quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - tomography for...
TRANSCRIPT
![Page 1: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/1.jpg)
Quantum algorithms forclassification
Alessandro Luongo
Twitter: @scinawa https://luongo.pro/qml
26-30-2018, Paris, International Conference on Quantum Computing
![Page 2: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/2.jpg)
![Page 3: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/3.jpg)
1 Prologue
2 Toolbox
3 Quantum algorithms for classification..Quantum Slow Feature AnalysisQuantum Frobenius Distance Classifierq-means
4 ..on real data
![Page 4: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/4.jpg)
Unsupervised methods
X ∈ Rn×d
Supervised methods
X ∈ Rn×d, Y ∈ Rn×m
• Anomaly detection
• Clustering
• Blind signal separation
• Text mining
• Regression
• Pattern recognition
• Time series forecasting
• Speech recognition
• Classification • Classification
![Page 5: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/5.jpg)
Unsupervised methods
X ∈ Rn×d
Supervised methods
X ∈ Rn×d, Y ∈ Rn×m
• Anomaly detection
• Clustering
• Blind signal separation
• Text mining
• Regression
• Pattern recognition
• Time series forecasting
• Speech recognition
• Classification • Classification
![Page 6: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/6.jpg)
![Page 7: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/7.jpg)
Unsupervised
![Page 8: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/8.jpg)
Unsupervised
![Page 9: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/9.jpg)
Unsupervised
![Page 10: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/10.jpg)
Supervised
![Page 11: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/11.jpg)
ML’s algos:runtime = O(poly(size)) = O(poly(n, d))...
... but size = O(2time)... ⇒ problem!
We need Quantum Machine Learning!runtime = O(polylog(size))
[HHL09] ...
![Page 12: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/12.jpg)
ML’s algos:runtime = O(poly(size)) = O(poly(n, d))...
... but size = O(2time)... ⇒ problem!
We need Quantum Machine Learning!runtime = O(polylog(size))
[HHL09] ...
![Page 13: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/13.jpg)
QML team @ IRIF
• Iordanis Kerenidis• Jonas Landman• Anupam Prakash
![Page 14: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/14.jpg)
Takeaways
• There is an efficient quantum procedure for superviseddimensionality reduction: Quantum Slow FeatureAnalysis
• There is an efficient quantum procedure for supervisedclassification and distance calculation: QuantumFrobenius Distance Estimator .
• There is a new efficient quantum procedure forunsupervised classification: q-means
• We simulate quantum algorithm on real data: they work!• QRAM based.
![Page 15: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/15.jpg)
1 - QRAMLet X ∈ Rn×d. There is a quantum algorithm that
|i⟩ |0⟩ → |i⟩ |xi⟩ |xi⟩ = ∥xi∥−1 |xi⟩
1√∑ni=0 ∥xi∥
2
n∑i=0
∥xi∥ |i⟩ |xi⟩
• Execution time: O(log nd)• Preparation time: O(nd log nd)• Size: O(nd log nd)
![Page 16: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/16.jpg)
1 - QRAMLet X ∈ Rn×d. There is a quantum algorithm that
|i⟩ |0⟩ → |i⟩ |xi⟩ |xi⟩ = ∥xi∥−1 |xi⟩
1√∑ni=0 ∥xi∥
2
n∑i=0
∥xi∥ |i⟩ |xi⟩
• Execution time: O(log nd)• Preparation time: O(nd log nd)• Size: O(nd log nd)
![Page 17: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/17.jpg)
1 - QRAMLet X ∈ Rn×d. There is a quantum algorithm that
|i⟩ |0⟩ → |i⟩ |xi⟩ |xi⟩ = ∥xi∥−1 |xi⟩
1√∑ni=0 ∥xi∥
2
n∑i=0
∥xi∥ |i⟩ |xi⟩
• Execution time: O(log nd)• Preparation time: O(nd log nd)• Size: O(nd log nd)
![Page 18: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/18.jpg)
![Page 19: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/19.jpg)
QRAM [[2,3,4],[5,6,7],[8,9,10]]
![Page 20: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/20.jpg)
QRAM + swaps... better
Thanks Alex Singh for the circuit
![Page 21: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/21.jpg)
2 - Q-BLAS- M ∈ Rd×d, s.t. ∥M∥2 = 1, in QRAM- x ∈ Rd in QRAM.There is a quantum algorithm that w.h.p. returns :
(i) |z⟩ such that∥∥|z⟩ − |M−1x⟩
∥∥ ≤ ϵ
in time O(κ(M)µ(M) log(1/ϵ))
(ii) |z⟩ such that ∥|z⟩ − |Mx⟩∥ ≤ ϵ
in time O(κ(M)µ(M) log(1/ϵ))
(iii) a state |M+≤θM≤θx⟩
in time O( µ(M)∥x∥δθ∥M+
≤θM≤θx∥)
Get estimates of ∥z∥ = f(M)x (with mult. error ϵ2, timeO() · ϵ−1
2 )
Gilyén, András, et al. ”Quantum singular value transformation and beyond: exponentialimprovements for quantum matrix arithmetics.” arXiv preprint arXiv:1806.01838 (2018).
