text
stringlengths 1
6.06k
| source
stringclasses 375
values | page
int64 0
1.49k
| book
stringclasses 375
values | chunk_index
int64 0
0
|
|---|---|---|---|---|
52 Op<unk>mizing support structure Branching Support Stuctures for 3D Prin<unk>ng: Ryan Schmidt and Nobuyuki Umetani These branching supports use 75% less material than the manufacture’s default support structure–reduce print <unk>me/cost. (Note: part orienta<unk>on mazers here too!)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/08_Digital_Fabrication.pdf
| 52
|
08_Digital_Fabrication
| 0
|
53 Minimize prin<unk>ng costs/maximize robustness Op<unk>mize printed design itself minimize printed volume without breaking maximize resilience for given print volume Cost-effec<unk>ve Prin<unk>ng of 3D Objects with Skin-Frame Structures [Wang et al. 2013]
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/08_Digital_Fabrication.pdf
| 53
|
08_Digital_Fabrication
| 0
|
Fabrica<unk>on services Shapeways iMaterialize makexyz 3D Hubs
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/08_Digital_Fabrication.pdf
| 54
|
08_Digital_Fabrication
| 0
|
54 55 Pushing 3D prin<unk>ng to the extremes Large scale: concrete printers Andrey Rudenko - Youtube 0:00 / 1:09
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/08_Digital_Fabrication.pdf
| 55
|
08_Digital_Fabrication
| 0
|
56 Pushing 3D prin<unk>ng to the extremes Small scale: nanophotonics nanoscribe Two photo lithography - nffa.eu microstructure with nanoscribe
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/08_Digital_Fabrication.pdf
| 56
|
08_Digital_Fabrication
| 0
|
Tableau de bord › Mes cours › COM-401 › 20 décembre - 26 décembre › Quiz 6 Question 1 Correct Note de 1,00 sur 1,00 Question 2 Correct Note de 1,00 sur 1,00 Commencé le jeudi, 23 décembre 2021, 10:15 État Terminé Terminé le jeudi, 23 décembre 2021, 10:24 Temps mis 9 min 33 s Note 5,00 sur 5,00 (100%) Let be a PKC with an encryption algorithm and a deterministic decryption algorithm. Tick the incorrect assertion. If is deterministic, then is not IND-CCA. If is probabilistic, then may not be IND-CCA. If is probabilistic, then may be IND-CPA. I do not answer. If is deterministic, then is IND-CPA. PKC Enc Enc PKC Enc PKC Enc PKC Enc PKC ! La réponse correcte est : If is deterministic, then is IND-CPA. Enc PKC Tick the correct assertion. I do not answer. ElGamal signatures have been proved secure for arbitrary public parameters. ElGamal and RSA signatures suffer from the same nonce-reuse attacks. ECDSA does not bring any benefits over DSA. ElGamal signatures suffered from nonce-reuse attacks. ! La réponse correcte est : ElGamal signatures suffered from nonce-reuse attacks. Question 3 Correct Note de 1,00 sur 1,00 Question 4 Correct Note de 1,00 sur 1,00 Question 5 Correct Note de 1,00 sur 1,00 Tick the incorrect assertion about elliptic curves and finite fields I do not answer. Any two isomorphic curves have the same -invariant. is a finite field if and only if . If a curve is defined over such that , its $j$-invariant is also in . A curve and its twist are not isomorphic over any finite dimentional extention of the base field. j [x]/⟨ + 1⟩ Fp x2 p= 3 mod 4 Fq gcd(q, 6) = 1 Fq ! La réponse correcte est : A curve and its twist are not isomorphic over any finite dimentional extention of the base field. Tick the correct assertion about
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/quizz_6.pdf
| 0
|
quizz_6
| 0
|
gcd(q, 6) = 1 Fq ! La réponse correcte est : A curve and its twist are not isomorphic over any finite dimentional extention of the base field. Tick the correct assertion about elliptic curve cryptography. Elliptic curve cryptography requires larger key sizes to provide same level of security. I do not answer. It is easy to pick random curves which are secure for cryptographic use. Singular curves (with ) are good for pairing-based cryptography. Let be an elliptic curve, then each -rational point of can be represented with bits. Δ = 0 E/Fp Fp E log(p) + 1 ! La réponse correcte est : Let be an elliptic curve, then each -rational point of can be represented with bits. E/Fp Fp E log(p) + 1 Tick the correct assertion. I do not answer. Plain RSA encryption is homomorphic ( ). RSA signatures are not message recovering. A 2048-bit RSA modulus is generally considered to be secure against quantum attacks. RSA-PSS is an encryption standard. Enc(xy) = Enc(x)Enc(y) ! La réponse correcte est : Plain RSA encryption is homomorphic ( ). Enc(xy) = Enc(x)Enc(y) ◀ Solution 11 Aller à... Contact EPFL CH-1015 Lausanne +41 21 693 11 11
