1.41k likes | 1.74k Views
Prelude. A pattern of activation in a NN is a vector A set of connection weights between units is a matrix Vectors and matrices have well-understood mathematical and geometric properties Very useful for understanding the properties of NNs. Operations on Vectors and Matrices. Outline.
E N D
Prelude • A pattern of activation in a NN is a vector • A set of connection weights between units is a matrix • Vectors and matrices have well-understood mathematical and geometric properties • Very useful for understanding the properties of NNs
Outline • The Players: Scalars, Vectors and Matrices • Vectors, matrices and neural nets • Geometric Analysis of Vectors • Multiplying Vectors by Scalars • Multiplying Vectors by Vectors • The inner product (produces a scalar) • The outer product (produces a matrix) • Multiplying Vectors by Matrices • Multiplying Matrices by Matrices
Scalars, Vectors and Matrices • Scalar: A single number (integer or real) • Vector: An ordered list of scalars [ 1 2 3 4 5 ] [ 0.4 1.2 0.07 8.4 12.3 ] [ 12 10 ] [ 2 ]
Scalars, Vectors and Matrices • Scalar: A single number (integer or real) • Vector: An ordered list of scalars [ 1 2 3 4 5 ] [ 0.4 1.2 0.07 8.4 12.3 ] [ 12 10 ] [ 2 ] [ 12 10 ] ≠ [ 10 12 ]
Scalars, Vectors and Matrices • Scalar: A single number (integer or real) • Vector: An ordered list of scalars [ 1 2 3 4 5 ] [ 0.4 1.2 0.07 8.4 12.3 ] [ 12 10 ] [ 2 ] Row vectors
1 2 3 4 5 1.5 0.3 6.2 12.0 17.1 Scalars, Vectors and Matrices • Scalar: A single number (integer or real) • Vector: An ordered list of scalars [ 1 2 3 4 5 ] [ 0.4 1.2 0.07 8.4 12.3 ] [ 12 10 ] [ 2 ] Row vectors Column Vectors
Scalars, Vectors and Matrices Scalar: A single number (integer or real) Vector: An ordered list of scalars Matrix: An ordered list of vectors: 1 2 6 1 7 8 2 5 9 0 0 3 3 1 5 7 6 3 2 7 9 3 3 1
Scalars, Vectors and Matrices Scalar: A single number (integer or real) Vector: An ordered list of scalars Matrix: An ordered list of vectors: 1 2 6 1 7 8 2 5 9 0 0 3 3 1 5 7 6 3 2 7 9 3 3 1 Row vectors
Scalars, Vectors and Matrices Scalar: A single number (integer or real) Vector: An ordered list of scalars Matrix: An ordered list of vectors: 1 2 6 1 7 8 2 5 9 0 0 3 3 1 5 7 6 3 2 7 9 3 3 1 Column vectors
Scalars, Vectors and Matrices Scalar: A single number (integer or real) Vector: An ordered list of scalars Matrix: An ordered list of vectors: 1 2 6 1 7 8 2 5 9 0 0 3 3 1 5 7 6 3 2 7 9 3 3 1 Matrices are indexed (row, column) M =
Scalars, Vectors and Matrices Scalar: A single number (integer or real) Vector: An ordered list of scalars Matrix: An ordered list of vectors: 1 2 6 1 7 8 2 5 9 0 0 3 3 1 5 7 6 3 2 7 9 3 3 1 Matrices are indexed (row, column) M(1,3) = 6 (row 1, column 3) M =
Scalars, Vectors and Matrices Scalar: A single number (integer or real) Vector: An ordered list of scalars Matrix: An ordered list of vectors: 1 2 6 1 7 8 2 5 9 0 0 3 3 1 5 7 6 3 2 7 9 3 3 1 Matrices are indexed (row, column) M(1,3) = 6 (row 1, column 3) M(3,1) = 3 (row 3, column 1) M =
Variable Naming Conventions • Scalars: Lowercase, italics x, y, z… • Vectors: Lowercase, bold u, v, w… • Matrices: Uppercase, bold M, N, O … • Constants: Greek , , , , …
Transposing Vectors If u is a row vector… u = [ 1 2 3 4 5 ] …then u’ (“u-transpose”) is a column vector 1 2 3 4 5 … and vice-versa. u’ =
Transposing Vectors If u is a row vector… u = [ 1 2 3 4 5 ] …then u’ (“u-transpose”) is a column vector 1 2 3 4 5 … and vice-versa. Why in the world would I care?? u’ = You
Transposing Vectors If u is a row vector… u = [ 1 2 3 4 5 ] …then u’ (“u-transpose”) is a column vector 1 2 3 4 5 … and vice-versa. Answer: It’ll matter when we come to vector multiplication. u’ =
Transposing Vectors If u is a row vector… u = [ 1 2 3 4 5 ] …then u’ (“u-transpose”) is a column vector 1 2 3 4 5 … and vice-versa. OK. u’ =
Vectors, Matrices & Neural Nets j1 j2 j3 Input units, j
Vectors, Matrices & Neural Nets i1 i2 Output units, i j1 j2 j3 Input units, j
Vectors, Matrices & Neural Nets i1 i2 Output units, i Connection weights, wij w23 w11 w13 w12 w22 w21 j1 j2 j3 Input units, j
Vectors, Matrices & Neural Nets i1 i2 Output units, i Connection weights, wij w23 w11 w13 w12 w22 w21 0.2 0.9 0.5 Input units, j The activations of the input nodes can be represented as a 3-dimensional vector: j = [ 0.2 0.9 0.5 ]
Vectors, Matrices & Neural Nets 1.0 0.0 Output units, i Connection weights, wij w23 w11 w13 w12 w22 w21 j1 j2 j3 Input units, j The activations of the output nodes can be represented as a 2-dimensional vector: i = [ 1.0 0.0 ]
Vectors, Matrices & Neural Nets i1 i2 Output units, i Connection weights, wij w23 w11 0.1 0.2 w13 w12 1.0 w22 w21 j1 j2 j3 Input units, j The weights leading into any output node can be represented as a 3-dimensional vector: w1j = [ 0.1 1.0 0.2 ]
Vectors, Matrices & Neural Nets i1 i2 Output units, i Connection weights, wij w23 -0.9 w11 0.1 0.2 w13 w12 1.0 w22 0.1 w21 1.0 j1 j2 j3 Input units, j The complete set of weights can be represented as a 3 (row) X 2 (column) matrix: 0.1 1.0 0.21.0 0.1 -0.9 W =
Vectors, Matrices & Neural Nets i1 i2 Output units, i Connection weights, wij Why in the world would I care?? w23 -0.9 w11 0.1 0.2 w13 w12 1.0 w22 0.1 w21 1.0 j1 j2 j3 Input units, j The complete set of weights can be represented as a 2 (row) X 3 (column) matrix: 0.1 1.0 0.21.0 0.1 -0.9 W =
Vectors, Matrices & Neural Nets Because the mathematics of vectors and matrices is well-understood. Because vectors have a very useful geometric interpretation. Because Matlab “thinks” in vectors and matrices. Because you are going to have to learn to think in Matlab. Why in the world would I care?? W
Vectors, Matrices & Neural Nets Because the mathematics of vectors and matrices is well-understood. Because vectors have a very useful geometric interpretation. Because Matlab “thinks” in vectors and matrices. Because you are going to have to learn to think in Matlab. OK.