![Page 22: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/22.jpg)
2 - Q-BLAS- M ∈ Rd×d, s.t. ∥M∥2 = 1, in QRAM- x ∈ Rd in QRAM.There is a quantum algorithm that w.h.p. returns :
(i) |z⟩ such that∥∥|z⟩ − |M−1x⟩
∥∥ ≤ ϵ
in time O(κ(M)µ(M) log(1/ϵ))
(ii) |z⟩ such that ∥|z⟩ − |Mx⟩∥ ≤ ϵ
in time O(κ(M)µ(M) log(1/ϵ))
(iii) a state |M+≤θM≤θx⟩
in time O( µ(M)∥x∥δθ∥M+
≤θM≤θx∥)
Get estimates of ∥z∥ = f(M)x (with mult. error ϵ2, timeO() · ϵ−1
2 )
Gilyén, András, et al. ”Quantum singular value transformation and beyond: exponentialimprovements for quantum matrix arithmetics.” arXiv preprint arXiv:1806.01838 (2018).
![Page 23: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/23.jpg)
2 - Q-BLAS- M ∈ Rd×d, s.t. ∥M∥2 = 1, in QRAM- x ∈ Rd in QRAM.There is a quantum algorithm that w.h.p. returns :
(i) |z⟩ such that∥∥|z⟩ − |M−1x⟩
∥∥ ≤ ϵ
in time O(κ(M)µ(M) log(1/ϵ))
(ii) |z⟩ such that ∥|z⟩ − |Mx⟩∥ ≤ ϵ
in time O(κ(M)µ(M) log(1/ϵ))
(iii) a state |M+≤θM≤θx⟩
in time O( µ(M)∥x∥δθ∥M+
≤θM≤θx∥)
Get estimates of ∥z∥ = f(M)x (with mult. error ϵ2, timeO() · ϵ−1
2 )
Gilyén, András, et al. ”Quantum singular value transformation and beyond: exponentialimprovements for quantum matrix arithmetics.” arXiv preprint arXiv:1806.01838 (2018).
![Page 24: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/24.jpg)
2 - Q-BLAS- M ∈ Rd×d, s.t. ∥M∥2 = 1, in QRAM- x ∈ Rd in QRAM.There is a quantum algorithm that w.h.p. returns :
(i) |z⟩ such that∥∥|z⟩ − |M−1x⟩
∥∥ ≤ ϵ
in time O(κ(M)µ(M) log(1/ϵ))
(ii) |z⟩ such that ∥|z⟩ − |Mx⟩∥ ≤ ϵ
in time O(κ(M)µ(M) log(1/ϵ))
(iii) a state |M+≤θM≤θx⟩
in time O( µ(M)∥x∥δθ∥M+
≤θM≤θx∥)
Get estimates of ∥z∥ = f(M)x (with mult. error ϵ2, timeO() · ϵ−1
2 )
Gilyén, András, et al. ”Quantum singular value transformation and beyond: exponentialimprovements for quantum matrix arithmetics.” arXiv preprint arXiv:1806.01838 (2018).
