chapter_linear-networks/softmax-regression.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "f5061218",
   "metadata": {
    "origin_pos": 0
   },
   "source": [
    "# Softmax Regression\n",
    ":label:`sec_softmax`\n",
    "\n",
    "In :numref:`sec_linear_regression`, we introduced linear regression,\n",
    "working through implementations from scratch in :numref:`sec_linear_scratch`\n",
    "and again using high-level APIs of a deep learning framework\n",
    "in :numref:`sec_linear_concise` to do the heavy lifting.\n",
    "\n",
    "Regression is the hammer we reach for when\n",
    "we want to answer *how much?* or *how many?* questions.\n",
    "If you want to predict the number of dollars (price)\n",
    "at which a house will be sold,\n",
    "or the number of wins a baseball team might have,\n",
    "or the number of days that a patient will remain hospitalized before being discharged,\n",
    "then you are probably looking for a regression model.\n",
    "\n",
    "In practice, we are more often interested in *classification*:\n",
    "asking not \"how much\" but \"which one\":\n",
    "\n",
    "* Does this email belong in the spam folder or the inbox?\n",
    "* Is this customer more likely *to sign up* or *not to sign up* for a subscription service?\n",
    "* Does this image depict a donkey, a dog, a cat, or a rooster?\n",
    "* Which movie is Aston most likely to watch next?\n",
    "\n",
    "Colloquially, machine learning practitioners\n",
    "overload the word *classification*\n",
    "to describe two subtly different problems:\n",
    "(i) those where we are interested only in\n",
    "hard assignments of examples to categories (classes);\n",
    "and (ii) those where we wish to make soft assignments,\n",
    "i.e., to assess the probability that each category applies.\n",
    "The distinction tends to get blurred, in part,\n",
    "because often, even when we only care about hard assignments,\n",
    "we still use models that make soft assignments.\n",
    "\n",
    "\n",
    "## Classification Problem\n",
    ":label:`subsec_classification-problem`\n",
    "\n",
    "To get our feet wet, let us start off with\n",
    "a simple image classification problem.\n",
    "Here, each input consists of a $2\\times2$ grayscale image.\n",
    "We can represent each pixel value with a single scalar,\n",
    "giving us four features $x_1, x_2, x_3, x_4$.\n",
    "Further, let us assume that each image belongs to one\n",
    "among the categories \"cat\", \"chicken\", and \"dog\".\n",
    "\n",
    "Next, we have to choose how to represent the labels.\n",
    "We have two obvious choices.\n",
    "Perhaps the most natural impulse would be to choose $y \\in \\{1, 2, 3\\}$,\n",
    "where the integers represent $\\{\\text{dog}, \\text{cat}, \\text{chicken}\\}$ respectively.\n",
    "This is a great way of *storing* such information on a computer.\n",
    "If the categories had some natural ordering among them,\n",
    "say if we were trying to predict $\\{\\text{baby}, \\text{toddler}, \\text{adolescent}, \\text{young adult}, \\text{adult}, \\text{geriatric}\\}$,\n",
    "then it might even make sense to cast this problem as regression\n",
    "and keep the labels in this format.\n",
    "\n",
    "But general classification problems do not come with natural orderings among the classes.\n",
    "Fortunately, statisticians long ago invented a simple way\n",
    "to represent categorical data: the *one-hot encoding*.\n",
    "A one-hot encoding is a vector with as many components as we have categories.\n",
    "The component corresponding to particular instance's category is set to 1\n",
    "and all other components are set to 0.\n",
