Presentation is loading. Please wait.

Presentation is loading. Please wait.

Of 19 03/21/2012CISS: Beliefs in Communication1 Efficient Semantic Communication & Compatible Beliefs Madhu Sudan Microsoft, New England Based on joint.

Similar presentations


Presentation on theme: "Of 19 03/21/2012CISS: Beliefs in Communication1 Efficient Semantic Communication & Compatible Beliefs Madhu Sudan Microsoft, New England Based on joint."— Presentation transcript:

1 of 19 03/21/2012CISS: Beliefs in Communication1 Efficient Semantic Communication & Compatible Beliefs Madhu Sudan Microsoft, New England Based on joint works with: Oded Goldreich (Weizmann) and Brendan Juba (Harvard)

2 of 19 03/21/2012CISS: Beliefs in Communication2 Part I: Background on Semantic Communication

3 of 19 Semantic Communication: Motivation First era of communication: Reliable Wires First era of communication: Reliable Wires Essentially done: wires are very reliable; performance can be enhanced (maybe) quantitatively but not qualitatively. Essentially done: wires are very reliable; performance can be enhanced (maybe) quantitatively but not qualitatively. Can we get endpoints to also be reliable? Can we get endpoints to also be reliable? Modern systems have smart endpoints. Modern systems have smart endpoints. Smart implies capability. Smart implies capability. Smart implies diversity. Smart implies diversity. But diversity implies (potential) misunderstanding. But diversity implies (potential) misunderstanding. Semantic Communication [Goldreich,Juba+S 10] Semantic Communication [Goldreich,Juba+S 10] Detect/Correct Misunderstanding. Detect/Correct Misunderstanding. 03/21/2012CISS: Beliefs in Communication3

4 of 19 Semantic Communication: Model - I 03/21/2012CISS: Beliefs in Communication4 A

5 of 19 Uncertainty about the receiver: Uncertainty about the receiver: (User doesnt know server; vice versa). (User doesnt know server; vice versa). Semantic Communication: Model - II 03/21/2012CISS: Beliefs in Communication5 A B Channel B2B2B2B2 AkAkAkAk A3A3A3A3 A2A2A2A2 A1A1A1A1 B1B1B1B1 B3B3B3B3 BjBjBjBj New Class of Problems New challenges Needs more attention!

6 of 19 Semantic Communication: Model - IIIa Goal-oriented communication: Goal-oriented communication: User attempting to reach some goal. User attempting to reach some goal. How to model this? How to model this? Classical approaches: Classical approaches: Some function of state of user, or some function of transcript of interaction etc. Some function of state of user, or some function of transcript of interaction etc. Fails in semantic/uncertainty setting. Fails in semantic/uncertainty setting. Our [GJS] approach: Introduce a (hypothetical) third agent. Our [GJS] approach: Introduce a (hypothetical) third agent. 03/21/2012CISS: Beliefs in Communication6

7 of 19 Semantic Communication: Model - IIIb March 1, 2011Semantic UCLA7UserServer Referee/Environment

8 of 19 Basic Definitions: Helpfulness, Universality, Sensing What makes a server helpful? What makes a server helpful? S is G-helpful, if there exists a user who can achieve goal (efficiently) for every starting state of S. S is G-helpful, if there exists a user who can achieve goal (efficiently) for every starting state of S. Universality: Universality: User U is universal if it achieves G with every G-helpful server. User U is universal if it achieves G with every G-helpful server. Sensing? Sensing? Roughly, Goal G can be sensed if there exists an efficient algorithm that scan use (with their inputs) to see if Referee will accept. Roughly, Goal G can be sensed if there exists an efficient algorithm that scan use (with their inputs) to see if Referee will accept. 03/21/2012CISS: Beliefs in Communication8

9 of 19 Principal Thesis and Theorem 03/21/2012CISS: Beliefs in Communication9

10 of 19 Proof + Insights Positive results by enumeration. Positive results by enumeration. Negative? Mostly by definition. Negative? Mostly by definition. Insights: Insights: Servers should know how to be interrupted. (How else can they function independent of complexity of their own state?) Servers should know how to be interrupted. (How else can they function independent of complexity of their own state?) Short interrupt signal helps. Short interrupt signal helps. Goals should be senseible. Goals should be senseible. 03/21/2012CISS: Beliefs in Communication10

11 of 19 03/21/2012CISS: Beliefs in Communication11 Part II: Beliefs & Compatibility

12 of 19 Motivation Why does natural (human) communication differ so much from designed communication? Why does natural (human) communication differ so much from designed communication? Languages are ambiguous Languages are ambiguous They violate their own grammatical rules They violate their own grammatical rules They are needlessly redundant at times, and noisily compressed at other times? They are needlessly redundant at times, and noisily compressed at other times? Can we use information theory to explain such phenomena (departures from information theory)? Can we use information theory to explain such phenomena (departures from information theory)? Use fact that natural communication deals with uncertainty about server. Use fact that natural communication deals with uncertainty about server. 03/21/2012CISS: Beliefs in Communication12

13 of 19 Does Semantic Communication help? Pros: Pros: Does deal with uncertainty about servers. Does deal with uncertainty about servers. Cons: Cons: Seems quite inefficient (user is enumerating all servers?). Seems quite inefficient (user is enumerating all servers?). Seems to throw away all knowledge about server (that might yield efficiency). Seems to throw away all knowledge about server (that might yield efficiency). Is universality really a goal? Is it not at odds with use of knowledge Is universality really a goal? Is it not at odds with use of knowledge 03/21/2012CISS: Beliefs in Communication13

14 of 19 Beliefs in Semantic Communication 03/21/2012CISS: Beliefs in Communication14

15 of 19 Compatibility of Beliefs 03/21/2012CISS: Beliefs in Communication15

16 of 19 Server Performance? 03/21/2012CISS: Beliefs in Communication16

17 of 19 Universality of Users under Beliefs 03/21/2012CISS: Beliefs in Communication17

18 of 19 Consequences/Conclusions Universality of communication is not at odds with efficiency. Universality of communication is not at odds with efficiency. Efficiency comes with compatibility of communicating players. Efficiency comes with compatibility of communicating players. Universality takes care of possibility of misunderstanding, at an affordable price. Universality takes care of possibility of misunderstanding, at an affordable price. 03/21/2012CISS: Beliefs in Communication18

19 of 19 Thank You 03/21/2012CISS: Beliefs in Communication19


Download ppt "Of 19 03/21/2012CISS: Beliefs in Communication1 Efficient Semantic Communication & Compatible Beliefs Madhu Sudan Microsoft, New England Based on joint."

Similar presentations


Ads by Google