![Page 25: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/25.jpg)
2.5 - Q-BLAS- A, B ∈ Rd×d in QRAM∥A∥2 = ∥B∥2 = 1, in QRAM- x ∈ Rd in QRAM.There is a quantum algorithm that w.h.p. returns :
(i) |z⟩ such that∥∥|z⟩ − |(AB)−1x⟩
∥∥ ≤ ϵ
(ii) |z⟩ such that ∥|z⟩ − |(AB)x⟩∥ ≤ ϵ
(iii) a state |(AB)+≤θ,δ(AB)≤θ,δx⟩
Get estimates of ∥z∥ = f(AB)x (with mult. error ϵ2, timeO() · ϵ−1
2 )
Gilyén, András, et al. ”Quantum singular value transformation and beyond: exponentialimprovements for quantum matrix arithmetics.” arXiv preprint arXiv:1806.01838 (2018).
![Page 26: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/26.jpg)
• Before: Quantum Singular Value Estimation∑i
αi |vi⟩ 7→∑i
αi |vi⟩ |σi⟩
• Now: Qubitization:
W = eiϕ0σzeiθσxeiϕ1σzeiθσx · · · eiϕkσzeiθσx
Family of possible W is large enough...
![Page 27: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/27.jpg)
3 - Compute distances
V ∈ Rn×d, C ∈ Rk×d in the QRAM, and ϵ > 0There is a quantum algorithm that w.h.p. and in time O
(ηϵ
)|i⟩ |j⟩ |0⟩ 7→ |i⟩ |j⟩ |d(vi, cj)⟩
where |d(vi, cj)− d(vi, cj)| ⩽ ϵ , where η = max∥vi∥min∥vi∥
.
Based on: Wiebe, N., Kapoor, A., & Svore, K. (2014). Quantum algorithmsfor nearest-neighbor methods for supervised and unsupervised learning.arXiv preprint arXiv:1401.2142.
![Page 28: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/28.jpg)
3 - sketch proof• Use Quantum Frobenius Dinstance to build:
∥vi∥√Zij
|i⟩ |j⟩ |0⟩ |vi⟩+∥∥cj∥∥√Zij
|i⟩ |j⟩ |1⟩ |cj⟩
• Hadamard on 3rd qubit.
p(1)ij =1
2Zij(∥vi∥2+
∥∥cj∥∥2−2 ∥vi∥∥∥cj∥∥ ⟨vi, cj⟩) = d(vi, cj)2
2Zij
• Perform amplitude estimation on L copies.
• Use Median Lemma (Wiebe et. al.)
• Invert circuit (garbage collection), multiply by 2Zij.
![Page 29: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/29.jpg)
4 - Tomography
For a pure quantum state |x⟩, there is atomography algorithm with sample and timecomplexity O(d log d/ϵ2) that produces an estimatex ∈ Rd with ∥x∥2 = 1 such that ∥x− x∥2 ≤ ϵ withprobability at least (1 − 1/d0.83).
Kerenidis, Iordanis, and Anupam Prakash. ”A quantuminterior point method for LPs and SDPs.” arXiv preprintarXiv:1808.09266 (2018).
![Page 30: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/30.jpg)
![Page 31: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/31.jpg)
PCA
![Page 32: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/32.jpg)
SFA || FLD
![Page 33: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/33.jpg)
Slow Feature Analysis (Supervised)
Input signal: x(i) ∈ Rd. Task: Learn K functions:
y(i) = [g1(x(i)), · · · , gK(x(i))]
Such that ∀j ∈ [K]. Minimize:
∆(yj) =1a
K∑k=1
∑s,t∈Tks<t
(gj(x(s))− gj(x(t))
)2
Constraints on output signal: average of components is 0,variance of components is 1, signals are decorrelated.
Def Cov. matrix B := XTX, Derivative cov. matrix A := XTX
AW = BWΛ
![Page 34: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/34.jpg)
Slow Feature Analysis (Supervised)
Input signal: x(i) ∈ Rd. Task: Learn K functions:
y(i) = [g1(x(i)), · · · , gK(x(i))]
Such that ∀j ∈ [K]. Minimize:
∆(yj) =1a
K∑k=1
∑s,t∈Tks<t
(gj(x(s))− gj(x(t))
)2
Constraints on output signal: average of components is 0,variance of components is 1, signals are decorrelated.
