Learning distributions and hypothesis testing via social learning
|
|
- Rudolph Elliott
- 5 years ago
- Views:
Transcription
1 UMich EECS / 48 Learning distributions and hypothesis testing via social learning Anand D. Department of Electrical and Computer Engineering, The State University of New Jersey September 29, 2015 (Joint work with Tara Javidi and Anusha Lalitha (UCSD)) Work sponsored by NSF under award CCF
2 UMich EECS / 48 Introduction
3 UMich EECS / 48 Some philosophical questions How we (as a network of social agents) make common choices or inferences about the world? If I want to help you learn, should I tell you my evidence or just my opinion? How much do we need to communicate with each other?
4 UMich EECS / 48 Which may have some applications (?) Distributed monitoring in networks (estimating a state). Hypothesis testing or detection using multi-modal sensors. Models for vocabulary evolution. Social learning in animals.
5 UMich EECS / 48 Estimation First simple model: estimate a histogram of local data. Each agent starts with a single color. Pass message to learn the histogram of initial colors or sample from that histogram. Main focus: simple protocols with limited communication.
6 UMich EECS / 48 Hypothesis testing Second simple model: estimate a global parameter θ. Each agent takes observations over time conditioned on θ. Can do local updates followed by communication with neighbors. Main focus: simple rule and rate of convergence.
7 UMich EECS / 48 Social learning i 4 Social learning focuses on simple models for how (human) networks can form consensus opinions: Consensus-based DeGroot model: gossip, average consensus etc. Bayesian social learning (Acemoglu et al., Bala and Goyal): agents make decisions and are observed by other agents. Opinion dynamics where agents change beliefs based on beliefs of nearby neighbors.
8 UMich EECS / 48 On limited messages Both of our problems involve some sort of average consensus step. In the first part we are interested in exchanging approximate messages.
9 UMich EECS / 48 On limited messages Both of our problems involve some sort of average consensus step. In the first part we are interested in exchanging approximate messages. Lots of work in quantized consensus (Aysal-Coates-Rabbat, Carli et al., Kashyap et al. Lavaei and Murray, Nedic et al, Srivastava and Nedic, Zhu and Martinez)
10 UMich EECS / 48 On limited messages Both of our problems involve some sort of average consensus step. In the first part we are interested in exchanging approximate messages. Lots of work in quantized consensus (Aysal-Coates-Rabbat, Carli et al., Kashyap et al. Lavaei and Murray, Nedic et al, Srivastava and Nedic, Zhu and Martinez) Time-varying network topologies (even more references).
11 UMich EECS / 48 On limited messages Both of our problems involve some sort of average consensus step. In the first part we are interested in exchanging approximate messages. Lots of work in quantized consensus (Aysal-Coates-Rabbat, Carli et al., Kashyap et al. Lavaei and Murray, Nedic et al, Srivastava and Nedic, Zhu and Martinez) Time-varying network topologies (even more references). Pretty mature area at this point.
12 UMich EECS / 48 A roadmap W 2,1 2 W 1,3 3 1 W 1,2 W 3,1 i W ji W ij j
13 UMich EECS / 48 A roadmap W 2,1 2 W 1,3 3 1 W 1,2 W 3,1 i W ji W ij j Social sampling and estimating histograms
14 UMich EECS / 48 A roadmap W 2,1 2 W 1,3 3 1 W 1,2 W 3,1 i W ji W ij j Social sampling and estimating histograms Distributed hypothesis testing and network divergence
15 UMich EECS / 48 A roadmap W 2,1 2 W 1,3 3 1 W 1,2 W 3,1 i W ji W ij j Social sampling and estimating histograms Distributed hypothesis testing and network divergence Some ongoing work and future ideas.
16 UMich EECS 2015 > Estimating Histograms 10 / 48 Social sampling and merging opinions A.D., T. Javidi, Distributed Learning of Distributions via Social Sampling, IEEE Transactions on Automatic Control 60(1): pp , January 2015.
17 UMich EECS 2015 > Estimating Histograms 11 / 48 Consensus and dynamics in networks Collection of individuals or agents
18 UMich EECS 2015 > Estimating Histograms 11 / 48 Consensus and dynamics in networks i 4 Collection of individuals or agents Agents observe part of a global phenomenon
19 UMich EECS 2015 > Estimating Histograms 11 / 48 Consensus and dynamics in networks i 4 Collection of individuals or agents Agents observe part of a global phenomenon Network of connections for communication
20 UMich EECS 2015 > Estimating Histograms 12 / 48 Phenomena vs. protocols
21 UMich EECS 2015 > Estimating Histograms 12 / 48 Phenomena vs. protocols Engineering: 7 9 Focus on algorithms Minimize communication cost How much do we lose vs. centralized?
22 UMich EECS 2015 > Estimating Histograms 12 / 48 Phenomena vs. protocols Engineering: 7 9 Focus on algorithms Minimize communication cost How much do we lose vs. centralized? Phenomenological: Focus on modeling Simple protocols What behaviors emerge?
23 UMich EECS 2015 > Estimating Histograms 13 / 48 Why simple protocols? i 4 We are more interested in developing simple models that can exhibit different phenomena. Simple source models. Simple communication that uses fewer resources. Simple update rules that are easier to analyze.
24 UMich EECS 2015 > Estimating Histograms 14 / 48 Communication and graph i 4 The n agents are arranged in a connected graph G.
25 UMich EECS 2015 > Estimating Histograms 14 / 48 Communication and graph i 4 The n agents are arranged in a connected graph G.
26 UMich EECS 2015 > Estimating Histograms 14 / 48 Communication and graph Y i (t) Y i (t) Y i (t) Y i (t) Y i (t) The n agents are arranged in a connected graph G. Agent i broadcasts to neighbors N i in the graph.
27 UMich EECS 2015 > Estimating Histograms 14 / 48 Communication and graph Y 1 (t) Y 2 (t) Y 3 (t) Y 5 (t) Y 4 (t) The n agents are arranged in a connected graph G. Agent i broadcasts to neighbors N i in the graph. Message Y i (t) lies in a discrete set.
28 UMich EECS 2015 > Estimating Histograms 15 / 48 The problem
29 UMich EECS 2015 > Estimating Histograms 15 / 48 The problem Each agent starts with θ i {1, 2,..., M}
30 UMich EECS 2015 > Estimating Histograms 15 / 48 The problem Each agent starts with θ i {1, 2,..., M} Agent i knows θ i (no noise)
31 UMich EECS 2015 > Estimating Histograms 15 / 48 The problem Each agent starts with θ i {1, 2,..., M} Agent i knows θ i (no noise) Maintain estimates Q i (t) of the empirical distribution Π of {θ i }
32 UMich EECS 2015 > Estimating Histograms 16 / 48 Social sampling We model the messages as random samples from local estimates. 1 Update rule from Q i (t 1) to Q i (t) : Q i (t) = W i (Q i (t 1), X i (t), Y i (t 1), {Y j (t 1) : j N i }, t). 2 Build a sampling distribution on {0, 1,..., M}: P i (t) = V i (Q i (t), t). 3 Sample message: Y i (t) P i (t).
