+ Compressed Sensing + Compressed Sensing Mobashir Mohammad Aditya Kulkarni Tobias Bertelsen Malay Singh Hirak Sarkar Nirandika Wanigasekara Yamilet Serrano Llerena Parvathy Sudhir 3 + Introduction Mobashir Mohammad + 4 The Data Deluge Sensors: Better… Stronger… Faster… Challenge: Exponentially increasing amounts of data Audio, Image, Video, Weather, … Global scale acquisition + 5 + 6 Sensing by Sampling Sample N + 7 Sensing by Sampling (2) N >> L Sample N Compress L JPEG … N >> L L Decompress N + 8 Compression: Toy Example + 9 Discrete Cosine Transformation Transformation + 10 Motivation Why go to so much effort to acquire all the data when most of the what we get will be thrown away? Cant we just directly measure the part that wont end up being thrown away? Donoho 2004 11 Outline • • • • • • • + Compressed Sensing Constructing Φ Sparse Signal Recovery Convex Optimization Algorithm Applications Summary Future Work 12 + Compressed Sensing Aditya Kulkarni + 13 What is compressed sensing? A paradigm shift that allows for the saving of time and space during the process of signal acquisition, while still allowing near perfect signal recovery when the signal is needed Analog Audio Signal High-rate Nyquist rate Sampling Compressed Compression (e.g. MP3) Sensing Low-rate + 14 Sparsity The concept that most signals in our natural world are sparse a. Original image c. Image reconstructed by discarding the zero coefficients + 15 How It Works + Dimensionality Reduction Problem 𝒚=𝚽𝒙 I. Measure 𝑦 II. Construct sensing matrix Φ III. Reconstruct 𝑥 16 + 17 Sampling Φ=I 𝑦 𝑥 𝑁× 1 𝑁×1 measurements sparse signal = 𝐾 nonzero entries 𝑁×𝑁 + 18 Φ 𝑦 𝑥 𝑁× 1 𝑀×1 measurements 𝐾<𝑀≪𝑁 sparse signal = 𝐾 nonzero entries 𝑀×𝑁 + 19 Ψ 𝑥 𝑁×1 𝑁 nonzero entries 𝛼 𝑁×1 = 𝐾 nonzero entries 𝑁×𝑁 + 20 Sparsity The concept that most signals in our natural world are sparse a. Original image c. Image reconstructed by discarding the zero coefficients + 21 Φ 𝑦 Ψ 𝛼 = 𝑀×1 𝑀×𝑁 𝑁×𝑁 𝑁×1 22 + Constructing Φ Tobias Bertelsen + 23 RIP - Restricted Isometry Property A matrix 𝛷 satisfies the RIP of order K if there exists a 𝛿 ∈ (0,1) such that: Φ𝑥1 − Φ𝑥2 22 1−𝛿 ≤ ≤1+𝛿 2 𝑥1 − 𝑥2 2 holds for all 𝐾-sparse vectors 𝑥1 and 𝑥2 Or equally Φ𝑥 22 1−𝛿 ≤ ≤1+𝛿 𝑥 22 holds for all 2K-sparse vectors 𝑥 The distance between two points are approximately the same in the signal-space and measure-space + 24 RIP - Restricted Isometry Property RIP ensures that measurement error does not blow up 1−𝛿 ≤ Φ𝑥1 −Φ𝑥2 22 𝑥1 −𝑥2 22 ≤1+𝛿 Image: http://www.brainshark.com/brainshark/brainshark.net/portal/title.aspx?pid=zCgzXgcEKz0z0 + 25 Randomized algorithm 1. Pick a sufficiently high 𝑀 2. Fill Φ randomly according to some random distribution Which How distribution? to pick 𝑀? What is the probability of satisfying RIP? + 26 Sub-Gaussian distribution Defined by 𝐸 𝑒 𝑋𝑡 ≤ 𝑒 𝑐2 𝑡2 2 Tails decay at least as fast as the Gaussian E.g.: The Gaussian