1 of 48

If you need more information, these references are helpful:�

  1. Kolda, T. G., & Bader, B. W. (2009). Tensor decompositions and applications. SIAM review, 51(3), 455-500.
  2. Matlab Tensor Toolbox: https://www.tensortoolbox.org/

The following slides contain material from Tammy Kolda, Charles Van Loan and Lars Elden

Tensor Decompositions

 

2 of 48

2

A Tensor is an 𝑑-way array

 

 

 

 

3 of 48

3

Grid Cells: Global positioning system that helps people navigate

4 of 48

4

5 of 48

5

Please Read!

6 of 48

6

Inside the Nucleus: Chromosome Territories

Tan, Longzhi, et al. "Three-dimensional genome structures of single diploid human cells." Science 361.6405 (2018): 924-928.

7 of 48

7

Some Mathematics….

8 of 48

8

The Kronecker Product

9 of 48

9

The Kronecker Product

10 of 48

10

All Possible Entry- Entry Products

11 of 48

11

Kronecker Products of Kronecker Products

12 of 48

12

You have seen them before

13 of 48

13

Block matrix

14 of 48

14

Block Matrices: Addition

15 of 48

15

Block Matrix: Multiplication

16 of 48

16

Block Matrix: Transposition

17 of 48

17

A block matrix can have block matrix entries

18 of 48

18

Reshape

19 of 48

19

Reshape

20 of 48

20

SVD Expansion

21 of 48

21

SVD

22 of 48

22

SVD

23 of 48

23

The Tucker product

Definition:

24 of 48

24

The Tucker product representation

Before…

Now…

25 of 48

25

The Higher-Order SVD (HOSVD)

26 of 48

26

The HOSVD of a matrix

27 of 48

27

The HOSVD of a matrix

28 of 48

28

Notations

29 of 48

29

Tensor Unfolding

30 of 48

30

A Common Framework for Tensor Computations

Tensor Unfolding

31 of 48

31

Illustration of row-wise and column-wise unfolding (flattening, matricizing) of a third-order tensor

Andrzej Cichocki, Rafal Zdunek, and Shun'ichi Amari. Nonnegative Matrix and Tensor Factorizations: Applications to Exploratory Multi-way Data Analysis and Blind Source (2009)

32 of 48

32

Unfolding (matricizing) of a third-order tensor

33 of 48

Rank One Tensors

Building block for decomposition: Vector outer products = Rank one tensors

Matrix version (2 – way)

Rank one matrix

Tensor version (3 – way)

Rank one tensor

34 of 48

Building block for decomposition: Vector outer products = Rank one tensors

 

Visualization

gets weird.. Math is still good!

Rank one Tensors

35 of 48

35

Data

Low-rank model

Matrix notation

Sum of squared errors

Low Rank Approximation

36 of 48

Tensor Decompositions

CP Model: Sum of d-way outer products, useful for interpretation

Tucker Model: Project onto high-variance subspaces to reduce dimensionality

37 of 48

CP Decomposition

The full name is CANDECOMP/PARAFAC Decomposition

CANDECOMP = Canonical Decomposition

PARAFAC = Parallel Factors Decomposition

CP Tensor Factorization (3-way): Detecting low-rank 3-way structure

Higher-order analogue of the SVD

CP first invented in 1927

F. L. Hitchcock, The Expression of a Tensor or a Polyadic as a Sum of Products, Journal of Mathematics and Physics, 1927

38 of 48

Tensor notation

Sum of squared errors

Factor matrices

Component

CP Decomposition

39 of 48

Example 1: Neuronal Activity

Williams, A. H., Kim, T. H., Wang, F., Vyas, S., Ryu, S. I., Shenoy, K. V., ... & Ganguli, S. (2018). Unsupervised discovery of demixed, low-dimensional neural dynamics across multiple timescales through tensor component analysis. Neuron, 98(6), 1099-1115.

40 of 48

Example 1: Neuronal activity

Mouse

in “maze”

Neural activity

One Column of Neuron x

Time matrix

 

× 600 trials (over 5 days)

Microscope by

Inscopix

41 of 48

Trials vary start position and strategies

  • 600 Trials over 5 days
  • Start West or East
  • Conditions swap twice
    • Always turn South
    • Always turn Right
    • Always turn South

Example: Neuronal activity

42 of 48

8-Component CP decomposition of data

43 of 48

8-Component CP decomposition of data

44 of 48

Got the treat! Happy neurons….

8-Component CP decomposition of data

45 of 48

8-Component CP decomposition of data

Clear separation. Starting position Neurons understand mouse has started. ….

46 of 48

8-Component CP decomposition of data

47 of 48

Example 2: Application to hazardous gas

 

Vergara, A., Fonollosa, J., Mahiques, J., Trincavelli, M., Rulkov, N., & Huerta, R. (2013). On the performance of gas sensor arrays in open sampling systems using Inhibitory Support Vector Machines. Sensors and Actuators B: Chemical, 185, 462-477.

48 of 48

Factors

Example 2: Application to hazardous gas