33 UMich EECS 2015 > Estimating Histograms 17 / 48 Social sampling X i i W ij trust j X j i Y j (t) update j W ji Y i (t) Q i (0) Q j (0) Q i (1) Q j (1) i Y j (t) messages j i W ij t!1 j Y i (t) W ji Q i (0) Q j (0) Q i (t) Q j (t)
34 UMich EECS 2015 > Estimating Histograms 18 / 48 Possible phenomena
35 UMich EECS 2015 > Estimating Histograms 18 / 48 Possible phenomena 1.00 Estimate at a single node vs. time Node estimate element Q1 Q2 Q3 Q time Coalescence: all agents converge to singletons
36 UMich EECS 2015 > Estimating Histograms 18 / 48 Possible phenomena Node estimate Node estimatess of a single bin vs. time time element true node3 node8 node13 node18 node23 Consensus: agents converge to common ˆΠ Π
37 UMich EECS 2015 > Estimating Histograms 18 / 48 Possible phenomena Node estimate Node estimatess of a single bin vs. time time element true node3 node8 node13 node18 node23 Convergence: agents converge to Π
38 UMich EECS 2015 > Estimating Histograms 19 / 48 Linear update rule Q i (t) = A i (t)q i (t 1) + B i (t)y i (t 1) + j N i W ij (t)y j (t 1) Linear update rule combining Y i P i and Q i.
39 UMich EECS 2015 > Estimating Histograms 19 / 48 Linear update rule Q i (t) = A i (t)q i (t 1) + B i (t)y i (t 1) + j N i W ij (t)y j (t 1) Linear update rule combining Y i P i and Q i. Exhibits different behavior depending on A i (t), B i (t), and W (t).
40 UMich EECS 2015 > Estimating Histograms 20 / 48 Convergence Main idea : massage the update rule into matrix form: Q(t + 1) = Q(t) + δ(t) [ ] HQ(t) + C(t) + M(t). with 1 Step size δ(t) = 1/t 2 Perturbation C(t) = O(δ(t)) 3 Martingale difference term M(t) This is a stochastic approximation: converges to a fixed point of H.
41 UMich EECS 2015 > Estimating Histograms 21 / 48 Example: censored updates Suppose we make distribution P i (t) a censored version of Q i (t): P i,m (t) = Q i,m (t) 1 (Q i,m (t) > δ(t)(1 W ii ))) M P i,0 (t) = Q i,m (t) 1 (Q i,m (t) δ(t)(1 W ii )) m=1
42 UMich EECS 2015 > Estimating Histograms 21 / 48 Example: censored updates Suppose we make distribution P i (t) a censored version of Q i (t): P i,m (t) = Q i,m (t) 1 (Q i,m (t) > δ(t)(1 W ii ))) M P i,0 (t) = Q i,m (t) 1 (Q i,m (t) δ(t)(1 W ii )) m=1 Agent sends Y i (t) = 0 if it samples a rare element in Q i.
43 UMich EECS 2015 > Estimating Histograms 22 / 48 Phenomena captured by censored model Node estimate Node estimatess of a single bin vs. time time element true node3 node8 node13 node18 node23
44 UMich EECS 2015 > Estimating Histograms 22 / 48 Phenomena captured by censored model Node estimate Node estimatess of a single bin vs. time time element true node3 node8 node13 node18 node23 Censored distribution P i (t) guards against marginal opinions.
45 UMich EECS 2015 > Estimating Histograms 22 / 48 Phenomena captured by censored model Node estimate Node estimatess of a single bin vs. time time element true node3 node8 node13 node18 node23 Censored distribution P i (t) guards against marginal opinions. Sampled messages Y i (t) P i (t) are simple messages.
46 UMich EECS 2015 > Estimating Histograms 22 / 48 Phenomena captured by censored model Node estimate Node estimatess of a single bin vs. time time element true node3 node8 node13 node18 node23 Censored distribution P i (t) guards against marginal opinions. Sampled messages Y i (t) P i (t) are simple messages. Decaying weights δ(t) represent solidifying of opinions.
47 UMich EECS 2015 > Estimating Histograms 22 / 48 Phenomena captured by censored model Node estimate Node estimatess of a single bin vs. time time element true node3 node8 node13 node18 node23 Censored distribution P i (t) guards against marginal opinions. Sampled messages Y i (t) P i (t) are simple messages. Decaying weights δ(t) represent solidifying of opinions. Result : all estimates converge almost surely to Π.
48 UMich EECS 2015 > Estimating Histograms 23 / 48 Future directions
49 UMich EECS 2015 > Estimating Histograms 23 / 48 Future directions Find the rate of convergence and dependence of the rate on the parameters
50 UMich EECS 2015 > Estimating Histograms 23 / 48 Future directions Find the rate of convergence and dependence of the rate on the parameters Investigate the robustness of the update rule to noise and perturbations
51 UMich EECS 2015 > Estimating Histograms 23 / 48 Future directions Find the rate of convergence and dependence of the rate on the parameters Investigate the robustness of the update rule to noise and perturbations Continuous distributions?
52 UMich EECS 2015 > Estimating Histograms 23 / 48 Future directions Find the rate of convergence and dependence of the rate on the parameters Investigate the robustness of the update rule to noise and perturbations Continuous distributions? Other message passing algorithms?
53 UMich EECS 2015 > Estimating Histograms 23 / 48 Future directions Find the rate of convergence and dependence of the rate on the parameters Investigate the robustness of the update rule to noise and perturbations Continuous distributions? Other message passing algorithms? Distributed optimization?
54 UMich EECS 2015 > Pseudo-Bayes 24 / 48 Non-Bayesian social learning A. Lalitha, T. Javidi, A., Social Learning and Distributed Hypothesis Testing, ArXiV report number arxiv: [math.st], October, 2014.
55 UMich EECS 2015 > Pseudo-Bayes 25 / 48 Model Set of n nodes.
56 UMich EECS 2015 > Pseudo-Bayes 25 / 48 Model Set of n nodes. Set of hypotheses Θ = {θ 1, θ 2,..., θ M }.
57 UMich EECS 2015 > Pseudo-Bayes 25 / 48 Model Set of n nodes. Set of hypotheses Θ = {θ 1, θ 2,..., θ M }. Observations X (t) i are i.i.d.
58 UMich EECS 2015 > Pseudo-Bayes 25 / 48 Model Set of n nodes. Set of hypotheses Θ = {θ 1, θ 2,..., θ M }. Observations X (t) i are i.i.d. Fixed known distributions {f i ( ; θ 1 ), f i ( ; θ 2 ),..., f i ( ; θ M )}.
59 UMich EECS 2015 > Pseudo-Bayes 25 / 48 Model Set of n nodes. Set of hypotheses Θ = {θ 1, θ 2,..., θ M }. Observations X (t) i are i.i.d. Fixed known distributions {f i ( ; θ 1 ), f i ( ; θ 2 ),..., f i ( ; θ M )}. θ Θ is fixed global unknown parameter
60 UMich EECS 2015 > Pseudo-Bayes 25 / 48 Model Set of n nodes. Set of hypotheses Θ = {θ 1, θ 2,..., θ M }. Observations X (t) i are i.i.d. Fixed known distributions {f i ( ; θ 1 ), f i ( ; θ 2 ),..., f i ( ; θ M )}. θ Θ is fixed global unknown parameter X (t) i f i ( ; θ ).