Def Cov. matrix B := XTX, Derivative cov. matrix A := XTX
AW = BWΛ
![Page 35: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/35.jpg)
Step 1: Whitening
Data is whitened (sphered) if B = XTX = I.
Whitening its just matrix (Moore-Penrose) inversionZ := X+X .. now ZTZ = IFreebie Theorem!There exists an efficient quantum algorithm forwhitening that builds |Z⟩
![Page 36: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/36.jpg)
Step 2: Projection
• Whiten data |X⟩ 7→ |Z⟩• Project data in slow feature space |Z⟩ 7→ |Y⟩
![Page 37: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/37.jpg)
New algo! QSFA
- Let X =∑
i σiuivTi ∈ Rn×d, X ∈ Rn log n×d QRAM.
- Let ϵ, θ, δ, η > 0.There exists a quantum algorithm that produces:
• |Y⟩ with | |Y⟩ − |A+≤θ,δA≤θ,δZ⟩ | ≤ ϵ in time
O((
κ(X)µ(X) log(1/ε) +(µ(X) + µ(X))
δθ
)× ||Z||
||A+≤θ,δA≤θ,δZ||
)
• ∥Y∥ s.t. |∥Y∥ − ∥Y∥ | ≤ η ∥Y∥ with anadditional 1/η factor.
![Page 38: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/38.jpg)
New algo! QFDC (Supervised)
Xk ∈ R|Tk|×d matrix of elements labeled kX0 ∈ R|Tk|×d repeats the row x0 for |Tk| times.
Fk(x0) =∥Xk − X0∥2
F
2(∥Xk∥2F + ∥X0∥2
F),
1√Nk
(|0⟩
∑i∈Tk
∥x(0)∥ |i⟩ |x(0)⟩+|1⟩∑i∈Tk
∥x(i)∥ |i⟩ |x(i)⟩)
h(x0) = mink{Fk(y0) = p(|1⟩)}
![Page 39: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/39.jpg)
Accuracy QSFA+QFDC
![Page 40: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/40.jpg)
From Wikipedia
![Page 41: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/41.jpg)
From Wikipedia
![Page 42: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/42.jpg)
From Wikipedia
![Page 43: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/43.jpg)
From Wikipedia
![Page 44: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/44.jpg)
From Wikipedia
![Page 45: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/45.jpg)
k-means (Unsupervised)
Find initial centroids cjRepeat until centroids are steady: |ctj − ct+1
j | ≤ τ
• Calculate distances between all points and all clusters
∀i ∈ [n], c ∈ [k] d(vi, ci)
• Assign points to closer cluster
l(vi) = arg minc∈[k]
d(vi, ci)
• Calculate new centroids
cj =1|Cj|
∑i∈Cj
vi
... is O(tndk) :(
![Page 46: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/46.jpg)
k-means (Unsupervised)
Find initial centroids cjRepeat until centroids are steady: |ctj − ct+1
j | ≤ τ
• Calculate distances between all points and all clusters
∀i ∈ [n], c ∈ [k] d(vi, ci)
• Assign points to closer cluster
l(vi) = arg minc∈[k]
d(vi, ci)
• Calculate new centroids
cj =1|Cj|
∑i∈Cj
vi
... is O(tndk) :(
![Page 47: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/47.jpg)
δ-k-means (Unsupervised)
Find initial centroids cjRepeat until centroids are steady: |ctj − ct+1
j | ≤ τ
• Calculate distances between all points and all clusters
∀i ∈ [n], c ∈ [k] d(vi, ci)
• Assign points to closer cluster
Lδ(vi) = {cp |d2(c∗i , vi)− d2(cp, vi)| ≤ δ }
l(vi) = rand(Lδ(vi))
• Calculate new centroids
cj =1|Cj|
∑i∈Cj
vi
![Page 48: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/48.jpg)
q-means (Unsupervised)
Find initial centroids cjRepeat until centroids are steady: |ctj − ct+1
j | ≤ τ
• Calculate distances between all points and all clustersK⊗j=0
n∑i=0
|i⟩ |j⟩ |d(vi, ci)⟩
• Assign points to closer clustern∑i=0
|i⟩ |l(i)⟩
• Calculate centroids againk∑j=1
√|Cj|N
|ct+1j ⟩ |j⟩
![Page 49: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/49.jpg)
Well-clusterable data
The data is (ξ, β, λ, η)-well clustered if there areξ > 0, β > 0, 0 ≪ λ < 1, η > 1 :
1 clusters’ separation: d(ci, cj) ≥ ξ ∀i, j ∈ [k]2 proximity to centroid: A fraction λn of pointsvi in the dataset verify: d(vi, cl(vi)) ≤ β.