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/quizz_6.pdf
| 0
|
quizz_6
| 0
|
Demystifying Bitcoin Prof R. Guerraoui EPFL
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 0
|
10. DemystifyingBitcoin
| 0
|
Demystifying Blockchain Bitcoin Ethereum Leader Proof of work Smart contracts Snapshot Broadcast Consensus
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 1
|
10. DemystifyingBitcoin
| 0
|
Perspectives (1) The journalist (2) The user (3) The participant (4) The engineer (5) The scientist
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 2
|
10. DemystifyingBitcoin
| 0
|
(1) The Journalist 2008: Financial crisis – Nakamoto (1/21m) From 1c to 10000$ through 20000$ (16.000$) 2014: Ethereum (CH) - Now 555 $ From trading hardware to general trading
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 3
|
10. DemystifyingBitcoin
| 0
|
Perspectives (1) The journalist (2) The user (3) The participant (4) The engineer (5) The scientist
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 5
|
10. DemystifyingBitcoin
| 0
|
(2) The User
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 6
|
10. DemystifyingBitcoin
| 0
|
(2) The User The wallet: 1 private key + several public keys Transaction validation Signing + gossiping + mining + chaining Transaction commitment After time t: thousands of users have seen it
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 7
|
10. DemystifyingBitcoin
| 0
|
(3) The Participant
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 8
|
10. DemystifyingBitcoin
| 0
|
P vs NP (Nash/GV 50 – Ford 70) ? * ? = 91 7 * 13 = ?
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 9
|
10. DemystifyingBitcoin
| 0
|
(3) The Participant
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 10
|
10. DemystifyingBitcoin
| 0
|
(3) The Participant To validate a transaction, a miner has to solve a puzzle including it Fairness and cooperation Total: 21 millions bitcoins Now: 18 millions Incentive: 6.25 bitcoins / puzzle 50 bitcoins 3 years ago
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 11
|
10. DemystifyingBitcoin
| 0
|
(4) The Engineer Joinning (a P2P network) Gossiping (the transaction) Mining (proof of work - nonce) Chaining (hash) Committing/Aborting Gathering (a block) Signing (a transaction) Gossiping (the block)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 12
|
10. DemystifyingBitcoin
| 0
|
The Big Picture Special Reward TX TX-2 (signed) TX-2452 (signed) ••• Prev # nonce This # Bitcoin block eward TX igned) (signed) ce This # Special Rew TX-2 (sig TX-2325 ( ••• Prev # nonc Mining: find such that < d nonce This # How? By trying different nonces (brute force)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 16
|
10. DemystifyingBitcoin
| 0
|
Perspectives (1) The journalist (2) The user (3) The participant (4) The engineer (5) The scientist
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 18
|
10. DemystifyingBitcoin
| 0
|
(5) The Scientist State Machine Replication (78)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 19
|
10. DemystifyingBitcoin
| 0
|
Every service can be implemented in a highly available manner using Consensus Safety: No two nodes must choose different values. The chosen value must have been proposed by a node. Liveness: Each node must eventually choose a value. Consensus Universality (78)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 20
|
10. DemystifyingBitcoin
| 0
|
Consensus is impossible in an asynchronous system Consensus Impossibility (84)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 21
|
10. DemystifyingBitcoin
| 0
|
« Computing’s central challenge is how not to make a mess of it ...» E. Dijkstra X000 implementations
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 22
|
10. DemystifyingBitcoin
| 0
|
Payment System Can we implement a payment system asynchronously?
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 23
|
10. DemystifyingBitcoin
| 0
|
P vs NP Asynchronous vs Synchronous ? * ? = 91 7 * 13 = ? « To understand a distributed computing problem: bring it to shared memory » T. Lannister Is payment an asynchronous problem?