61 UMich EECS 2015 > Pseudo-Bayes 25 / 48 Model Set of n nodes. Set of hypotheses Θ = {θ 1, θ 2,..., θ M }. Observations X (t) i are i.i.d. Fixed known distributions {f i ( ; θ 1 ), f i ( ; θ 2 ),..., f i ( ; θ M )}. θ Θ is fixed global unknown parameter X (t) i f i ( ; θ ). GOAL Parametric inference of unknown θ
62 UMich EECS 2015 > Pseudo-Bayes 26 / 48 Hypothesis Testing
63 UMich EECS 2015 > Pseudo-Bayes 26 / 48 Hypothesis Testing If θ is globally identifiable, then collecting all observations X (t) = {X (t) 1, X(t) 2,..., X(t) n } at a central locations yields a centralized hypothesis testing problem.
64 UMich EECS 2015 > Pseudo-Bayes 26 / 48 Hypothesis Testing If θ is globally identifiable, then collecting all observations X (t) = {X (t) 1, X(t) 2,..., X(t) n } at a central locations yields a centralized hypothesis testing problem. Exponentially fast convergence to the true hypothesis
65 UMich EECS 2015 > Pseudo-Bayes 26 / 48 Hypothesis Testing If θ is globally identifiable, then collecting all observations X (t) = {X (t) 1, X(t) 2,..., X(t) n } at a central locations yields a centralized hypothesis testing problem. Exponentially fast convergence to the true hypothesis Can this be achieved locally with low dimensional observations?
66 UMich EECS 2015 > Pseudo-Bayes 27 / 48 Example: Low-dimensional Observations If all observations are not collected centrally, node 1 individually cannot learn θ.
67 UMich EECS 2015 > Pseudo-Bayes 27 / 48 Example: Low-dimensional Observations If all observations are not collected centrally, node 1 individually cannot learn θ. = nodes must communicate.
68 UMich EECS 2015 > Pseudo-Bayes 28 / 48 Distributed Hypothesis Testing Define Θ i = {θ Θ : f i ( ; θ) = f i ( ; θ )}.
69 UMich EECS 2015 > Pseudo-Bayes 28 / 48 Distributed Hypothesis Testing Define Θ i = {θ Θ : f i ( ; θ) = f i ( ; θ )}. θ Θ i = θ and θ are observationally equivalent for node i.
70 UMich EECS 2015 > Pseudo-Bayes 28 / 48 Distributed Hypothesis Testing Define Θ i = {θ Θ : f i ( ; θ) = f i ( ; θ )}. θ Θ i = θ and θ are observationally equivalent for node i. Suppose {θ } = Θ 1 Θ 2... Θ n.
71 UMich EECS 2015 > Pseudo-Bayes 28 / 48 Distributed Hypothesis Testing Define Θ i = {θ Θ : f i ( ; θ) = f i ( ; θ )}. θ Θ i = θ and θ are observationally equivalent for node i. Suppose {θ } = Θ 1 Θ 2... Θ n. GOAL Parametric inference of unknown θ
72 UMich EECS 2015 > Pseudo-Bayes 29 / 48 Learning Rule
73 UMich EECS 2015 > Pseudo-Bayes 29 / 48 Learning Rule At t = 0, node i begins with initial estimate vector q (0) i > 0, where components of q (t) i form a probability distribution on Θ.
74 UMich EECS 2015 > Pseudo-Bayes 29 / 48 Learning Rule At t = 0, node i begins with initial estimate vector q (0) i > 0, where components of q (t) i form a probability distribution on Θ. At t > 0, node i draws X (t) i.
75 UMich EECS 2015 > Pseudo-Bayes 30 / 48 Learning Rule Node i computes belief vector,, via Bayesian update b (t) i b (t) i (θ) = ( f i θ Θ f i X (t) i ) ; θ q (t 1) i (θ) ( X (t) i ; θ ) q (t 1) i (θ ).
76 UMich EECS 2015 > Pseudo-Bayes 30 / 48 Learning Rule Node i computes belief vector,, via Bayesian update b (t) i b (t) i (θ) = ( f i θ Θ f i X (t) i ) ; θ q (t 1) i (θ) ( X (t) i ; θ ) q (t 1) i (θ ). Sends message Y (t) i = b (t) i.
77 UMich EECS 2015 > Pseudo-Bayes 31 / 48 Learning Rule Receives messages from its neighbors at the same time.
78 UMich EECS 2015 > Pseudo-Bayes 31 / 48 Learning Rule Receives messages from its neighbors at the same time. Updates q (t) i via averaging of log beliefs, ( n ) exp j=1 W ijlog b (t) j (θ) q (t) i (θ) = θ Θ exp ( n j=1 W ijlog b (t) j (θ ) where weight W ij denotes the influence of node j on estimate of node i. ),
79 UMich EECS 2015 > Pseudo-Bayes 31 / 48 Learning Rule Receives messages from its neighbors at the same time. Updates q (t) i via averaging of log beliefs, ( n ) exp j=1 W ijlog b (t) j (θ) q (t) i (θ) = θ Θ exp ( n j=1 W ijlog b (t) j (θ ) where weight W ij denotes the influence of node j on estimate of node i. Put t = t + 1 and repeat. ),
80 UMich EECS 2015 > Pseudo-Bayes 32 / 48 In a picture Q i (,t) X i (t) f i ( ) local observations Bayesian Update b i (,t)= f i (X i (t) ) P f 0 i (X i (t) )Q i (,t) Average log-beliefs Q i (,t+ 1) = messages from {b j (,t)} neighbors Pn exp j=1 W ij log b j (,t) P 0 2 exp Pn. j=1 W ij log b j ( 0,t)
81 UMich EECS 2015 > Pseudo-Bayes 33 / 48 An example
82 UMich EECS 2015 > Pseudo-Bayes 33 / 48 An example When connected in a network, using the proposed learning rule node 1 learns θ.
83 UMich EECS 2015 > Pseudo-Bayes 34 / 48 Assumptions Assumption 1 For every pair θ θ, f i ( ; θ ) f i ( ; θ) for at least one node, i.e the KL-divergence D (f i ( ; θ ) f i ( ; θ)) > 0.
84 UMich EECS 2015 > Pseudo-Bayes 34 / 48 Assumptions Assumption 1 For every pair θ θ, f i ( ; θ ) f i ( ; θ) for at least one node, i.e the KL-divergence D (f i ( ; θ ) f i ( ; θ)) > 0. Assumption 2 The stochastic matrix W is irreducible.
85 UMich EECS 2015 > Pseudo-Bayes 34 / 48 Assumptions Assumption 1 For every pair θ θ, f i ( ; θ ) f i ( ; θ) for at least one node, i.e the KL-divergence D (f i ( ; θ ) f i ( ; θ)) > 0. Assumption 2 The stochastic matrix W is irreducible. Assumption 3 For all i [n], the initial estimate q (0) i (θ) > 0 for every θ Θ.