Geometric Analysis of Vectors Dimensionality: The number of numbers in a vector
Geometric Analysis of Vectors Dimensionality: The number of numbers in a vector
Geometric Analysis of Vectors Dimensionality: The number of numbers in a vector
Geometric Analysis of Vectors • Implications for neural networks • Auto-associative nets • State of activation at time t is a vector (a point in a space) • As activations change, vector moves through that space • Will prove invaluable in understanding Hopfield nets • Layered nets (“perceptrons”) • Input vectors activate output vectors • Points in input space map to points in output space • Will prove invaluable in understanding perceptrons and back-propagation learning
Multiplying a Vector by a Scalar 4 5 [ 5 4 ] * 2 =
Multiplying a Vector by a Scalar 8 4 10 5 [ 5 4 ] * 2 = [ 10 8 ] Lengthens the vector but does not change its orientation
Adding a Vector to a Scalar 4 5 [ 5 4 ] + 2 =
Adding a Vector to a Scalar 4 5 [ 5 4 ] + 2 = NAN Is Illegal.
Adding a Vector to a Vector 6 4 5 3 [ 5 4 ] + [ 3 6 ] =
Adding a Vector to a Vector 10 6 4 5 3 8 [ 5 4 ] + [ 3 6 ] = [ 8 10 ] Forms a parallelogram.
Multiplying a Vector by a Vector 1:The Inner Product (aka “Dot Product”) If u and v are both row vectors of the same dimensionality… u = [ 1 2 3 ] v = [ 4 5 6 ]
Multiplying a Vector by a Vector 1:The Inner Product (aka “Dot Product”) If u and v are both row vectors of the same dimensionality… u = [ 1 2 3 ] v = [ 4 5 6 ] … then the product u ·v =
Multiplying a Vector by a Vector 1:The Inner Product (aka “Dot Product”) If u and v are both row vectors of the same dimensionality… u = [ 1 2 3 ] v = [ 4 5 6 ] … then the product u ·v = NAN Is undefined.
Multiplying a Vector by a Vector 1:The Inner Product (aka “Dot Product”) If u and v are both row vectors of the same dimensionality… u = [ 1 2 3 ] v = [ 4 5 6 ] … then the product u ·v = NAN Is undefined. Huh?? Why?? That’s BS!
Multiplying a Vector by a Vector 1:The Inner Product (aka “Dot Product”) I told you you’d eventually care about transposing vectors… ?
Multiplying a Vector by a Vector 1:The Inner Product (aka “Dot Product”) • The Mantra: “Rows by Columns” • Multiply rows (or row vectors) by columns (or column vectors)
Multiplying a Vector by a Vector 1:The Inner Product (aka “Dot Product”) • The Mantra: “Rows by Columns” • Multiply rows (or row vectors) by columns (or column vectors) u = [ 1 2 3 ] v = [ 4 5 6 ]
Multiplying a Vector by a Vector 1:The Inner Product (aka “Dot Product”) • The Mantra: “Rows by Columns” • Multiply rows (or row vectors) by columns (or column vectors) 4 5 6 u = [ 1 2 3 ] v = [ 4 5 6 ] v’ =
Multiplying a Vector by a Vector 1:The Inner Product (aka “Dot Product”) • The Mantra: “Rows by Columns” • Multiply rows (or row vectors) by columns (or column vectors) 4 5 6 u = [ 1 2 3 ] v’ = u ·v’ = 32
Multiplying a Vector by a Vector 1:The Inner Product (aka “Dot Product”) • The Mantra: “Rows by Columns” • Multiply rows (or row vectors) by columns (or column vectors) v’ Imagine rotating your row vector into a (pseudo) column vector… 4 5 6 1 2 3 u = [ 1 2 3 ] u ·v’ = 32
Multiplying a Vector by a Vector 1:The Inner Product (aka “Dot Product”) • The Mantra: “Rows by Columns” • Multiply rows (or row vectors) by columns (or column vectors) v’ Now multiply corresponding elements and add up the products… 4 5 6 1 2 3 4 u = [ 1 2 3 ] u ·v’ = 32