Federated Generalized Bayesian Learning via Distributed Stein Variational Gradient Descent

Kassab, Rahif; Simeone, Osvaldo

doi:10.1109/tsp.2022.3168490

Cited by 21 publications

(21 citation statements)

References 27 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Unlike the LMC, the Stein Variational Gradient Descent (SVGD) algorithm applies gradient descent directly to KL(•|π) (see Section 2.2 for the complete definition). SVGD is an important alternative to the Langevin algorithm and already has been used extensively in different settings of machine learning, such as variational auto-encoders [Pu et al, 2017], reinforcement learning [Liu et al, 2017], sequential decision making [Zhang et al, 2018[Zhang et al, , 2019b, generative adversarial networks [Tao et al, 2019] and federated learning [Kassab and Simeone, 2022].…”

Section: Svgdmentioning

confidence: 99%

Convergence of Stein Variational Gradient Descent under a Weaker Smoothness Condition

Lukang¹,

Karagulyan²,

Richtárik³

2022

Preprint

View full text Add to dashboard Cite

Stein Variational Gradient Descent (SVGD) is an important alternative to the Langevin-type algorithms for sampling from probability distributions of the form π(x) ∝ exp(−V (x)). In the existing theory of Langevin-type algorithms and SVGD, the potential function V is often assumed to be L-smooth. However, this restrictive condition excludes a large class of potential functions such as polynomials of degree greater than 2. Our paper studies the convergence of the SVGD algorithm for distributions with (L 0 , L 1 )-smooth potentials. This relaxed smoothness assumption was introduced by Zhang et al. [2019a] for the analysis of gradient clipping algorithms. With the help of trajectory-independent auxiliary conditions, we provide a descent lemma establishing that the algorithm decreases the KL divergence at each iteration and prove a complexity bound for SVGD in the population limit in terms of the Stein Fisher information.

show abstract

Section: Svgdmentioning

confidence: 99%

Convergence of Stein Variational Gradient Descent under a Weaker Smoothness Condition

Lukang¹,

Karagulyan²,

Richtárik³

2022

Preprint

View full text Add to dashboard Cite

show abstract

“…Furthermore, in an FL framework, the "collapse" of uncertainty in the model parameter space -also known as epistemic uncertainty -to a single model parameter vector prevents agents from properly communicating their respective states of knowledge about the problem. This, in turn, can yield slower convergence [7].…”

Section: Introductionmentioning

confidence: 99%

“…SVGD approximates the posterior distribution using a set of particles, like MC sampling, while also benefiting from the faster convergence of VI through deterministic optimization, rather than sampling. The Distributed SVGD (DSVGD) protocol introduced in [7] extends SVGD to FL (see Fig. 1a).…”

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

“…Previous work [7] has assumed the possibility to transfer an unlimited amount of information at each iteration round. Therefore, the advantages highlighted in [7] of Bayesian FL were obtained at the cost of larger per-iteration communication overhead. In fact, in DSVGD, agents need to exchange multiple particles at each iteration, rather than a single model parameter vector as in frequentist FL.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Compressed Particle-Based Federated Bayesian Learning and Unlearning

Gong¹,

Simeone²,

Kang³

2022

Preprint

Self Cite

View full text Add to dashboard Cite

Conventional frequentist FL schemes are known to yield overconfident decisions. Bayesian FL addresses this issue by allowing agents to process and exchange uncertainty information encoded in distributions over the model parameters. However, this comes at the cost of a larger per-iteration communication overhead. This letter investigates whether Bayesian FL can still provide advantages in terms of calibration when constraining communication bandwidth. We present compressed particlebased Bayesian FL protocols for FL and federated "unlearning" that apply quantization and sparsification across multiple particles. The experimental results confirm that the benefits of Bayesian FL are robust to bandwidth constraints.

show abstract

Fundamentals and Preliminaries of Federated Learning

Chen,

Cui

2023

Wireless Networks

View full text Add to dashboard Cite

Federated Generalized Bayesian Learning via Distributed Stein Variational Gradient Descent

Cited by 21 publications

References 27 publications

Convergence of Stein Variational Gradient Descent under a Weaker Smoothness Condition

Convergence of Stein Variational Gradient Descent under a Weaker Smoothness Condition

Compressed Particle-Based Federated Bayesian Learning and Unlearning

Fundamentals and Preliminaries of Federated Learning

Contact Info

Product

Resources

About