86 UMich EECS 2015 > Pseudo-Bayes 35 / 48 Convergence Results Let θ be the unknown fixed parameter. Suppose assumptions 1 3 hold. The eigenvector centrality v = [v 1, v 2,..., v n ] is the left eigenvector of W for eigenvalue 1.
87 UMich EECS 2015 > Pseudo-Bayes 35 / 48 Convergence Results Let θ be the unknown fixed parameter. Suppose assumptions 1 3 hold. The eigenvector centrality v = [v 1, v 2,..., v n ] is the left eigenvector of W for eigenvalue 1. Theorem: Rate of rejecting θ θ Every node i s estimate of θ θ almost surely converges to 0 exponentially fast. Mathematically, 1 lim log q(t) t i (θ) = K(θ, θ) P-a.s. t where K(θ, θ) = n j=1 v jd (f j ( ; θ ) f j ( ; θ)).
88 UMich EECS 2015 > Pseudo-Bayes 36 / 48 Example: Network-wide Learning Θ = {θ 1, θ 2, θ 3, θ 4 } and θ = θ 1. If i and j are connected, 1 W ij = degree of node i, otherwise 0. v = [ 1 12, 1 8, 1 12, 1 8, 1 6, 1 8, 1 12, 1 8, 1 12 ].
89 UMich EECS 2015 > Pseudo-Bayes 37 / 48 Example Θ 1 = {θ }, Θ i = Θ i 1 Θ5 = {θ }, Θ i = Θ i 5
90 UMich EECS 2015 > Pseudo-Bayes 38 / 48 Corollaries Theorem: Rate of rejecting θ θ Every node i s estimate of θ θ almost surely converges to 0 exponentially fast. Mathematically, 1 lim log q(t) t i (θ) = K(θ, θ) P-a.s. t where K(θ, θ) = n j=1 v jd (f j ( ; θ ) f j ( ; θ)). Lower bound on rate of convergence to θ For every node i, the rate at which error in the estimate of θ goes to zero can be lower bounded as 1 ( ) lim t t log 1 q (t) i (θ ) = min θ θ K(θ, θ) P-a.s.
91 UMich EECS 2015 > Pseudo-Bayes 39 / 48 Corollaries Lower bound on rate of learning The rate of learning λ across the network can be lower bounded as, λ min θ Θ min θ θ K(θ, θ) P-a.s. where, and λ = lim inf t 1 t log e t, e t = 1 2 n i=1 q (t) i ( ) 1 θ (.) 1 = n i=1 q (t) i θ θ (θ).
92 UMich EECS 2015 > Pseudo-Bayes 40 / 48 Example: Periodicity node 1 can distinguish node 2 can distinguish Θ = {θ 1, θ 2, θ 3, θ 4 } and θ = θ 1. Underlying graph is periodic, ( ) 0 1 W =. 1 0
93 UMich EECS 2015 > Pseudo-Bayes 41 / 48 Example: Networks with Large Mixing Times node 1 can distinguish node 2 can distinguish Θ = {θ 1, θ 2, θ 3, θ 4 } and θ = θ 1. Underlying graph is aperiodic, ( ) W =
94 UMich EECS 2015 > Pseudo-Bayes 42 / 48 Concentration Result Assumption 4 For k [n], X X k, and for any given θ i, θ j Θ such that θ i θ j, is bounded, denoted by L. log f k( ;θ i) f k ( ;θ j) Theorem Under Assumptions 1 4, for every ɛ > 0 there exists a T such that for all t T and for every θ θ and i [n] we have and ( Pr ( Pr log q (t) i log q (t) i ) (θ) (K(θ, θ) ɛ)t γ(ɛ, L, t), ) (θ) (K(θ, θ) + ɛ)t γ( ɛ, L, t), 2 where L is a finite constant and γ(ɛ, L, t) = 2 exp ( ) ɛ2 t 2L 2 d.
95 UMich EECS 2015 > Pseudo-Bayes 43 / 48 Related Work and Contribution Jadbabaie et al. use local Bayesian update of beliefs followed by averaging the beliefs. Show exponential convergence with no closed form of convergence rate. [ 12] Provide an upper bound on learning rate. [ 13] We average the log beliefs instead. Provide a lower bound on learning rate λ. Lower bound on learning rate is greater than the upper bound = Our learning rule converges faster.
96 UMich EECS 2015 > Pseudo-Bayes 43 / 48 Related Work and Contribution Jadbabaie et al. use local Bayesian update of beliefs followed by averaging the beliefs. Show exponential convergence with no closed form of convergence rate. [ 12] Provide an upper bound on learning rate. [ 13] We average the log beliefs instead. Provide a lower bound on learning rate λ. Lower bound on learning rate is greater than the upper bound = Our learning rule converges faster. Shahrampour and Jadbabaie, 13 formulated a stochastic optimization learning problem; obtained a dual-based learning rule for doubly stochastic W, Provide closed-form lower bound on rate of identifying θ. Using our rule we achieve the same lower bound (from corollary 1) min θ θ ( 1 n ) n D(f j( ; θ ) f j( ; θ)). j=1
97 UMich EECS 2015 > Pseudo-Bayes 44 / 48 Related Work and Contribution An update rule similar to ours was used in Rahnama Rad and Tahbaz-Salehi, 2010 to Show that node s belief converges in probability to the true parameter. However, under certain analytic assumptions. For general model and discrete parameter spaces we show almost-sure exponentially fast convergence.
98 UMich EECS 2015 > Pseudo-Bayes 44 / 48 Related Work and Contribution An update rule similar to ours was used in Rahnama Rad and Tahbaz-Salehi, 2010 to Show that node s belief converges in probability to the true parameter. However, under certain analytic assumptions. For general model and discrete parameter spaces we show almost-sure exponentially fast convergence. Shahrampour et. al. and Nedic et. al. (independently) showed that our learning rule coincides with distributed stochastic optimization based learning rule (W irreducible and aperiodic)
99 UMich EECS 2015 > Summing up 45 / 48 Social sampling to estimate histograms X i i W ij trust j X j i Y j (t) update j W ji Y i (t) Qi(0) Qj(0) Qi(1) Qj(1) i Y j (t) messages j i W ij t!1 j Y i (t) W ji Qi(0) Qj(0) Q i(t) Q j(t) Simple model of randomized message exchange. Unified analysis captures different qualitative behaviors. Censoring rule to achieve consensus to true histogram.
100 UMich EECS 2015 > Summing up 46 / 48 Hypothesis testing and semi-bayes Q i (t) Get Local Send Social Q i (t + 1) Data Update Msg Update X i (t) {Y j (t)} Combination of local Bayesian updates and averaging. Network divergence: an intuitive measure for the rate of convergence. Posterior consistency gives a Bayesio-frequentist analysis.
101 UMich EECS 2015 > Summing up 47 / 48 Looking forward i 4 Continuous distributions and parameters. Applications to distributed optimization. Time-varying case.
