Semantic Communication for SWIPT systems Nizar Khalfet, Costas Psomas, Symeon Chatzinotas, and Ioannis Krikidis IRIDA Research center for Communication Technologies Department of Electrical and Computer Engineering University of Cyprus IRIDA Workshop 1 / 24 1 Introduction/Motivation 2 Contributions 3 DM channel with SWIPT and semantics Achievable region for the DM channel Converse Semantic information-energy region for DM Channel 4 Gaussian channel with SWIPT and semantics 5 Numerical results 6 Conclusion/Extension 2 / 24 Introduction/Motivation It is predicted that the global amount of data generated by network nodes will increase to 175 zetta-bytes in 2025. SemCom: transmit only the key information to the destination: safely remove the information irrelevant to the specific task without causing any performance degradation. Existing research contributions reveals that SemCom is promising to be used when the SNR is low and/or the available wireless resource is limited. SemCom generally requires less power/bandwidth than BitCom. Introducing SemTrans: enables users to achieve a comparable performance as BitTrans but consmuing low transmit power, and causing less interference. 2 / 24 Contributions A novel framework is proposed that enables the study of the fundamental limits of SWIPT with semantic communication for a discrete memoryless (DM) channel and a Gaussian channel. An achievable information-energy region as well its converse for both the DM channel and the Gaussian channel are characterized by including the semantic context into the communication. A higher performance is observed in comparison to conventional communication approaches (i.e., without semantics) by considering a low semantic ambiguity code. 3 / 24 DM channel with SWIPT and semantics Semantic channel: P(Q|W ) PY |X Transmitter Y IR X PS|X ER S The output y is observed at the receiver, with probability P(Y = y | X = x) = n Y P(yi |xi ), (1) i=1 P(yi |xi ) is the transition probability distribution. Q: semantic context random variable, with respect to a probability distribution P(Q|W ) X P (Q = q|W = w ) = 1. (2) q∈Q 4 / 24 DM channel with SWIPT and semantics Semantic channel: P(Q|W ) PY |X Transmitter Y IR X PS|X ER S semantic distance between two words as d(w , ŵ ) = 1 − sim(w , ŵ ), (3) 0 ≤ sim(w , ŵ ) ≤ 1: semantic similarity between w and ŵ the average semantic error denoted by PSE is given by X P(Y = y |X = x) PSE (q) = w ∈W,q∈Q,y ∈Y n ,x∈X n ×P(Q = q|W = w )P(W = w )d(w , ŵ ) (4) 5 / 24 DM channel with SWIPT and semantics Semantic channel: P(Q|W ) PY |X Transmitter Y IR X PS|X ER S The output s at the EH is observed at the receiver, with probability P(S = s | X = x) = n Y P(si |xi ), (5) i=1 E(s): the average harvested energy function given by E(s) = n 1X g (si ), n i=1 (6) g : S → R+ : the energy harvested from the output symbols. 6 / 24 DM channel with SWIPT and semantics Semantic channel: P(Q|W ) PY |X Transmitter Y IR X PS|X ER S Achievable rates From a semantic communication standpoint, an information-energy rate (R, b) is achievable, if the probability of the miss-interpretation of message w , given the context Q, PSE , satisfies the limit PSE (q) → 0, for n → ∞ and the energy shortage probability, PES (b) satisfies PES [E(y ) ≤ b] → 0, for n → ∞. 7 / 24 Achievable region for the DM channel Semantic Information energy achievable region The information energy capacity in the case of non-colocated EH is upper bounded by the function C : [b1 , b0 ] → R+ , is C(b) ≥ max ρ.P(X |W )|E(y )≥b I (X ; Y ) − H(X |W ) + H(Q), I (X ; Y ) = H(X ) − H(X |Y ) is the mutual information between X and Y . H(X |W ) is the equivocation of the semantic encoder, given the semantic context Q. H(Q) measures the semantic context source, or the local information available at the transmitter and receiver. A higher H(Q) means strong ability of the receiver to interpret received messages. 8 / 24 Proof of Achievable region: Jointly typical sequences Definition Let An jointly typical sequences {(x n , y n )} with respect to p(x, y ) is the set of n sequences with empirical entropy close to the true entropies, i.e., An = {(x n , y n )} ∈ X n × Y n : 1 − log p(x n ) − H(X ) < n 1 − log p(y n ) − H(Y ) < n 1 − log p(x n , y n ) − H(X , Y ) < n p(x n , y n ) = Qn i=1 (7) (8) (9) (10) p(xi , yi ) 9 / 24 Proof of Achievable region: Jointly typical sequences Theorem P{(X n , Y n ) ∈ An } → 1 as n → ∞ |An | ≤ 2n(H(X ,Y )+) if (X n , Y n ) has a distribution p(x n )p(y n ) then P{(X n , Y n ) ∈ An } ≤ 2n(I (X ;Y )−3) (11) 10 / 24 Proof of the Achievable region A semantic error appears if a received message is not decoded by the receiver using the context Q. Let n be sufficiently large number, Q1 , Q2 , . . . , QN is the sequence of the observed context, X1 , X2 , . . . , XN is the sequence of the transmitted signals, Y1 , X2 , . . . , XN is the sequence of the received signals. According to AEP: there are 2nH(Q) typical sequence of context According to AEP: and the channel coding Theorem : there are 2(I (X ;Y )−R)N typical input sequence. According to AEP: there are 2−NH(X |W ) typical sequence of X given the context. Hence there are : 2(I (X ;Y )−H(X |W )+H(Q))N typical sequence of input given the context. 