distribution, any bounded distribution Satisfies the concentration of measure property (not RIP): For any vector 𝑥 and a matrix Φ with sub-Gaussian entries, there exists a 𝛿 ∈ [0,1] such that Φ𝑥 22 1−𝛿 ≤ ≤ 1+𝛿 𝑥 22 holds with exponentially high probability 1 − 2−𝑑 𝛿 𝑀 where 𝑑(𝛿) is a constant only dependent on 𝛿 + 27 Johnson-Lidenstrauss Lemma Generalization to a discrete set of 𝑃 vectors For any vector the magnitude are preserved with: 𝑝𝑟𝑜𝑏 = 1 − 2−𝑑 𝛿 𝑀 For all P vectors the magnitudes are preserved with: 𝑝𝑟𝑜𝑏 = 1 − 𝑃2−𝑑 𝛿 𝑀 = 1 − 2log 𝑃−𝑑 𝛿 𝑀 To account for this 𝑀 must grow with 𝑂 log 𝑃 + 28 Generalizing to RIP Φ𝑥 22 𝑥 22 RIP: 1 − 𝛿 ≤ We want to approximate all 2𝐾-sparse vectors with 𝑃 unit vectors ≤ 1+𝛿 , 𝑥 0 ≤ 2𝐾 𝑁 The space of all 2𝐾-sparse vectors is made up of 2𝐾 2𝐾-dimensional subspaces – one for each position of non-zero entries in 𝑥 We sample points on the unit-sphere of each subspace 𝑁 ≤ 2𝐾 𝑁𝑒 2𝐾 2𝐾 𝑁 𝐾 𝐾 𝑃=𝑂 𝑀 = 𝑂 log 𝑃 = 𝑂 𝐾 log 𝑁 𝐾 + 29 Randomized algorithm Use sub-Gaussian distribution Pick 𝑀 = 𝑂 𝐾 log Exponentially Formal 𝑁 𝐾 high probability of RIP proofs and specific formulas for constants exists + 30 Sparse in another base We assumed the signal itself was sparse What if the signal is sparse in another base, i.e. 𝑥 = Ψ𝛼 is sparse. ΦΨ must have the RIP As long as Ψ is an orthogonal basis, the random construction works. + 31 Characteristics of Random Φ Stable Universal Works with any orthogonal basis Democratic Robust to noise, since it satisfies RIP Any element in has equal importance Robust to data loss Other Methods Random Fourier submatrix Fast JL transform 32 + Sparse Signal Recovery Malay Singh + 33 The Hyperplane of 𝑦 = Φ𝑥 + 34 𝐿𝑝 Norms for N dimensional vector x 1 𝑝 𝑁 𝑥𝑗 𝑥 𝑝 = 𝑝 𝑗=1 𝑠𝑢𝑝𝑝(𝑥) max 𝑥𝑗 𝑗=1,2,…,𝑁 Unit Sphere of 𝐿1 norm if 𝑝 > 0 Unit Sphere of 𝐿2 norm if 𝑝 = 0 if 𝑝 = ∞ Unit Sphere of 𝐿∞ norm Unit Sphere of 𝐿0.5 quasinorm + 35 𝐿𝑝 Balls in higher dimensions + 36 How about 𝐿0 minimization 𝑥 = arg min 𝑥 𝑥∈𝐵(𝑦) 0 𝐵 𝑦 = 𝑥 ∶ Φ𝑥 = 𝑦 But the problem is non-convex and very hard to solve + 37 We do the 𝐿2 minimization 𝑥 = arg min 𝑥 𝑥∈𝐵(𝑦) 2 𝐵 𝑦 = 𝑥 ∶ Φ𝑥 = 𝑦 We are minimizing the Euclidean distance. But the arbitrary angle of hyperplane matters + 38 What if we convexify the 𝐿0 to 𝐿1 𝑥 = arg min 𝑥 𝑥∈𝐵(𝑦) 1 𝐵 𝑦 = 𝑥 ∶ Φ𝑥 = 𝑦 + 39 Issues with 𝐿0 minimization ⋅ 0 is non-convex and 𝐿0 minimization is potentially very difficult to solve. We convexify the problem by replacing ⋅ ⋅ 1 . This leads us to 𝐿1 Minimization. 0 by Minimizing 𝐿2 results in small values in some dimensions but not necessarily zero. 