102 UMich EECS 2015 > Summing up 48 / 48 Thank You!
Consensus and Distributed Inference Rates Using Network Divergence
DIMACS August 2017 1 / 26 Consensus and Distributed Inference Rates Using Network Divergence Anand D. Department of Electrical and Computer Engineering, The State University of New Jersey August 23, 2017
More informationAsymptotics, asynchrony, and asymmetry in distributed consensus
DANCES Seminar 1 / Asymptotics, asynchrony, and asymmetry in distributed consensus Anand D. Information Theory and Applications Center University of California, San Diego 9 March 011 Joint work with Alex
More informationarxiv: v1 [math.oc] 24 Dec 2018
On Increasing Self-Confidence in Non-Bayesian Social Learning over Time-Varying Directed Graphs César A. Uribe and Ali Jadbabaie arxiv:82.989v [math.oc] 24 Dec 28 Abstract We study the convergence of the
More informationInformation Aggregation in Complex Dynamic Networks
The Combined 48 th IEEE Conference on Decision and Control and 28 th Chinese Control Conference Information Aggregation in Complex Dynamic Networks Ali Jadbabaie Skirkanich Associate Professor of Innovation
More informationDistributed Optimization over Networks Gossip-Based Algorithms
Distributed Optimization over Networks Gossip-Based Algorithms Angelia Nedić angelia@illinois.edu ISE Department and Coordinated Science Laboratory University of Illinois at Urbana-Champaign Outline Random
More informationDISTRIBUTED GENERALIZED LIKELIHOOD RATIO TESTS : FUNDAMENTAL LIMITS AND TRADEOFFS. Anit Kumar Sahu and Soummya Kar
DISTRIBUTED GENERALIZED LIKELIHOOD RATIO TESTS : FUNDAMENTAL LIMITS AND TRADEOFFS Anit Kumar Sahu and Soummya Kar Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh
More informationConvergence Rates in Decentralized Optimization. Alex Olshevsky Department of Electrical and Computer Engineering Boston University
Convergence Rates in Decentralized Optimization Alex Olshevsky Department of Electrical and Computer Engineering Boston University Distributed and Multi-agent Control Strong need for protocols to coordinate
More informationDistributed Optimization over Random Networks
Distributed Optimization over Random Networks Ilan Lobel and Asu Ozdaglar Allerton Conference September 2008 Operations Research Center and Electrical Engineering & Computer Science Massachusetts Institute
More informationBayesian Learning in Social Networks
Bayesian Learning in Social Networks Asu Ozdaglar Joint work with Daron Acemoglu, Munther Dahleh, Ilan Lobel Department of Electrical Engineering and Computer Science, Department of Economics, Operations
More informationHegselmann-Krause Dynamics: An Upper Bound on Termination Time
Hegselmann-Krause Dynamics: An Upper Bound on Termination Time B. Touri Coordinated Science Laboratory University of Illinois Urbana, IL 680 touri@illinois.edu A. Nedić Industrial and Enterprise Systems
More informationBAYESIAN ESTIMATION OF UNKNOWN PARAMETERS OVER NETWORKS
BAYESIAN ESTIMATION OF UNKNOWN PARAMETERS OVER NETWORKS Petar M. Djurić Dept. of Electrical & Computer Engineering Stony Brook University Stony Brook, NY 11794, USA e-mail: petar.djuric@stonybrook.edu
More informationA Note to Robustness Analysis of the Hybrid Consensus Protocols
American Control Conference on O'Farrell Street, San Francisco, CA, USA June 9 - July, A Note to Robustness Analysis of the Hybrid Consensus Protocols Haopeng Zhang, Sean R Mullen, and Qing Hui Abstract
More informationQuantized Average Consensus on Gossip Digraphs
Quantized Average Consensus on Gossip Digraphs Hideaki Ishii Tokyo Institute of Technology Joint work with Kai Cai Workshop on Uncertain Dynamical Systems Udine, Italy August 25th, 2011 Multi-Agent Consensus
More informationALMOST SURE CONVERGENCE OF RANDOM GOSSIP ALGORITHMS
ALMOST SURE CONVERGENCE OF RANDOM GOSSIP ALGORITHMS Giorgio Picci with T. Taylor, ASU Tempe AZ. Wofgang Runggaldier s Birthday, Brixen July 2007 1 CONSENSUS FOR RANDOM GOSSIP ALGORITHMS Consider a finite
More informationEconomics of Networks Social Learning
Economics of Networks Social Learning Evan Sadler Massachusetts Institute of Technology Evan Sadler Social Learning 1/38 Agenda Recap of rational herding Observational learning in a network DeGroot learning
More informationAsymmetric Randomized Gossip Algorithms for Consensus
Asymmetric Randomized Gossip Algorithms for Consensus F. Fagnani S. Zampieri Dipartimento di Matematica, Politecnico di Torino, C.so Duca degli Abruzzi, 24, 10129 Torino, Italy, (email: fabio.fagnani@polito.it).
More informationDistributed Estimation and Detection for Smart Grid
Distributed Estimation and Detection for Smart Grid Texas A&M University Joint Wor with: S. Kar (CMU), R. Tandon (Princeton), H. V. Poor (Princeton), and J. M. F. Moura (CMU) 1 Distributed Estimation/Detection
More information6.207/14.15: Networks Lectures 22-23: Social Learning in Networks
6.207/14.15: Networks Lectures 22-23: Social Learning in Networks Daron Acemoglu and Asu Ozdaglar MIT December 2 end 7, 2009 1 Introduction Outline Recap on Bayesian social learning Non-Bayesian (myopic)
More informationReview. DS GA 1002 Statistical and Mathematical Models. Carlos Fernandez-Granda
Review DS GA 1002 Statistical and Mathematical Models http://www.cims.nyu.edu/~cfgranda/pages/dsga1002_fall16 Carlos Fernandez-Granda Probability and statistics Probability: Framework for dealing with
More informationIntroduction to Machine Learning CMU-10701
Introduction to Machine Learning CMU-10701 Markov Chain Monte Carlo Methods Barnabás Póczos & Aarti Singh Contents Markov Chain Monte Carlo Methods Goal & Motivation Sampling Rejection Importance Markov
More informationDiscrete-time Consensus Filters on Directed Switching Graphs
214 11th IEEE International Conference on Control & Automation (ICCA) June 18-2, 214. Taichung, Taiwan Discrete-time Consensus Filters on Directed Switching Graphs Shuai Li and Yi Guo Abstract We consider
More informationE-Companion to The Evolution of Beliefs over Signed Social Networks
OPERATIONS RESEARCH INFORMS E-Companion to The Evolution of Beliefs over Signed Social Networks Guodong Shi Research School of Engineering, CECS, The Australian National University, Canberra ACT 000, Australia
More informationBayesian estimation of the discrepancy with misspecified parametric models
Bayesian estimation of the discrepancy with misspecified parametric models Pierpaolo De Blasi University of Torino & Collegio Carlo Alberto Bayesian Nonparametrics workshop ICERM, 17-21 September 2012
More informationCOMP90051 Statistical Machine Learning