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 24
|
10. DemystifyingBitcoin
| 0
|
The infinitely big The infinitely small
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 25
|
10. DemystifyingBitcoin
| 0
|
Message Passing p1 p2 p3 Send Receive
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 26
|
10. DemystifyingBitcoin
| 0
|
Shared Memory p1 p2 Write() Read() 1 1 ó Message Passing Registers
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 27
|
10. DemystifyingBitcoin
| 0
|
Atomic Shared Memory p1 p2 p3 write(1) - ok read() - 1 read() - 1
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 28
|
10. DemystifyingBitcoin
| 0
|
Atomic Shared Memory p1 p2 p3 write(1) - ok read() - 1 read() - 0
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 29
|
10. DemystifyingBitcoin
| 0
|
Message Passing ó Shared Memory p1 p2 p3 Write(1) Ok Read() 1 Quorums (asynchrony)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 30
|
10. DemystifyingBitcoin
| 0
|
« Optimization is the source of all evil » D. Knuth « To understand a distributed computing problem: bring it to shared memory » T. Lannister ó
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 31
|
10. DemystifyingBitcoin
| 0
|
P vs NP Asynchronous vs Synchronous ? * ? = 91 7 * 13 = ?
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 32
|
10. DemystifyingBitcoin
| 0
|
Atomicity Wait-freedom Payment System Can we implement a payment system asynchronously?
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 33
|
10. DemystifyingBitcoin
| 0
|
Counter: Specification A counter has two operations inc() and read(); it maintains an integer x init to 0 read(): return(x) inc(): x := x + 1; return(ok)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 34
|
10. DemystifyingBitcoin
| 0
|
Counter: Algorithm The processes share an array of registers Reg[1,..,N] inc(): Reg[i].write(Reg[i].read() +1); return(ok) read(): sum := 0; for j = 1 to N do sum := sum + Reg[j].read(); return(sum)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 35
|
10. DemystifyingBitcoin
| 0
|
Counter*: Specification Counter* has, in addition, operation dec() dec(): if x > 0 then x := x - 1; return(ok) else return(no) Can we implement Counter* asynchronously?
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 36
|
10. DemystifyingBitcoin
| 0
|
2-Consensus with Counter* § Registers R0 and R1 and Counter* C - initialized to 1 § Process pI: § propose(vI) § RI.write(vI) § res := C.dec() § if(res = ok) then ü return(vI) ü else return(R{1-I}.read())
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 37
|
10. DemystifyingBitcoin
| 0
|
Impossibility [FLP85,LA87] § Corollary: no asynchronous algorithm implements Counter* among two processes using registers § Theorem: no asynchronous algorithm implements consensus among two processes using registers
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 38
|
10. DemystifyingBitcoin
| 0
|
Sperner’s Lemma § Theorem: no asynchronous algorithm implements set-agreement using registers
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 39
|
10. DemystifyingBitcoin
| 0
|
§ The consensus number of an object is the maximum number of processes than can solve consensus with it
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 40
|
10. DemystifyingBitcoin
| 0
|
Payment Object (PO): Specification Pay(a,b,x): transfer amount x from a to b if a > x (return ok; else return no) NB. Only the owner of a invokes Pay(a,*,*) § Questions: can PO be implemented asynchronously? what is the consensus number of PO?
