AI Solutions for Biomedical and Industrial Data From Data to Decisions

Transcription

1 AI Soutions for Biomedica and Industria Data From Data to Decisions Moncef Gabbouj Laboratory of Signa Processing Tampere University of Technoogy

2 From Data to Decision Our Approach Our Approach Surveiance Traffic RQ-3 RQ-2 RQ- Moncef Gabbouj

3 I) Optimization in Machine Learning: Muti-dimensiona Partice Swarm Optimization (MD-PSO) Goa: A new optimization agorithm based on partice swarms which finds the optima soution at the optima dimension (it can be appied to optimization in mutidimensiona spaces where the dimension of the soution space is not know a priori). 7 xd ( t) xd9( t) 3 d=2 d=3 gbest(3) gbest(2) S. Kiranyaz, T. Ince, A. Yidirim and M. Gabbouj, Fractiona Partice Swarm Optimization in Muti-Dimensiona Search Space, IEEE Trans. on Systems, Man, and Cybernetics Part B, pp , vo. 40, No. 2, Apri 200. MD PSO (dbest) Go to d =23 a OK! xd a ( t) 23 Optimizing over dimensions 3

4 How to dea with such compex functions? SPHERE GIUNTA RASTRIGIN GRIEWANK DEJONG ROSENBROCK 4

5 Learn to Synthesize Optima Features Signa Processing and Optimization Approach x (FS-) y (,0) 2 2 x x, x, 2 xx D à 3D Feature engineering: Feature extraction Feature seection Feature synthesis cass- cass-2 (FS-2) sin( 2 fx) D à D (,0) 0 y cass- cass-2 Ground Truth Image Database FeX FV MD-PSO based Feature Synth. Synt. FV () Synt. FV (R-) MD-PSO based Feature Synth. Synt. FV (R) Fitness Eva. (-AP) Raitoharju, Neura Computing and Appications, 206. Moncef Gabbouj

6 Network Architecture Design: Evoutionary Artificia Neura Networks Question: How to design optima neura networks? MD-PSO xx a xd a ( t) ( t) 0 { w jk },{ w jk },{ k},{ w O O,...,{ w jk },{ k },{ k 2 jk O 2 },{ } k } Kiranyaz, Neura Networks, (top 5 th downoaded paper from Esevier Journa) Moncef Gabbouj

7 Beat Cass Type Patient-specific Cassification of ECG Data by Evoutionary ANNs Common data: 200 beats Patient X Patient-specific data: first 5 min. beats Expert Labeing Training Labes per beat MD PSO: Evoution + Training Data Acquisition Beat Detection Morph. Feature Extraction (TI-DWT) Dimension Reduction (PCA) Tempora Features ANN Space Ince, IEEE Transactions on Biomedica Engineering,

8 Advanced Patter Recognition and Machine Learning: Coective Network of Binary Cassifiers Supervised sematic cassifier Evoutionary cassifier (deep rooted in mathematica optimization) Scaabe wrt both casses and features (suitabe for Big Data) Incrementa as opposed to static cassifiers Dataset Features { FV0, FV,..., FV N } FV0 FV FV N FV0 FV FV N FV0 FV FV N BC0 BC BC N 0 FV0 FV FV BC BC N N BC BC0 BC BC N Fuser NBC 0 Fuser NBC Fuser NBC C CNBC CV 0 CV CV C Cass Seection Kiranyaz, Neura Networks, 202 Kiranyaz, IEEE Transactions on Systems, Man and Cybernetics - Part B: Cybernetics, 202 { c * } Gabbouj/ 2/4/208 8

9 EEG Cassification by Incrementa CNBC Evoution EEG Record EEG Labes Neuroogist Labeing CNBC Evoution Patient X BC 0 BC NBC 0 CV 0 EEG CNBC An Eary EEG Record BC N Feature Extraction + Norm. Normaized Feature Vectors BC 0 BC BC N NBC CV BC 0 NBC 7 BC CV 7 BC N Kiranyaz, Journa of Biomedica Informatics,

10 Synthetic Aperture Radar (SAR) Data Anaysis and Cassification Kiranyaz, IEEE Trans. on Systems, Man, and Cybernetics Part B, 202. Uhmann, IEEE Trans. on Geoscience & Remote Sensing, 204.