11 / 24 Converse region for the DM channel Semantic Information energy converse region The information energy capacity in the case of non-colocated EH is upper bounded by the function C : [b1 , b0 ] → R+ , is X C(b) ≤ max H(X ) − H(X |W ) − H(X |W , Y ) + p(W )H(Q|W ) ρ.P(X |W )|E (ρ)≥b W 12 / 24 Sketch of the Proof By using Fano inequality: NR ≤ N X NI (Xi ; Yi ) + o(N) (12) i=1 X , Y and W are random variables such that W → X → Y that form a Markov chain, i.e., W and Y are independent given X I (Y ; W |X ) = 0. (13) I (X ; W ) ≥ I (Y ; W ) (14) Data procession inequality, i.e, 13 / 24 SWIPT Gaussian channel with semantics Semantic channel: P(Q|W ) zt Transmitter xt ⊗ h1 ⊕ y1,t IR h2 y2,t ER Information Decoder: y1,t = h1 xt + zt , Energy harvester: y2,t = h2 xt zt ∼ N (0, 1) is the Gaussian noise with unit variance. The conditional probability: (y − h1 x 2 ) 1 p(y |x) = √ exp − . 2 2π (15) 14 / 24 SWIPT Gaussian channel with semantics Semantic loss for the Guassian case by using the maximum likelihood (ML) detector: s n n kxi − xj k2 1X X , (16) PSE (q) = d(xi , xj )Q P n i=1 2 j=1,j6=i d(xi , xj ) = 1 − sim(xi , xj ) sim(xi , xj ) = BΦ (mi )BΦ (mj )T , kBΦ (mi )k kBΦ (mj )k (17) BΦ (·) : the pretrained bidirectional encoder representation from transformers (BERT) model The average harvested energy: E(Y2 ) = n 1X 2 Y2,t . n t=1 (18) 15 / 24 Gaussian channel with SWIPT and semantics Semantic channel: P(Q|W ) zt Transmitter xt ⊗ h1 ⊕ y1,t IR h2 Y2,t ER Achievable region for Gaussian channel The information energy capacity is lower bounded by the function C : R+ → R+ , is C(b) ≥ max P(X |W )|E(Y2 )≥b 1 log(1 + λ1 P) − h(X |W ) 2 1 + log(1 + λ2 P), 2 (19) where λ1 ≥ 0 and λ2 ≥ 0 denotes the fraction of the power dedicated to the information signal component and the semantic context signals, respectively. 16 / 24 Proof of the achievable region for the Gaussian channel Context random variable satisfies, Q ∼ N (0, λ2 P) and X ∼ N (0, λ1 P) λ1 + λ2 = 1, (20) Codewords are chosen to be i.i.d, with variance P − and satisfies E[I0 (Bh2 |Xi |)] = b + . The context Q is sent through a feedback link using a faction λ2 of the power P. The receiver observes the codeword list and the sequence of the context Q1 , Q2 , . . . , Qn generated via feedback by the transmitter, and decides on w if {X n (w ), Y } are jointly typical, {Q n (w )} is a typical sequence, and {X n (w )|W } is a typical sequence given the context. using the AEP, the semantic error probability is upper bounded as follows PSE ≤ 2N(−I (X :Y )+H(X |Q)−H(Q)−3) , (21) 17 / 24 Gaussian channel with SWIPT and semantics Semantic channel: P(Q|W ) zt Transmitter xt ⊗ h1 ⊕ y1,t IR h2 Y2,t ER Converse region for Gaussian channel The information-energy capacity region for the SWIPT semantic communication is upper bounded by the function C : R+ → R+ , i.e., C(b) ≤ max 1 E(Y2 )≥b 2 log(1 + µ1 Preq ) + 1 log(1 + µ2 Preq ), 2 where µ1 ≥ 0 and µ2 ≥ 0 denote the fraction of the power dedicated to the input X and the semantic context Q, respectively. 18 / 24 Proof of the converse region for the Gaussian channel From the assumption that the message indice W ∈ {1, 2, . . . , 2nR } are i.i.d. following a uniform distribution,and using Fano’s inequality nR = H(W ) = I (W ; Ŵ ) + H(W |Ŵ ) n X ≤ h(Yi |Qi ) − h(Z ) + n, (22) (23) i=1 → 0 as PSE → 0. By using a power splitting technique between the information component and the semantic context Q, i.e., Y = µ1 X + µ2 Q + Z , (24) µ1 + µ2 = 1. By applying Jensen’s inequality, we obtain R≤ 1 1 log(1 + µ1 P) + log(1 + µ2 P). 2 2 (25) 19 / 24 Example: Semantic through binary symmetric channel Binary symmetric channel (BSC) with a crossover probability of ρ p(Y = y | X = x) = ρl(y,x) (1 − ρ)n−l(y,x) (26) l(y, x) is the Hamming distance between y and x. Two sets of contexts Q = {q1 , q2 } q1 = things originating from non-living beings q2 = things originating from living beings. 1, P(Q = q1 |W = w ) = 0, 0.5. E (ρ) = if w = car,automobile if w = bird if w = Crane (27) n 1X [(1 − ρ)P(xn = 1) + ρP(xn = 0)]b1 n n=1 +[ρP(xn = 1) + (1 − ρ)P(xn = 0)]b0 , (28) 20 / 24 Numerical results: Achievable semantic information-energy region versus converse region over a DM channel 21 / 24 Achievable semantic information-energy region versus conventional capacity over a DM channel 22 / 24 Impact of semantic context on the information-energy capacity region over the Gaussian channel 23 / 24 Conclusion/Extension For the point-to-point case, we have developed an information theoretic framework to characterize achievable information-energy regions and their converses, considering both the DM and the Gaussian channels. Numerical results indicate that employing low semantic ambiguity codes can lead to improved performance compared to conventional communication approaches. Extension: The Gaussian MAC case has been considered: A system with a semantic transmitter and a conventional transmitter, operating under EH constraints at the receiver has been considered. 24 / 24