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 41
|
10. DemystifyingBitcoin
| 0
|
Snapshot: Specification A snapshot has operations update() and scan(); it maintains an array x of size N scan(): return(x) update(i,v): x[i] := v; return(ok)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 42
|
10. DemystifyingBitcoin
| 0
|
Algorithm? The processes share one array of N registers Reg[1,..,N] scan(): for j = 1 to N do x[j] := Reg[j].read(); return(x) update(i,v): Reg[i].write(v); return(ok)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 43
|
10. DemystifyingBitcoin
| 0
|
Atomicity? p1 p2 p3 update(1,1) - ok scan() - [1,0,2] update(3,2) - ok
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 44
|
10. DemystifyingBitcoin
| 0
|
Atomicity? p1 p2 p3 update(1,1) - ok scan() - [1,0,2] update(3,2) - ok
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 45
|
10. DemystifyingBitcoin
| 0
|
Atomicity? p1 p2 p3 scan() - [0,0,10] update(2,1) - ok update(3,10) - ok
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 46
|
10. DemystifyingBitcoin
| 0
|
Key idea for atomicity To scan, a process keeps reading the entire snapshot (i.e., collecting), until two arrays are the same To update, scan then write the value and the scan Key idea for wait-freedom To scan, a process keeps collecting and returns a collect if it did not change, or some collect returned by a concurrent scan
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 47
|
10. DemystifyingBitcoin
| 0
|
The Payment Object: Algorithm Every process stores the sequence of its outgoing payments in its snapshot location To pay, the process scans, computes its current balance: if bigger than the transfer, updates and returns ok, otherwise returns no To read, scan and return the current balance
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 48
|
10. DemystifyingBitcoin
| 0
|
PO can be implemented Asynchronously Consensus number of PO is 1 Consensus number of PO(k) is k
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 49
|
10. DemystifyingBitcoin
| 0
|
Payment System (AT2) AT2_S AT2_D AT2_R Number of lines of code: one order of magnitude less Latency: seconds (at most)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 50
|
10. DemystifyingBitcoin
| 0
|
Rachid Guerraoui, Petr Kuznetsov, Matteo Monti, Matej Pavlovic, Dragos-Adrian Seredinschi: The Consensus Number of a Cryptocurrency. PODC 2019: 307-316 Rachid Guerraoui, Petr Kuznetsov, Matteo Monti, Matej Pavlovic, Dragos-Adrian Seredinschi: Scalable Byzantine Reliable Broadcast. DISC 2019: 1-16 (Best Paper Award) Daniel Collins, Rachid Guerraoui, Jovan Komatovic, Petr Kuznetsov, Matteo Monti, Matej Pavlovic, Yvonne Anne Pignolet, Dragos-Adrian Seredinschi, Andrei Tonkikh, Athanasios Xygkis: Online Payments by Merely Broadcasting Messages. DSN 2020: 26-38 (Runner for the Best Paper Award) References
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 51
|
10. DemystifyingBitcoin
| 0
|
References
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/10. DemystifyingBitcoin.pdf
| 52
|
10. DemystifyingBitcoin
| 0
|
Resource sharing Giovanni De Micheli Integrated Systems Laboratory This presentation can be used for non-commercial purposes as long as this note and the copyright footers are not removed © Giovanni De Micheli – All rights reserved
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 0
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 2 Module 1 N Objectives L Motivation and problem formulation L Flat and hierarchical graphs L Functional and memory resources L Extension to module selection
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 1
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 3 Allocation and binding N Allocation: L Number of resources available N Binding: L Relation between operations and resources N Sharing: L Many-to-one relation N Optimum binding/sharing: L Minimize the resource usage
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 2
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 4 Binding NLimiting cases: LDedicated resources M One resource per operation M No sharing LOne multi-task resource M ALU LOne resource per type
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 3
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 5 Optimum sharing problem NScheduled sequencing graphs LOperation concurrency is well defined NConsider operation types independently LProblem decomposition LPerform analysis for each resource type
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 4
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 6 Compatibly and conflicts N Operation compatibility: L Same type L Non concurrent N Compatibility graph: L Vertices: operations L Edges: compatibility relation N Conflict graph: L Complement of compatibility graph t1 x=a+b y=c+d 1 2 t2 s=x+y t=x-y 3 4 t3 z=a+t 5 1 2 3 4 5 Compatibility graph Conflict graph 1 2 3 4 5
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 5
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 7 Example t1 x=a+b y=c+d 1 2 t2 s=x+y t=x-y 3 4 t3 z=a+t 5 Conflict 1 2 3 4 5 1 2 3 4 5 Compatibility Partitioning Coloring ALU1: 1,3,5 ALU2: 2,4
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 6