11 The CNBC test-bed appication GUI showing a sampe user-defined ground truth set over San Francisco Bay area. Mutimedia Group

12 CET- Lighthouse Stadium Mutimedia Group Prof. S. Kiranyaz

13 Patient-Specific Seizure Detection Using Noninear Dynamics and Nucides Average sensitivity 9.5%; average specificity 95.6% on CHB-MIT Database Zabihi, Journa of Biomedica and Heath Informatics, 208 (under review)

14 FACE SEGMENTATION IN THUMBNAIL IMAGES BY DATA-ADAPTIVE CONVOLUTIONAL SEGMENTATION NETWORKS Input Layer R 2nd Layer (56, 56) G (54, 54) 3rd Layer Input Image B (56, 56) (52, 52) Output Layer Lapacian (56, 56) (400, 300) (56, 56) (50, 50) -R (56, 56) -G (56, 56) -B (56, 56) Kiranyaz, Proc. IEEE ICIP, 206, Phoenix, Arizona. Moncef Gabbouj

15 Image L2S Truth Image L2S Truth Some Initia Resuts on the Test dataset 2/4/208 5

16 Tabe : Mean performances of the FCN- 32s VGG [9] and the proposed L2S with singe and 5 CSNs. Kiranyaz, Proc. IEEE ICIP, 206, Phoenix, Arizona. 2/4/208 6

17 Combining Quatum Mechanics and Spectra Graph Theory: Quantum-Cut for Saient Object Detection Aytekin Pattern Recognition, 206 Aytekin Pattern Recognition Letters, 206 Aytekin ICPR 204. IBM Best Paper Award Aytekin, Best Nordic Thesis Award, 207 Moncef Gabbouj 7

18 Object Proposas and Labeing Image Labeing

19 Combining Deep Learning and Signa Processing Object Proposas Waris, Neurocomputing, 207. Moncef Gabbouj

20 Object(s) Recognition & Locaization

21 Person Car Bicyce & Bus Detection Gabbouj/ 2/4/208

22 Rea-Time Motor Faut Detection by -D Convoutiona Neura Networks Rea-time Motor Condition Monitoring MC Data & Labes Offine Training D CNN Back-Propagation Pre- Processing motor current signa Motor Cass Type Kiranyaz, IEEE Transaction on Industria Eectronics, /4/208 22

23 Rea-time Motor Faut Detection by D CNN C, CNN i C,2 C,N U i UN i C 2, D i DN i C 2,2 C 2,N2 U o UN o C 3, D o DN o C 3,2 C 3,N3 CNN o 240 msec to detect and cassify fauts from a 2-sec signa samped at 2.8kHz (5min for training) Kiranyaz, IEEE Transaction on Industria Eectronics, 206. Moncef Gabbouj

24 Rea-Time Patient-Specific ECG Cassification by D CNN Raw Beat Sampes Training (Offine) Common data: 200 beats Patient X Patient-specific data: first 5 min. beats Training Labes per beat D CNN Back-Propagation 0.5 Data Acquisition Beat Detection Beat Cass Type Upoad Rea-time Monitoring Patient X Aert Kiranyaz, IEEE Transactions on Biomedica Engineering, 205. Moncef Gabbouj

25 Rea-Time Patient-Specific ECG Cassification by D CNN Kiranyaz, IEEE Transactions on Biomedica Engineering, 206. Moncef Gabbouj