COMP90051 Statistical Machine Learning Semester 2, 2017 Lecturer: Trevor Cohn 17. Bayesian inference; Bayesian regression Training == optimisation (?) Stages of learning & inference: Formulate model Regression
More informationConvergence of Rule-of-Thumb Learning Rules in Social Networks
Convergence of Rule-of-Thumb Learning Rules in Social Networks Daron Acemoglu Department of Economics Massachusetts Institute of Technology Cambridge, MA 02142 Email: daron@mit.edu Angelia Nedić Department
More informationBayesian Learning in Undirected Graphical Models
Bayesian Learning in Undirected Graphical Models Zoubin Ghahramani Gatsby Computational Neuroscience Unit University College London, UK http://www.gatsby.ucl.ac.uk/ Work with: Iain Murray and Hyun-Chul
More informationQuantized average consensus via dynamic coding/decoding schemes
Proceedings of the 47th IEEE Conference on Decision and Control Cancun, Mexico, Dec 9-, 2008 Quantized average consensus via dynamic coding/decoding schemes Ruggero Carli Francesco Bullo Sandro Zampieri
More information13 : Variational Inference: Loopy Belief Propagation and Mean Field
10-708: Probabilistic Graphical Models 10-708, Spring 2012 13 : Variational Inference: Loopy Belief Propagation and Mean Field Lecturer: Eric P. Xing Scribes: Peter Schulam and William Wang 1 Introduction
More informationRate of Convergence of Learning in Social Networks
Rate of Convergence of Learning in Social Networks Ilan Lobel, Daron Acemoglu, Munther Dahleh and Asuman Ozdaglar Massachusetts Institute of Technology Cambridge, MA Abstract We study the rate of convergence
More informationDistributed Particle Filters: Stability Results and Graph-based Compression of Weighted Particle Clouds
Distributed Particle Filters: Stability Results and Graph-based Compression of Weighted Particle Clouds Michael Rabbat Joint with: Syamantak Datta Gupta and Mark Coates (McGill), and Stephane Blouin (DRDC)
More informationFinite-Time Distributed Consensus in Graphs with Time-Invariant Topologies
Finite-Time Distributed Consensus in Graphs with Time-Invariant Topologies Shreyas Sundaram and Christoforos N Hadjicostis Abstract We present a method for achieving consensus in distributed systems in
More informationAgreement algorithms for synchronization of clocks in nodes of stochastic networks
UDC 519.248: 62 192 Agreement algorithms for synchronization of clocks in nodes of stochastic networks L. Manita, A. Manita National Research University Higher School of Economics, Moscow Institute of
More informationSocial network analysis: social learning
Social network analysis: social learning Donglei Du (ddu@unb.edu) Faculty of Business Administration, University of New Brunswick, NB Canada Fredericton E3B 9Y2 October 20, 2016 Donglei Du (UNB) AlgoTrading
More informationBayesian Learning in Undirected Graphical Models
Bayesian Learning in Undirected Graphical Models Zoubin Ghahramani Gatsby Computational Neuroscience Unit University College London, UK http://www.gatsby.ucl.ac.uk/ and Center for Automated Learning and
More informationAverage-Consensus of Multi-Agent Systems with Direct Topology Based on Event-Triggered Control
Outline Background Preliminaries Consensus Numerical simulations Conclusions Average-Consensus of Multi-Agent Systems with Direct Topology Based on Event-Triggered Control Email: lzhx@nankai.edu.cn, chenzq@nankai.edu.cn
More informationON SPATIAL GOSSIP ALGORITHMS FOR AVERAGE CONSENSUS. Michael G. Rabbat
ON SPATIAL GOSSIP ALGORITHMS FOR AVERAGE CONSENSUS Michael G. Rabbat Dept. of Electrical and Computer Engineering McGill University, Montréal, Québec Email: michael.rabbat@mcgill.ca ABSTRACT This paper
More informationSTAT 518 Intro Student Presentation
STAT 518 Intro Student Presentation Wen Wei Loh April 11, 2013 Title of paper Radford M. Neal [1999] Bayesian Statistics, 6: 475-501, 1999 What the paper is about Regression and Classification Flexible
More informationDistributed online optimization over jointly connected digraphs
Distributed online optimization over jointly connected digraphs David Mateos-Núñez Jorge Cortés University of California, San Diego {dmateosn,cortes}@ucsd.edu Mathematical Theory of Networks and Systems
More informationConsensus of Hybrid Multi-agent Systems
Consensus of Hybrid Multi-agent Systems Yuanshi Zheng, Jingying Ma, and Long Wang arxiv:52.0389v [cs.sy] 0 Dec 205 Abstract In this paper, we consider the consensus problem of hybrid multi-agent system.
More informationADMM and Fast Gradient Methods for Distributed Optimization
ADMM and Fast Gradient Methods for Distributed Optimization João Xavier Instituto Sistemas e Robótica (ISR), Instituto Superior Técnico (IST) European Control Conference, ECC 13 July 16, 013 Joint work
More informationIntroduction to Systems Analysis and Decision Making Prepared by: Jakub Tomczak
Introduction to Systems Analysis and Decision Making Prepared by: Jakub Tomczak 1 Introduction. Random variables During the course we are interested in reasoning about considered phenomenon. In other words,
More informationConsensus-Based Distributed Optimization with Malicious Nodes
Consensus-Based Distributed Optimization with Malicious Nodes Shreyas Sundaram Bahman Gharesifard Abstract We investigate the vulnerabilities of consensusbased distributed optimization protocols to nodes
More informationFinite-Field Consensus
Finite-Field Consensus Fabio Pasqualetti, Domenica Borra, and Francesco Bullo Abstract This work studies consensus networks over finite fields, where agents process and communicate values from the set
More informationZero-Gradient-Sum Algorithms for Distributed Convex Optimization: The Continuous-Time Case
0 American Control Conference on O'Farrell Street, San Francisco, CA, USA June 9 - July 0, 0 Zero-Gradient-Sum Algorithms for Distributed Convex Optimization: The Continuous-Time Case Jie Lu and Choon
More informationStatistical Measures of Uncertainty in Inverse Problems
Statistical Measures of Uncertainty in Inverse Problems Workshop on Uncertainty in Inverse Problems Institute for Mathematics and Its Applications Minneapolis, MN 19-26 April 2002 P.B. Stark Department
More informationPriors for the frequentist, consistency beyond Schwartz
Victoria University, Wellington, New Zealand, 11 January 2016 Priors for the frequentist, consistency beyond Schwartz Bas Kleijn, KdV Institute for Mathematics Part I Introduction Bayesian and Frequentist
More informationDecentralized Stochastic Control with Partial Sharing Information Structures: A Common Information Approach