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 8 Compatibility and conflicts N Compatibility graph: L Partition the graph into a minimum number of cliques L Find clique cover number k ( G+ ) N Conflict graph: L Color the vertices by a minimum number of colors. L Find the chromatic number х ( G_ ) N NP-complete problems: L Heuristic algorithms
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 7
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 9 Data-flow graphs (flat sequencing graphs) N The compatibility/conflict graphs have special properties: L Compatibility M Comparability graph L Conflict M Interval graph N Polynomial time solutions: L Golumbic’s algorithm L Left-edge algorithm
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 8
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 10 Perfect graphs N Comparability graph: L Graph G (V, E ) has an orientation G ( V, F ) with the transitive property (vi, vj) є F and (vj, vk) є F →(vi, vk) є F N Interval graph: L Vertices correspond to intervals L Edges correspond to interval intersection L Subset of chordal graphs M Every loop with more than three edges has a chord
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 9
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 11 Example TIME 1 TIME 2 TIME 3 TIME 4 * * + < - - * * * * + NOP NOP 0 1 2 3 4 5 6 7 8 9 10 11 n 3 1 8 7 6 2 4 10 5 11 9
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 10
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 12 Example TIME 1 TIME 2 TIME 3 TIME 4 * * + < - - * * * * + NOP NOP 0 1 2 3 4 5 6 7 8 9 10 11 n 1 4 5 9 10 2 3 6 7 8 11
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 11
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 13 Left-edge algorithm NInput: LSet of intervals with left and right edge M Start and Stop times LA set of colors (initially one color) NRationale: LSort intervals in a list by left edge LAssign non overlapping intervals to first color using the list LWhen possible intervals are exhausted, increase color counter and repeat
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 12
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 14 Example 0 1 2 3 4 5 6 7 1 6 4 7 8 2 3 5 1 0 1 2 3 4 5 6 7 8 2 3 6 7 5 4 1 6 7 4 2 3 5 Conflict graph Intervals 6 7 4 2 1 3 5 Colored conflict graph Coloring
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 13
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 15 Left-edge algorithm LEFT_EDGE(I) { Sort elements of I in a list L in ascending order of li; c = 0; while (some interval has not been colored) do { S = Ø; r = 0; while ( exists s є L such that ls > r) do { s = First element in the list L with ls > r; S = S U {s}; r = rs; Delete s from L; } c = c + 1; Label elements of S with color c; } }
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 14
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 17 Hierarchical sequencing graphs NHierarchical conflict/compatibility graphs: LEasy to compute LPrevent sharing across hierarchy NFlatten hierarchy: LBigger graphs LDestroy nice properties
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 15
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 18 Example TIME 1 TIME 2 TIME 3 TIME 4 TIME 5 TIME 6 TIME 7 a a * * * 2 3 4 a 2 3 4 a + a * a + a * 2 * 3 * 4 * (a) (b) (c)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 16
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 19 Example (a) (b) (c) a d c b a b c d TIME 1 TIME 2 TIME 3 TIME 4 a BR c b NOP NOP d NOP NOP
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 17
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 20 Register binding problem N Given a schedule: L Lifetime intervals for variables L Lifetime overlaps N Conflict graph (interval graph): L Vertices variables L Edges overlaps L Interval graph N Compatibility graph (comparability graph): L Complement of conflict graph
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 18
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 21 Register sharing in data-flow graphs N Given: L Variable lifetime conflict graph N Find: L Minimum number of registers storing all the variables N Key point: L Interval graph M Left-edge algorithm (polynomial-time complexity)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 19
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 22 Example * * * * * - - TIME 1 TIME 2 TIME 3 TIME 4 1 2 3 4 5 6 7 z1 z2 z3 z4 z5 z6 z1 z3 z5 z2 z4 z6 z1 z2 z3 z4 z5 z6 (a) (b) (c)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 20
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 23 Register sharing general case N Iterative conflicts: L Preserve values across iterations L Circular-arc conflict graph M Coloring is intractable N Hierarchical graphs: L General conflict graphs M Coloring is intractable N Heuristic algorithms
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 21
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 24 Example TIME 1 TIME 2 TIME 3 TIME 4 < * * * * + * * + - - 3 x u dx 3 y u dx x dx dx y u u y c a 4 5 3 7 9 1 2 6 8 10 11 z1 z2 z3 z4 z5 z6 z7 x y u z1 z2 z3 z4 z5 z6 u y u y z7 x x (a) (b)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 22