26 Advance Warning for Cardiac Arrhythmia Kiranyaz, Scientific Reports, 207. (Springer Nature) Moncef Gabbouj

27 Advance Warning for Cardiac Arrhythmia a) Modeing common causes of cardiac arrhythmia in the signa domain. b) Symboic iustration of an abnorma S beat synthesis for Person-Y using the degrading system designed from the ECG data of severa Patient(s)-X. c) Iustration of the overa system, where a dedicated CNN is trained by Backpropagation over the training dataset created for Person-Y (top). Once the D CNN is trained, it can then be used as a continuous cardiac heath monitoring and advance warning system (bottom) for Person-Y. Moncef Gabbouj

28 Resuts 34 patient with 63,34 ECG beats were used in the evauation. Sen = 82.5%, Spe = 99.55%, Ppr = 95.55%, Acc = 97.75% and FAR = 0.45%, Average probabiity of missing the first abnorma beat is 0.74 The average probabiity of missing three consecutive abnorma beats is around Therefore, detecting one or more abnorma beat(s) among the first three occurrences is highy probabe (>99.4%) Moncef Gabbouj

29 Advanced Network Architecture Design: Generaized Operationa Perceptrons (GOPS) 2 j N- Input 2 y y Layer - 2 y i y N Layer- 2 ( y k ( y k ( y k ( y k Layer-2 2, w k, w, w ) 2k i ik N Layer-3 2, w ) ) N k ) Layer-4 6 x POPmax hmax = 4 k k b k k y k x k b f k Layer f k ( k xk ) k k y k..., k ( y, w i ik P, f F, Output k PF in GOPmin() (min. MSE = 0.22) Input 2 Layer- 2 k Output We proposed a Progressive Operationa Feedforward Neura network earning approach ),... Data-driven network s architecture Data-driven network s parameters tuning PF in GOPmin(2) (min. MSE = 0.02) Input 2 Kiranyaz, Neurocomputing, 207. Layer-2 Output PF in GOPmin(3) (min. MSE = 0-4 ) Input Layer-3 2 Output Input 2 Layer- 2 Layer-2 Fina POP Layer-3 2 Output Kiranyaz (206) IEEE ICIP, Phoenix, Arizona

30 AARHUS UNIVERSITET Peter Gorm Larsen Professor Efficient CNN Design (Deep Learning). OKTOBER 200. Aexandros Iosifidis Tensor-based CNN fiters modeing: Empoying mutiinear projection as the primary feature extractor Lower number of network parameters (reduced memory footprint) Faster cassification (reduced computationa cost) LR: Low-Rank Regression, Tai arxiv 205 D.T. Thanh, A. Iosifidis and M. Gabbouj, Improving Efficiency in Convoutiona Neura Network with Mutiinear Fiters, Neura Networks, 208 (arxiv 207)

31 AARHUS UNIVERSITET Peter Gorm Larsen Professor Long- and Short-memory Neura BoFs. OKTOBER Tempora-aware NN BoF mode for time-series data anaysis N. Passais, A. Tefas, J. Kanniainen, M. Gabbouj and A. Iosifidis, Tempora Bag-of-Features earning for Predicting Mid Price Movements using High Frequency Limit Order Book Data, IEEE Transactions on Emerging Topics in Computationa Inteigence, 209

32 AARHUS UNIVERSITET Attention in Muti-inear Networks (AMLN) Peter Gorm Larsen Professor Aexandros Iosifidis. OKTOBER Date AMLN architecture incorporates biinear projection and an attention mechanism to detect and focus on tempora information. AMLN is highy interpretabe, abe to highight the importance and contribution of each tempora instance, two-hidden-ayer network achives state-of-the-art performance on a argescae Limit Order Book (LOB) dataset D.T. Tran, A. Iosifidis, J. Kanniainen and M. Gabbouj, Tempora Attention augmented Biinear Network for Financia Time-Series Data Anaysis, IEEE Transactions on Neura Networks and Learning Systems, 209