Decentralized Stochastic Control with Partial Sharing Information Structures: A Common Information Approach 1 Ashutosh Nayyar, Aditya Mahajan and Demosthenis Teneketzis Abstract A general model of decentralized
More informationThe Particle Filter. PD Dr. Rudolph Triebel Computer Vision Group. Machine Learning for Computer Vision
The Particle Filter Non-parametric implementation of Bayes filter Represents the belief (posterior) random state samples. by a set of This representation is approximate. Can represent distributions that
More informationBayesian Machine Learning
Bayesian Machine Learning Andrew Gordon Wilson ORIE 6741 Lecture 4 Occam s Razor, Model Construction, and Directed Graphical Models https://people.orie.cornell.edu/andrew/orie6741 Cornell University September
More informationConvergence Rate for Consensus with Delays
Convergence Rate for Consensus with Delays Angelia Nedić and Asuman Ozdaglar October 8, 2007 Abstract We study the problem of reaching a consensus in the values of a distributed system of agents with time-varying
More informationSTAT 499/962 Topics in Statistics Bayesian Inference and Decision Theory Jan 2018, Handout 01
STAT 499/962 Topics in Statistics Bayesian Inference and Decision Theory Jan 2018, Handout 01 Nasser Sadeghkhani a.sadeghkhani@queensu.ca There are two main schools to statistical inference: 1-frequentist
More informationAn Upper Bound on the Convergence Time for Quantized Consensus
An Upper Bound on the Convergence Time for Quantized Consensus 1 Shang Shang, Paul W. Cuff, Pan Hui and Sanjeev R. Kulkarni Department of Electrical Engineering, Princeton University arxiv:1208.0788v2
More informationComputer Vision Group Prof. Daniel Cremers. 14. Sampling Methods
Prof. Daniel Cremers 14. Sampling Methods Sampling Methods Sampling Methods are widely used in Computer Science as an approximation of a deterministic algorithm to represent uncertainty without a parametric
More informationBayesian Methods for Machine Learning
Bayesian Methods for Machine Learning CS 584: Big Data Analytics Material adapted from Radford Neal s tutorial (http://ftp.cs.utoronto.ca/pub/radford/bayes-tut.pdf), Zoubin Ghahramni (http://hunch.net/~coms-4771/zoubin_ghahramani_bayesian_learning.pdf),
More informationInformation Cascades in Social Networks via Dynamic System Analyses
Information Cascades in Social Networks via Dynamic System Analyses Shao-Lun Huang and Kwang-Cheng Chen Graduate Institute of Communication Engineering, National Taiwan University, Taipei, Taiwan Email:{huangntu,
More informationSensor Tasking and Control
Sensor Tasking and Control Sensing Networking Leonidas Guibas Stanford University Computation CS428 Sensor systems are about sensing, after all... System State Continuous and Discrete Variables The quantities
More informationAlternative Characterization of Ergodicity for Doubly Stochastic Chains
Alternative Characterization of Ergodicity for Doubly Stochastic Chains Behrouz Touri and Angelia Nedić Abstract In this paper we discuss the ergodicity of stochastic and doubly stochastic chains. We define
More informationGAUSSIAN PROCESS TRANSFORMS
GAUSSIAN PROCESS TRANSFORMS Philip A. Chou Ricardo L. de Queiroz Microsoft Research, Redmond, WA, USA pachou@microsoft.com) Computer Science Department, Universidade de Brasilia, Brasilia, Brazil queiroz@ieee.org)
More informationSpread of (Mis)Information in Social Networks
Spread of (Mis)Information in Social Networks Daron Acemoglu, Asuman Ozdaglar, and Ali ParandehGheibi May 9, 009 Abstract We provide a model to investigate the tension between information aggregation and
More informationDistributed Consensus and Linear Functional Calculation in Networks: An Observability Perspective
Distributed Consensus and Linear Functional Calculation in Networks: An Observability Perspective ABSTRACT Shreyas Sundaram Coordinated Science Laboratory and Dept of Electrical and Computer Engineering
More information13: Variational inference II
10-708: Probabilistic Graphical Models, Spring 2015 13: Variational inference II Lecturer: Eric P. Xing Scribes: Ronghuo Zheng, Zhiting Hu, Yuntian Deng 1 Introduction We started to talk about variational
More informationConditional probabilities and graphical models
Conditional probabilities and graphical models Thomas Mailund Bioinformatics Research Centre (BiRC), Aarhus University Probability theory allows us to describe uncertainty in the processes we model within
More information+ + ( + ) = Linear recurrent networks. Simpler, much more amenable to analytic treatment E.g. by choosing
Linear recurrent networks Simpler, much more amenable to analytic treatment E.g. by choosing + ( + ) = Firing rates can be negative Approximates dynamics around fixed point Approximation often reasonable
More informationConnections Between Cooperative Control and Potential Games Illustrated on the Consensus Problem
Proceedings of the European Control Conference 2007 Kos, Greece, July 2-5, 2007 Connections Between Cooperative Control and Potential Games Illustrated on the Consensus Problem Jason R. Marden, Gürdal
More informationAsymmetric Information Diffusion via Gossiping on Static And Dynamic Networks
Asymmetric Information Diffusion via Gossiping on Static And Dynamic Networks Ercan Yildiz 1, Anna Scaglione 2, Asuman Ozdaglar 3 Cornell University, UC Davis, MIT Abstract In this paper we consider the
More informationConfident Bayesian Sequence Prediction. Tor Lattimore
Confident Bayesian Sequence Prediction Tor Lattimore Sequence Prediction Can you guess the next number? 1, 2, 3, 4, 5,... 3, 1, 4, 1, 5,... 1, 0, 0, 0, 1, 0, 1, 1, 1, 1, 0, 1, 0, 0, 1, 0, 1, 0, 1, 0, 1,
More informationTime Synchronization in WSNs: A Maximum Value Based Consensus Approach
211 5th IEEE Conference on Decision and Control and European Control Conference (CDC-ECC) Orlando, FL, USA, December 12-15, 211 Time Synchronization in WSNs: A Maximum Value Based Consensus Approach Jianping
More informationAsymptotic Learning on Bayesian Social Networks
Asymptotic Learning on Bayesian Social Networks Elchanan Mossel Allan Sly Omer Tamuz January 29, 2014 Abstract Understanding information exchange and aggregation on networks is a central problem in theoretical
More informationProbability Models of Information Exchange on Networks Lecture 3
Probability Models of Information Exchange on Networks Lecture 3 Elchanan Mossel UC Berkeley All Right Reserved The Bayesian View of the Jury Theorem Recall: we assume 0/1 with prior probability (0.5,0.5).