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 25 Example Variable-lifetimes and circular-arc conflict graph z1 z2 z3 z4 z5 z6 u z7 x y x 1 2 3 4 z5 z6 z7 z4 z3 z1 z2 u y
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 23
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 30 Bus sharing and binding NFind the minimum number of busses to accommodate all data transfer NFind the maximum number of data transfers for a fixed number of busses NSimilar to memory binding problem
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 24
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 31 Example NOne bus: L3 variables can be transferred NTwo busses: LAll variables can be transferred * * * * * - - TIME 1 TIME 2 TIME 3 TIME 4 1 2 3 4 5 6 7 z1 z2 z3 z4 z5 z6 z1 z3 z5 z2 z4 z6 z1 z2 z3 z4 z5 z6 (a) (b) (c)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 25
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 32 Module selection problem N Extension of resource sharing L Library of resources: L More than one resource per type N Example: L Ripple-carry adder (small and slow) L Carry look-ahead adder (big and fast) N Resource modeling: L Resource subtypes with M ( area, delay ) parameters
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 26
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 33 Module selection solution N ILP formulation: L Decision variables M Select resource sub-type M Determine ( area, delay ) N Heuristic algorithm L Determine minimum latency with fastest resource subtypes L Recover area by using slower resources on non-critical paths
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 27
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 34 Example N Multipliers with: L ( Area, delay ) = ( 5,1 ) and ( 2,2 ) N Latency bound of 5 * * + < - - * * * * + NOP NOP 0 1 2 3 4 5 6 7 8 9 10 11 n TIME 1 TIME 2 TIME 3 TIME 4 TIME 5 (1,1) (1,2) (2,1) (2,2) Slow multipliers save area
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 28
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 35 Example 2 N Latency bound of 4 L Fast multipliers for { v1 , v2 , v3 } L Slower multiplier can be used elsewhere M Less sharing N Minimum-latency design uses fast multipliers only L Impossible to use slow multipliers * * + < - - * * * * + NOP NOP 0 1 2 3 4 5 6 7 8 9 10 11 n TIME 1 TIME 2 TIME 3 TIME 4 (1,1) (1,2) (2,1) (2,2)
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 29
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 36 Module 2 N Objectives L Data path generation L Control synthesis
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 30
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 37 Data path synthesis NApplied after resource binding NConnectivity synthesis: LConnection of resources to multiplexers busses and registers LControl unit interface LI/O ports NPhysical data path synthesis LSpecific techniques for regular datapath design M Regularity extraction
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 31
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 38 Example * ALU DATA-PATH CONTROL-UNIT r2 r1 u y x dx 3 a REGISTERS enable Mux control ALU control (+,-,<) c
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 32
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 39 Control synthesis NSynthesis of the control unit NLogic model: LSynchronous FSM NPhysical implementation: LHard-wired or distributed FSM LMicrocode
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 33
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 40 Example * * * - - * * * + + < NOP NOP 1 2 3 4 5 6 7 8 9 10 11 0 n TIME 1 TIME 2 TIME 3 TIME 4 reset reset’ 4 s1 s4 s3 s2 reset’ reset 5 reset reset’ 1,2,6,8,10 3,7,9,11
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 34
|
DT6 (sharing)
| 0
|
(c) Giovanni De Micheli 41 Summary NResource sharing is reducible to vertex coloring or to clique covering: LSimple for flat graphs LIntractable, but still easy in practice, for other graphs LResource sharing has several extensions: M Module selection NData path design and control synthesis are conceptually simple but still important steps in synthesis LGenerated data path is an interconnection of blocks LControl is one or more finite-state machines
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/DT6 (sharing).pdf
| 35
|
DT6 (sharing)
| 0
|
Contents Contents 1 1 Speech Recognition 1 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 ASR statistical formalism . . . . . . . . . . . . . . . . . . 4 1.3 Pattern classification with realistic data . . . . . . . . . . 7 1.4 Finite State Automata . . . . . . . . . . . . . . . . . . . . 9 1.4.1 Deterministic finite state automata . . . . . . . . . 10 1.4.2 Stochastic finite state automata and Markov models 11 1.4.3 Hidden Markov models (HMM) . . . . . . . . . . . 15 1.4.4 Hybrid HMM/ANN systems . . . . . . . . . . . . 21 1.5 Acoustic Modeling . . . . . . . . . . . . . . . . . . . . . . 23 1.5.1 Acoustic model . . . . . . . . . . . . . . . . . . . . 23 1.5.2 Acoustic model (HMM) training . . . . . . . . . . 24 1.6 Language modeling . . . . . . . . . . . . . . . . . . . . . . 26 1.6.1 Language model . . . . . . . . . . . . . .