33 Mutimoda and Cross-moda Learning Gao, IEEE Transactions on Cybernetics, 208. Moncef Gabbouj

34 IEEE NER 205 BCI CHALLENGE 205: (3 rd PLACE) OBJECTIVE: P300-Speer is a we-known brain-computer interface paradigm and has great potentia to hep individuas with neuromuscuar disabiities to communicate. The goa of this chaenge was to detect errors during the speing task by anayzing the subjects EEG recording. METHODOLOGY: The BCI chaenge@ner205 dataset is used. 56 EEG channes are used. 3 different features of approximation coefficients of Daubechies 4 waveets in 5-eve decomposition are extracted (base-eve features). The base-eve features are cassified by 56 different inear discriminant anaysis (LDA) cassifiers (for each channe) to obtain the metaeve features which are the posterior probabiities (of the LDA cassifiers). For cassification of (56-dimention) feature vectors a feedforward neura network with two hidden ayers; 5 neurons in the first and 7 neurons in the second hidden ayer is used. RESULTS: The area under the ROC curve of 0.78 was achieved for the proposed approach (this was 0.69 in the private eaderboard). Moncef Gabbouj

35 PHYSIONET CHALLENGE 206: (2 nd PLACE AMONG 48 TEAMS) OBJECTIVE: Heart sound has a great potentia to be used as a diagnostic test in ambuatory monitoring. The goa was to detect heart anomaies by anayzing the subject's heart sound waves. METHODOLOGY: The Physionet chaenge 206 PCG dataset is used. 8 features is seected among 40 features from time, frequency, time-frequency domains. Wrapper-based feature seection scheme using sequentia forward seection search agorithm is used for feature seection. 20 ANN were used with two hidden ayers in each, and 25 hidden neurons at each ayer. RESULTS: Train Dataset 0-fod crossvaidation Evauation Metrics Sen (%) Spe (%) Score (%) Ave Train Evauation Metrics Dataset 0-fod crossvaidation Sen (%) Spe (%) Score (%) Ave Moncef Gabbouj

36 PHYSIONET CHALLENGE 207: ( st PLACE AMONG 75 TEAMS) OBJECTIVE: The goa of this chaenge was to detect Atria Fibriation (AF) rhythm using hand-hed ECG monitoring devices, in addition to three other casses: norma or sinus rhythm, other rhythms, and too noisy to anayze. METHODOLOGY: The Physionet chaenge 207 ECG dataset is used. 49 hand-crafted muti-domain features are extracted. 50 features are seected using random forest. Hybrid cassification (base-eve + meta-eve) earning is used: RESULTS: Evauation Metrics Ave. on Train Dataset (%) 0-fod crossvaidation Ave. on Test Dataset (%) Fn (Norma) Fa (AF) Fo (Other) Fp (Noisy) 6. F(Tota) Moncef Gabbouj

37 University Partners NSF-Business Finand-IUCRC-CVDI Industry Advisory Board (IAB) Financia Support, Research Direction, Guidance Coaborative Research Approach Research Expertise & Institutiona Support (Infrastructure, Students, etc. ) Visua Anaytics Research Projects Predictive Anaytics Data Ceaning Interactive Visuaization Areas of Focus Data Summarization Socia Media Anaytics Core and Suppementa Funds, IUCRC Governance Funding Agencies Research Technoogy Portfoio Vaue Created Cooperative Technoogy Transfer Royaty Free Licenses Pubication Access Technoogy Breakthroughs Coaborative Resuts University Aignments Facuty/Student Access Industry Partnerships Investment in Future Trained Workforce Recruitment Opportunities Professiona Deveopment Vaue Return Research Cost Savings 90% funds dedicated to research >40: Return on Investment Founded in 202

38 Recent books in the fied

39 Summary Nove methods and agorithms for artificia inteigence deepy rooted in signa processing and pattern recognition, New machine earning techniques deveoped based on the specific properties of the probems at hand, Data-to-Decision Research Community in TUT gathered a critica mass to make sizeabe contributions to the fied of AI, Moncef Gabbouj

40 Mutimedia Research Group Moncef Gabbouj