Abstract Minimizing envy in distributed discrete resource or task allocation, is an unusual distributed optimization challenge, since the quality of the allocation for each of the agents is dependent, not only on its own allocation, but on the allocation of others as well. Thus, in order to perform distributed search for allocations with minimal envy there is a need to design innovative algorithms that can cope with the challenging constraint structure of an envy minimization problem. Distributed methods for minimizing envy among agents in indivisible resource allocation problems are presented. First, Distributed Envy Minimization Problems (DEMP) are formulated as Distributed Constraint Reasoning problems. When the DEMPs are large, and cannot be solved by a complete search an incomplete local search algorithm is presented. Each transfer of a good from one agent to another involves the change of state of more than one agent. Thus, a minimizing envy local search algorithm must build upon actions (transfers) that include multiple agents. Since DEMPs are particularly susceptible to local minima during local search, the paper proposes an algorithm that alternates between two different hill climbing search phases. The first phase uses one-transfer steps while the other exploits envy cycle elimination steps. An algorithm that minimizes envy while preserving efficiency, is proposed. The proposed algorithm finds a Pareto optimal allocation with low envy. In the context of resource allocation problems, a Pareto optimal solution is particularly desirable since it presents a stable solution. The proposed algorithm first finds a divisible Pareto optimal envy-free allocation using a Fisher market equilibrium. This allocation is transferred into an indivisible allocation of goods while maintaining the Pareto optimal characteristic of the allocation and a low envy level among agents.

Details DOI

AIJ Journal 2012 Journal Article

Concurrent forward bounding for distributed constraint optimization problems

Arnon Netzer
Alon Grubshtein
Amnon Meisels

A distributed search algorithm for solving Distributed Constraints Optimization Problems (DCOPs) is presented. The new algorithm scans the search space by using multiple search processes (SPs) that run on all agents concurrently. SPs search in non-intersecting parts of the global search space and perform Branch & Bound search. Each search process (SP) uses the mechanism of forward bounding (FB) to prune efficiently its part of the global search space. The Concurrent Forward-Bounding (ConcFB) algorithm enables all SPs to share their upper bound across all parts of the global search space. The number of concurrent SPs is controlled dynamically by the ConcFB algorithm, by performing dynamic splitting. Within each SP a dynamic variable ordering is employed in order to help control the balance of computational load among all agents and across different SPs. The ConcFB algorithm is evaluated experimentally and compared to all state of the art DCOP algorithms. The number of Non-Concurrent Logical Operations, Non-Concurrent Steps, the total number of messages sent and CPU time are used as performance metrics. The evaluation procedure considers different DCOP problem types with a varying number of agents and different constraint graphs. As problems become larger and denser, ConcFB is shown to outperform all other evaluated algorithms by 2–3 orders of magnitude in all performance measures. Further evaluations comparing different variants of ConcFB provide important insights into the working of the algorithm and reveals the contribution of its different components.

Details DOI