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/3 ppur_asr.pdf
| 0
|
3 ppur_asr
| 0
|
. . . . . . . . 26 1.6.1 Language model . . . . . . . . . . . . . . . . . . . 26 1.6.2 Language model raining . . . . . . . . . . . . . . . 28 1.7 Decoding . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 2 Speaker Verification 35 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 35 2.2 Acoustic parameters . . . . . . . . . . . . . . . . . . . . . 36 2.3 Similarity measures . . . . . . . . . . . . . . . . . . . . . . 37 2.4 Text-dependent speaker verification . . . . . . . . . . . . . 40 2.5 Text-independent speaker verification . . . . . . . . . . . 41 2.6 Text prompted speaker verification . . . . . . . . . . . . . 43 2.7 Identification, verification and the decision threshold . . . 44 Bibliography . . . . . . . .
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/3 ppur_asr.pdf
| 0
|
3 ppur_asr
| 0
|
. . . . . . 43 2.7 Identification, verification and the decision threshold . . . 44 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/3 ppur_asr.pdf
| 0
|
3 ppur_asr
| 0
|
Chapter 1 Speech Recognition 1.1 Introduction Automatic speech recognition (ASR) attempts to use a machine to de- rive the linguistic message from a speech signal. The linguistic message originates in a speaker’s mind, and the speaker converts it to speech sounds, considering the intended recipient of the message and any feed- back the recipient might be providing. Human listeners can generally de- termine a functional equivalent to the speaker’s intended message quite effortlessly. However, this task represents a difficult challenge for a ma- chine. Unrestricted ASR would most likely require emulation of human in- telligence. While some believe this can be achieved in the foreseeable future, others believe that we are still far from a general solution to the problem, despite apparent advances (and commercial products) since the criticism of this field by John Pierce in the late 1960’s [24]. What could be the reason for this continuing skepticism ? The general problem could be seen as finding a mapping from the space of all allowable acoustic signals to that of all meaningful messages. Each space grows with the length of the message. Even when restric- ting the maximum length of the message, the mapping is many-to-one, since the speech signal carries information from many sources and many different speech utterances may carry the same linguistic message. The message may be produced by different talkers, a talker may speak slower or faster, the acoustic environment may change, the health or emotions of the talker may be reflected in changes of the talker’s voice, and so on. Thus, ASR must deal with significant non-linguistic variability in the data. We have limited understanding of how this is done by human listeners. ASR is thus a difficult for several reasons : (1) speech is originally an analog signal which, when typically sampled between 8 kHz (for tele- phone speech) and 10 to 16 kHz (for microphone speech), contains a lot
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/3 ppur_asr.pdf
| 2
|
3 ppur_asr
| 0
|
2 Introduction of redundant information that cannot be treated as such for recognition, (2) the acoustical signal is often degraded by additive and convolutional noise, e.g., due to environmental noise, telephone bandwidth, different microphone and filter characteristics, distorted acoustics, etc., (3) speech is a sequential signal that is affected by temporal and frequency distor- tions, (4) pronunciation of words and phonemes is strongly dependent on the context (which is usually referred to as coarticulation), (5) there are always some intra-speaker variabilities in the way words and pho- nemes are pronounced but also large inter-speaker variabilities including the effects of regional dialect, and (6) talkers will use words that are out of the recognizer vocabulary. In theory, a typical ASR system should generate all possible acoustic sequences for all legal linguistic messages. The model is trained using a large number of different acoustic signals that are associated with their corresponding word sequences. Recognition then could be done by tes- ting all possible word sequence models to find the one that would produce the acoustic signal best matching the actual signal being recognized. Fig. 1.1 Overall plan of a speech recognition system Given the complexity of the problem, and as illustrated in Figure 1.1, the resulting structure of current (state-of-the-art) ASR systems consists
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/3 ppur_asr.pdf
| 3
|
3 ppur_asr
| 0
|
