Last modified: 2026-11-19
Every result in the notes, listed by module and page in the order the notes present them. Each row numbers the result within its page and across the whole list, so the last overall number counts every result. The list is built from the notes themselves, so it cannot fall out of step with them. The definitions are listed the same way.
| On page | Overall | Result |
|---|---|---|
| 1 | 1 | The chain rule of probability |
| On page | Overall | Result |
|---|---|---|
| 1 | 32 | The Frobenius objective is a sum over columns |
| 2 | 33 | Multivariate normal equations |
| On page | Overall | Result |
|---|---|---|
| 1 | 34 | Under squared-error loss, the empirical risk is the sample mean squared error (sMSE) |
| On page | Overall | Result |
|---|---|---|
| 1 | 35 | Hoeffding’s inequality for a held-out set |
| On page | Overall | Result |
|---|---|---|
| 1 | 87 | Separated data have no maximum likelihood estimate |
| On page | Overall | Result |
|---|---|---|
| 1 | 88 | The two-class multinomial model is logistic regression |
| 2 | 89 | Gradient of the multinomial log-likelihood |
| On page | Overall | Result |
|---|---|---|
| 1 | 90 | The posterior from a prior and class densities |
| 2 | 91 | The linear discriminant function |
| 3 | 92 | LDA implies a logistic model |
| 4 | 93 | The quadratic discriminant function |
| 5 | 94 | Quadratic discriminant boundaries are quadric surfaces |
| 6 | 95 | Parameter counts of the classifiers |
| On page | Overall | Result |
|---|---|---|
| 1 | 96 | Micro-averaged precision and recall equal accuracy |
| 2 | 97 | A score independent of the label has ROC AUC one half |
| On page | Overall | Result |
|---|---|---|
| 1 | 98 | Number of models that best subset selection fits |
| On page | Overall | Result |
|---|---|---|
| 1 | 99 | A gradient step on the ridge objective is weight decay |
| 2 | 100 | The coordinate update for the lasso |
| 3 | 101 | Ridge and the lasso with orthonormal features |
| 4 | 102 | The coefficients of a dimension reduction regression |
| On page | Overall | Result |
|---|---|---|
| 1 | 103 | A larger model family never has a larger minimum empirical risk |
| 2 | 104 | The variance of the cross-validation estimate in terms of fold correlation |
| On page | Overall | Result |
|---|---|---|
| 1 | 118 | XOR is not linearly separable |
| 2 | 119 | The product feature separates XOR |
| 3 | 120 | Number of polynomial features |
| 4 | 121 | Parameters of a one-hidden-layer network |
| On page | Overall | Result |
|---|---|---|
| 1 | 122 | The gradient of the last layer’s weights is an outer product |
| 2 | 123 | The gradient of the first layer’s weights is an outer product |
| 3 | 124 | Derivative of the hyperbolic tangent |
| On page | Overall | Result |
|---|---|---|
| 1 | 125 | \(K\)-means never raises the distortion |
| On page | Overall | Result |
|---|---|---|
| 1 | 126 | The VC generalization bound |
| On page | Overall | Result |
|---|---|---|
| 1 | 127 | Parameters of a cubic spline |
| On page | Overall | Result |
|---|---|---|
| 1 | 128 | The backfitting update for one component |
| On page | Overall | Result |
|---|---|---|
| 1 | 129 | Conditioning a scalar Gaussian |
| 2 | 130 | Gaussian process regression |
| On page | Overall | Result |
|---|---|---|
| 1 | 131 | Output size of a convolution |
| 2 | 132 | Jacobian of the state of a tanh recurrent layer |
| 3 | 133 | Jacobian of a tanh recurrent layer |
| 4 | 134 | The backward recursion and the shared-weight gradient |
| On page | Overall | Result |
|---|---|---|
| 1 | 135 | Perplexity does not depend on the base |