3 dataset’s width: All the norms are between 1and η = maxi (∥vi∥)
![Page 50: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/50.jpg)
![Page 51: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/51.jpg)
New algo! q-means
For a (ξ, β, λ, η)-well clusterable dataset V ∈ Rn×d
in QRAM, there is a quantum algorithm that returnsin t steps the k centroids that cluster the datasetconsistently with the classical δ-k-means algorithmin time O
(t · k
3dη3
δ3
).
![Page 52: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/52.jpg)
Accuracy q-means
![Page 53: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/53.jpg)
λmax/λmin: more data
Condition number by increasing the number of elements intraining set
![Page 54: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/54.jpg)
λmax/λmin: more features
Condition number by increasing the features (pixels)
![Page 55: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/55.jpg)
µ(X): more data
µ(X) and µ(X) by increasing the number of elements intraining set.
![Page 56: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/56.jpg)
µ(X): more features
µ(X) and µ(X) by increasing the number of features.
![Page 57: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/57.jpg)
Conclusions
Quantum Slow Feature Analysis and Quantum FrobeniusDistance Classifier on MNIST
• ∝ 150 || 200 (Logical) qubits
• Classifying the test set (104 vectors) with quantum algosis ≃ 100 times faster
• Possibility to extract the model classically!|i⟩ |i⟩ 7→ |i⟩ |gi⟩.
• Not only fast but might be more accurate!
![Page 58: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/58.jpg)
Conclusions
q-means• Exponentially faster in the number of data
points: O(n) → O(log n).• Finding new datasets to apply q-means.• We recover the centroids classically
![Page 59: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/59.jpg)
... Yet an uneven comparison?
From: https://hdbscan.readthedocs.io/en/latest/performance_and_scalability.html
![Page 60: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/60.jpg)
#TODOs
• Generalizations...• Experiments...• Code...• New algos...• Compositions...• Adversarial QML...• Privacy preserving QML...
![Page 61: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/61.jpg)
Thanks for your timethere is never enough.
(cit. Dan Geer)
@scinawa
• Quantum Machine Learning ⇒ https://luongo.pro/qml
• QSFA + QFDC ⇒ https://arxiv.org/abs/1805.08837
• q-means ⇒ stay tuned...
![Page 62: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/62.jpg)
• Projective Simulation
• Quantum Recommendation Systems
• Quantum SVM
• Quantum Anomaly detection
• Quantum Gradient Descent
•
• Quantum PCA
• ...
![Page 63: Quantum algorithms for classificationicoqc.sciencesconf.org/data/pages/luongo.pdf4 - Tomography For a pure quantum state jx , there is a tomography algorithm with sample and time complexity](https://reader034.vdocuments.site/reader034/viewer/2022043010/5f9f12b5148b3d7de956d6c9/html5/thumbnails/63.jpg)
1 Ewin Tang. A quantum-inspired classicalalgorithm for recommendation systems.arXiv:1807.04271, 2018. (undergraduate thesis,advised by Scott Aaronson)
2 Ewin Tang. Quantum-inspired classicalalgorithms for principal componentanalysis and supervised clustering.arXiv:1811.00414, 2018.
3 Ewin Tang et al. Quantum-inspired low-rankstochastic regression with logarithmicdependence on the dimension arXiv:1811.04909, 2018