Speech Recognition 3 of five major components : a) Signal processing : A feature extraction module that transforms the speech signal into a sequence of acoustic feature vectors. Degradations due to the factors (1) and (2) [as well as (3) to some extent] mentio- ned above can be reduced using a better feature extractor or “front end”. This issue has been addressed in CHAPTER..... Typically, this output of this module will represents a sentence as a sequence of acoustic vectors X = {x1, x2, . . . , xn, . . . , xN}. b) Probability estimator : A local hypothesis generator that will pro- duce a set of local hypothesis likelihoods, associated with subword (sound) or phonetic classes or some phonetic hypotheses about a speech segment (associated with one or more acoustic vectors xn). This is usually based on speech unit models (typically word or pho- neme models) which are trained on a large amount of speech data containing many occurrences of the speech units in many different contexts ; this will address problems (3) and (4) mentioned above. Performing this training on one specific speaker or on a large popula- tion of speakers will allow the models to take intra and inter-speaker variabilities [problem (5) mentioned above] into account. Typically, this module will thus generate consist a vector of likelihoods P(xn|qk), for every possible (phonetic) class qk (k = 1, . . . , K) and every instant (time frame) n. c) Grammar : A language module that should interact with the pattern matching module to help the recognizer to incorporate syntactic, se- mantic and pragmatic constraints. Sometimes, this can be used to address problem (6) mentioned above or to solve the ambiguity pro- blems. This has been addressed in CHAPTER... and will be briefly recalled in Section 1.6. d) Pronunciation Lexicon : Small vocabulary ASR systems may use mo- dels for whole words or even complete utterances. However, it is hard if not impossible to train models for all words in large vocabulary systems, since many examples of
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/3 ppur_asr.pdf
| 4
|
3 ppur_asr
| 0
|
ation Lexicon : Small vocabulary ASR systems may use mo- dels for whole words or even complete utterances. However, it is hard if not impossible to train models for all words in large vocabulary systems, since many examples of each are required in order to effec- tively learn the model parameters. Subsequently, there is usually a need for sub-word based models, which can be dynamically assigned to different words in the anticipated vocabulary. The most often-used sub-word units are phonemes or phoneme-like units. Each sequence of these units corresponds to“beads on a string”, or a sequence in which the feature emission statistics are derived from a single probability density function. As further discussed in Section 1.4, in order to better approximate the dynamics of speech
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/3 ppur_asr.pdf
| 4
|
3 ppur_asr
| 0
|
4 ASR statistical formalism production, HMMs for ASR are usually composed of multiple states for each phonemic sound unit (e.g., 2-5). Furthermore, the phoneme classes typically are further subdivided into different contextual classes (e.g., with particular neighbors to the left and right). This is done in order to better represent the wide variability in sounds that could be associated with any particular phoneme, for which a primary cause is the coarticulation between neighboring speech sounds due to the iner- tia in the vocal apparatus. Since using all allowable contexts would lead to a large number of the context dependent phoneme classes, the context phonemes are often clustered into larger classes and in this way the number of sub-word units is kept reasonably small. In large systems, acoustic probabilities from differing degrees of context and clustering are integrated together, much as was referred to earlier for the language models. e) Decoder : A time alignment and pattern matching module, that trans- forms the local hypothesis into a global decision for word or sentence recognition. This will be discussed in the companion paper (Bour- lard and Morgan, this volume) and is generally used to address the temporal distortion problem [problem (3) above]. This will be further discussed in Section 1.7. The difficulty of ASR varies greatly with the complexity of the task. As of this writing, the task may range from recognition of a few antici- pated voice commands uttered by a single user, to transcribing conver- sational speech of an arbitrary talker in realistic noisy environments. Speech signal processing and acoustic feature extraction was discus- sed in CHAPTER.... In this chapter, we mainly discuss the basics of automatic speech recognition. For more information, we refer the reader to the most recent (and probably most complete) books on the subject, including : [8], more oriented towards speech processing and recognition, and [12], slightly more oriented towards language processing. 1.2 ASR statistical formalism Much of the work in ASR focuses on the statistical modeling of speech, using either Hidden Markov Models (HMMs), Artificial Neu- ral Networks (ANNs), or a combination of both to enhance recognition
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/3 ppur_asr.pdf
| 5
|
3 ppur_asr
| 0
|
of the work in ASR focuses on the statistical modeling of speech, using either Hidden Markov Models (HMMs), Artificial Neu- ral Networks (ANNs), or a combination of both to enhance recognition performance. As discussed above, the goal of ASR is to find the most likely sequence of symbols representing the linguistic message in the speech signal given
|
/home/ricoiban/GEMMA/mnlp_chatsplaining/RAG/3 ppur_asr.pdf
| 5
|
3 ppur_asr
| 0
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.