𝐿1 provides a better result because in its solution most of the dimensions are zero. 40 + Convex Optimization Hirak Sarkar + 41 What it is all about … Find a sparse representation 𝑥 𝑦 = Φ𝑥 Here Φ ∈ ℝ𝑀×𝑁 and 𝑥 ∈ ℝ𝑁 , Moreover 𝑀 ≪ 𝑁 Two ways to solve (P1) min 𝐽 𝑥 : 𝑦 = Φ𝑥 𝑥 where 𝐽 𝑥 is a measure of sparseness (P2) min{𝐽 𝑥 : 𝐻 Φ𝑥, 𝑦 ≤ 𝜖} 𝑥 + 42 How to chose 𝐻 𝑥 and 𝐽 𝑥 Take the simplest convex function 𝐽 𝑥 = 𝑥 1 𝐻 Φ𝑥, 𝑦 = 1 2 A simple 𝐻 𝑥 Φ𝑥 − 𝑦 Final unconstrained version min{𝜇 𝑥 𝑥 1 + 𝐻(𝑥)} 2 2 + 43 Versions of the same problem 1 min{𝜇 𝑥 1 + Φ𝑥 − 𝑦 𝑥 2 2 2} 1 min 𝑥 1 𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜 Φ𝑥 − 𝑦 𝑥 2 2 2 ≤𝜖 1 min Φ𝑥 − 𝑦 𝑥 2 1 ≤𝑡 2 2 𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜 𝑥 + 44 Formalize Nature Convex Differentiable 𝛻𝐻 𝑥 ′ = Φ𝑇 (Φ𝑥 ′ − 𝑦) Basic of 𝐻 Intuition Take an arbitrary 𝑥 ′ Calculate 𝑥 ′ − 𝜏𝛻𝐻(𝑥′) Use the shrinkage operator Make corrections and iterate + 45 Shrinkage operator We define the shrinkage operator as follows 𝑥−𝛼 𝑠ℎ𝑟𝑖𝑛𝑘 𝑥, 𝛼 = 0 𝑥+𝛼 if 𝛼 < 𝑥 if − 𝛼 ≤ 𝑥 ≤ 𝛼 if 𝑥 < −𝛼 + 46 Algorithm Input: Matrix Φ, Signal measurement 𝑦, parameter sequence 𝜇𝑛 Output: Signal estimate 𝑥 Initialization: 𝑥0 = 0, 𝑟 = 𝑦, 𝑘 = 0 + 47 Performance For closed and convex function 𝐽 𝑥 any 𝜇 > 0 the algorithm converges within finite steps For 𝐽 𝑥 = 𝑥 1 and a moderate 𝜇 number of iterations needed is less than 5 48 + Single Pixel Camera Nirandika Wanigasekara + 49 Single Pixel Camera What is a single pixel camera An optical computer sequentially measures the 𝑦[𝑖] Directly acquires 𝑀 random linear measurements without first collecting the 𝑁 pixel values + 50 Single Pixel Camera- Architecture + 51 Single Pixel Camera- DMD Array Digital Micro mirror Device A type of a reflective spatial light modulator Selectively redirects parts of the light beam Consisting of an array of N tiny mirrors Each mirror can be positioned in one of two states(+/-10 degrees) Orients the light towards or away from the second lens + 52 Single Pixel Camera- Architecture + 53 Single Pixel Camera- Photodiode Find the focal point of the second lens Place a photodiode at this point Measure the output voltage of the photodiode The voltage equals 𝑦𝑗 , which is the inner product between 𝜙𝑗 and the desired image 𝑥. + 54 Single Pixel Camera- Architecture + 55 Single Pixel Camera- measurements A random number generator (RNG) sets the mirror orientations in a pseudorandom 1/0 pattern Repeats the above process for 𝑀 times Obtains the measurement vector 𝑦 and ϕ Now we can construct the system in the = 𝑦𝑗 Φj 𝑥 + 56 Single Pixel Camera- Architecture + 57 Sample image reconstructions 256*256 conventional image of black and white ‘R’ Image reconstructed from 𝑀 = 1300 How can we improve the reconstruction further? M 256 ⇥ 256 256 ⇥ 256 = 1300 256 ⇥ 256 M = 1300 + 58 Utility This device is useful when measurements are expensive Low Light Imager Conventional Photomultiplier tube/ avalanche photodiode Single Pixel Camera Single photomultiplier Original 800 1600 65536 pixels from 6600 + 59 Utility CS Infrared Imager IR photodiode CS Hyperspectral Imager 60 + Compressed Sensing MRI Yamilet Serrano Llerena + 61 Compressed Sensing MRI Magnetic Resonance Imaging (MRI) Essential medical imaging tool