More informationIntroduction to Bayesian Statistics
Bayesian Parameter Estimation Introduction to Bayesian Statistics Harvey Thornburg Center for Computer Research in Music and Acoustics (CCRMA) Department of Music, Stanford University Stanford, California
More informationA Gossip Algorithm for Aggregative Games on Graphs
A Gossip Algorithm for Aggregative Games on Graphs Jayash Koshal, Angelia Nedić and Uday V. Shanbhag Abstract We consider a class of games, termed as aggregative games, being played over a distributed
More informationFoundations of Nonparametric Bayesian Methods
1 / 27 Foundations of Nonparametric Bayesian Methods Part II: Models on the Simplex Peter Orbanz http://mlg.eng.cam.ac.uk/porbanz/npb-tutorial.html 2 / 27 Tutorial Overview Part I: Basics Part II: Models
More informationProbabilistic and Bayesian Machine Learning
Probabilistic and Bayesian Machine Learning Day 4: Expectation and Belief Propagation Yee Whye Teh ywteh@gatsby.ucl.ac.uk Gatsby Computational Neuroscience Unit University College London http://www.gatsby.ucl.ac.uk/
More informationLecture : Probabilistic Machine Learning
Lecture : Probabilistic Machine Learning Riashat Islam Reasoning and Learning Lab McGill University September 11, 2018 ML : Many Methods with Many Links Modelling Views of Machine Learning Machine Learning
More informationApril 20th, Advanced Topics in Machine Learning California Institute of Technology. Markov Chain Monte Carlo for Machine Learning
for for Advanced Topics in California Institute of Technology April 20th, 2017 1 / 50 Table of Contents for 1 2 3 4 2 / 50 History of methods for Enrico Fermi used to calculate incredibly accurate predictions
More informationPattern Recognition and Machine Learning. Bishop Chapter 2: Probability Distributions
Pattern Recognition and Machine Learning Chapter 2: Probability Distributions Cécile Amblard Alex Kläser Jakob Verbeek October 11, 27 Probability Distributions: General Density Estimation: given a finite
More informationMachine Learning. Lecture 4: Regularization and Bayesian Statistics. Feng Li. https://funglee.github.io
Machine Learning Lecture 4: Regularization and Bayesian Statistics Feng Li fli@sdu.edu.cn https://funglee.github.io School of Computer Science and Technology Shandong University Fall 207 Overfitting Problem
More informationMulti-Robotic Systems
CHAPTER 9 Multi-Robotic Systems The topic of multi-robotic systems is quite popular now. It is believed that such systems can have the following benefits: Improved performance ( winning by numbers ) Distributed
More informationAlmost Sure Convergence to Consensus in Markovian Random Graphs
Proceedings of the 47th IEEE Conference on Decision and Control Cancun, Mexico, Dec. 9-11, 2008 Almost Sure Convergence to Consensus in Markovian Random Graphs Ion Matei, Nuno Martins and John S. Baras
More informationDistributed Average Consensus with Dithered Quantization
1 Distributed Average Consensus with Dithered Quantization Tuncer C. Aysal, Mark J. Coates, and Michael G. Rabbat Telecommunications and Signal Processing Computer Networks Lab. Department of Electrical
More informationAccuracy and Decision Time for Decentralized Implementations of the Sequential Probability Ratio Test
21 American Control Conference Marriott Waterfront, Baltimore, MD, USA June 3-July 2, 21 ThA1.3 Accuracy Decision Time for Decentralized Implementations of the Sequential Probability Ratio Test Sra Hala
More informationarxiv: v1 [math.oc] 29 Sep 2018
Distributed Finite-time Least Squares Solver for Network Linear Equations Tao Yang a, Jemin George b, Jiahu Qin c,, Xinlei Yi d, Junfeng Wu e arxiv:856v mathoc 9 Sep 8 a Department of Electrical Engineering,
More informationSC7/SM6 Bayes Methods HT18 Lecturer: Geoff Nicholls Lecture 2: Monte Carlo Methods Notes and Problem sheets are available at http://www.stats.ox.ac.uk/~nicholls/bayesmethods/ and via the MSc weblearn pages.
More informationEECS 750. Hypothesis Testing with Communication Constraints
EECS 750 Hypothesis Testing with Communication Constraints Name: Dinesh Krithivasan Abstract In this report, we study a modification of the classical statistical problem of bivariate hypothesis testing.
More informationMachine Learning for Data Science (CS4786) Lecture 24
Machine Learning for Data Science (CS4786) Lecture 24 Graphical Models: Approximate Inference Course Webpage : http://www.cs.cornell.edu/courses/cs4786/2016sp/ BELIEF PROPAGATION OR MESSAGE PASSING Each
More informationZeno-free, distributed event-triggered communication and control for multi-agent average consensus
Zeno-free, distributed event-triggered communication and control for multi-agent average consensus Cameron Nowzari Jorge Cortés Abstract This paper studies a distributed event-triggered communication and
More informationStatistical Inference
Statistical Inference Robert L. Wolpert Institute of Statistics and Decision Sciences Duke University, Durham, NC, USA Spring, 2006 1. DeGroot 1973 In (DeGroot 1973), Morrie DeGroot considers testing the
More informationClassical and Bayesian inference
Classical and Bayesian inference AMS 132 Claudia Wehrhahn (UCSC) Classical and Bayesian inference January 8 1 / 11 The Prior Distribution Definition Suppose that one has a statistical model with parameter
More informationCSC321 Lecture 18: Learning Probabilistic Models
CSC321 Lecture 18: Learning Probabilistic Models Roger Grosse Roger Grosse CSC321 Lecture 18: Learning Probabilistic Models 1 / 25 Overview So far in this course: mainly supervised learning Language modeling
More informationNetwork Optimization with Heuristic Rational Agents
Network Optimization with Heuristic Rational Agents Ceyhun Eksin and Alejandro Ribeiro Department of Electrical and Systems Engineering, University of Pennsylvania {ceksin, aribeiro}@seas.upenn.edu Abstract
More informationConsensus on multi-time Scale digraphs
Consensus on multi-time Scale digraphs Qiong Wu a,, Chenguang Xi b a Department of Mathematics, Tufts University, 503 Boston Ave, Medford, MA 0255 b Department of Electrical and Computer Engineering, Tufts
More informationConsensus, Flocking and Opinion Dynamics
Consensus, Flocking and Opinion Dynamics Antoine Girard Laboratoire Jean Kuntzmann, Université de Grenoble antoine.girard@imag.fr International Summer School of Automatic Control GIPSA Lab, Grenoble, France,
More informationUniformly Most Powerful Bayesian Tests and Standards for Statistical Evidence
Uniformly Most Powerful Bayesian Tests and Standards for Statistical Evidence Valen E. Johnson Texas A&M University February 27, 2014 Valen E. Johnson Texas A&M University Uniformly most powerful Bayes
More informationRECENT years have witnessed an intense interest on distributed
356 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL 6, NO, NOVEMBER 06 Distributed Detection: Finite-Time Analysis and Impact of Network Topology Shahin Shahrampour, Alexander Rakhlin, and Ali Jadbabaie, Fellow,
More informationSTATS 200: Introduction to Statistical Inference. Lecture 29: Course review
STATS 200: Introduction to Statistical Inference Lecture 29: Course review Course review We started in Lecture 1 with a fundamental assumption: Data is a realization of a random process. The goal throughout
More informationDistributed Algorithms for Optimization and Nash-Games on Graphs
Distributed Algorithms for Optimization and Nash-Games on Graphs Angelia Nedić angelia@illinois.edu ECEE Department Arizona State University, Tempe, AZ 0 Large Networked Systems The recent advances in
More informationDecentralized decision making with spatially distributed data
Decentralized decision making with spatially distributed data XuanLong Nguyen Department of Statistics University of Michigan Acknowledgement: Michael Jordan, Martin Wainwright, Ram Rajagopal, Pravin Varaiya
More informationStatistics: Learning models from data
DS-GA 1002 Lecture notes 5 October 19, 2015 Statistics: Learning models from data Learning models from data that are assumed to be generated probabilistically from a certain unknown distribution is a crucial
More informationEfficient Inference in Fully Connected CRFs with Gaussian Edge Potentials
Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials by Phillip Krahenbuhl and Vladlen Koltun Presented by Adam Stambler Multi-class image segmentation Assign a class label to each
More informationLearning in Linear Games over Networks
Fiftieth Annual Allerton Conference Allerton House, UIUC, Illinois, USA October -, 0 Learning in Linear Games over Networks Ceyhun Eksin, Pooya Molavi, Alejandro Ribeiro, and Ali Jadbabaie Dept. of Electrical
More information