with slow data acquisition process. Applying Compressed Sensing (CS) to MRI offers that: • We can send much less information reducing the scanned time • We are still able to reconstruct the image in based on they are compressible + 62 Compressed Sensing MRI Scan Process + 63 Scan Process Signal Received K-space Space where MRI data is stored K-space trajectories: K-space is 2D Fourier transform of the MR image + 64 In the context of CS Φ: • • • y=Φx Is depends on the acquisition device Is the Fourier Basis Is an M x N matrix y: • Is the measured k-space data from the scanner x: + 65 Recall ... The heart of CS is the assumption that x has a sparse representation. Medical Images are naturally compressible by sparse coding in an appropriate transform domain (e.g. Wavelet Transform) Not significant + 66 Compressed Sensing MRI Scan Process + 67 Image Reconstruction CS uses only a fraction of the MRI data to reconstruct image. + 68 Image Reconstruction + 69 Benefits of CS w.r.t Resonance Allow for faster image cardiac/pediatric imaging) acquisition (essential for Using same amount of k-space data, CS can obtain higher Resolution Images. 70 + Summary Parvathy Sudhir Pillai + 71 Summary Motivation Data deluge Directly acquiring useful part of the signal Key idea: Reduce the number of samples Implications Dimensionality reduction Low redundancy and wastage + 72 Open Problems ‘Good’ sensing matrices Adaptive? Deterministic? Nonlinear compressed sensing Numerical algorithms Hardware design Coefficients (𝛼) Intensity (x) Phase (𝜃) + 73 Impact Data generation and storage Conceptual achievements Exploit minimal complexity efficiently Information theory framework Numerous application areas Legacy - Trans disciplinary research Information Hardware CS Software Complexity + 74 Ongoing Research New mathematical framework for evaluating CS schemes Spectrum sensing Not so optimal Data transmission - wireless sensors (EKG) to wired base stations. 90% power savings + 75 In the news + 76 References Emmanuel Candes, Compressive Sensing - A 25 Minute Tour, First EU-US Frontiers of Engineering Symposium, Cambridge, September 2010 David Schneider, Camera Chip Makes Already-Compressed Images, IEEE Spectrum, Feb 2013 T.Strohmer. Measure what should be measured: Progress and Challenges in Compressive Sensing. IEEE Signal Processing Letters, vol.19(12): pp.887-893, 2012. Larry Hardesty, Toward practical compressed sensing, MIT news, Feb 2013 Tao Hu and Mitya Chklovvskii, Reconstruction of Sparse Circuits Using Multi-neuronal Excitation (RESCUME), Advances in Neural Information Processing Systems, 2009 http://inviewcorp.com/technology/compressive-sensing/ http://ge.geglobalresearch.com/blog/the-beauty-of-compressive-sensing/ http://www.worldindustrialreporter.com/bell-labs-create-lensless-camera-throughcompressive-sensing-technique/ http://www.lablanche-and-co.com/ + 77 THANK YOU