    "In our case, a label $y$ would be a three-dimensional vector,\n",
    "with $(1, 0, 0)$ corresponding to \"cat\", $(0, 1, 0)$ to \"chicken\",\n",
    "and $(0, 0, 1)$ to \"dog\":\n",
    "\n",
    "$$y \\in \\{(1, 0, 0), (0, 1, 0), (0, 0, 1)\\}.$$\n",
    "\n",
    "## Network Architecture\n",
    "\n",
    "In order to estimate the conditional probabilities associated with all the possible classes,\n",
    "we need a model with multiple outputs, one per class.\n",
    "To address classification with linear models,\n",
    "we will need as many affine functions as we have outputs.\n",
    "Each output will correspond to its own affine function.\n",
    "In our case, since we have 4 features and 3 possible output categories,\n",
    "we will need 12 scalars to represent the weights ($w$ with subscripts),\n",
    "and 3 scalars to represent the biases ($b$ with subscripts).\n",
    "We compute these three *logits*, $o_1, o_2$, and $o_3$, for each input:\n",
    "\n",
    "$$\n",
    "\\begin{aligned}\n",
    "o_1 &= x_1 w_{11} + x_2 w_{12} + x_3 w_{13} + x_4 w_{14} + b_1,\\\\\n",
    "o_2 &= x_1 w_{21} + x_2 w_{22} + x_3 w_{23} + x_4 w_{24} + b_2,\\\\\n",
    "o_3 &= x_1 w_{31} + x_2 w_{32} + x_3 w_{33} + x_4 w_{34} + b_3.\n",
    "\\end{aligned}\n",
    "$$\n",
    "\n",
    "We can depict this calculation with the neural network diagram shown in :numref:`fig_softmaxreg`.\n",
    "Just as in linear regression, softmax regression is also a single-layer neural network.\n",
    "And since the calculation of each output, $o_1, o_2$, and $o_3$,\n",
    "depends on all inputs, $x_1$, $x_2$, $x_3$, and $x_4$,\n",
    "the output layer of softmax regression can also be described as fully-connected layer.\n",
    "\n",
    "![Softmax regression is a single-layer neural network.](../img/softmaxreg.svg)\n",
    ":label:`fig_softmaxreg`\n",
    "\n",
    "To express the model more compactly, we can use linear algebra notation.\n",
    "In vector form, we arrive at\n",
    "$\\mathbf{o} = \\mathbf{W} \\mathbf{x} + \\mathbf{b}$,\n",
    "a form better suited both for mathematics, and for writing code.\n",
    "Note that we have gathered all of our weights into a $3 \\times 4$ matrix\n",
    "and that for features of a given data example $\\mathbf{x}$,\n",
    "our outputs are given by a matrix-vector product of our weights by our input features\n",
    "plus our biases $\\mathbf{b}$.\n",
    "\n",
    "\n",
    "## Parameterization Cost of Fully-Connected Layers\n",
    ":label:`subsec_parameterization-cost-fc-layers`\n",
    "\n",
    "As we will see in subsequent chapters,\n",
    "fully-connected layers are ubiquitous in deep learning.\n",
    "However, as the name suggests,\n",
    "fully-connected layers are *fully* connected\n",
    "with potentially many learnable parameters.\n",
    "Specifically,\n",
    "for any fully-connected layer\n",
    "with $d$ inputs and $q$ outputs,\n",
    "the parameterization cost is $\\mathcal{O}(dq)$,\n",
    "which can be prohibitively high in practice.\n",
    "Fortunately,\n",
    "this cost \n",
    "of transforming $d$ inputs into $q$ outputs\n",
    "can be reduced to $\\mathcal{O}(\\frac{dq}{n})$,\n",
    "where the hyperparameter $n$ can be flexibly specified\n",
    "by us to balance between parameter saving and model effectiveness in real-world applications :cite:`Zhang.Tay.Zhang.ea.2021`.\n",
    "\n",
    "\n",
    "\n",
    "\n",
    "\n",
    "## Softmax Operation\n",
    ":label:`subsec_softmax_operation`\n",
    "\n",
    "The main approach that we are going to take here\n",
    "is to interpret the outputs of our model as probabilities.\n",
    "We will optimize our parameters to produce probabilities\n",
    "that maximize the likelihood of the observed data.\n",
    "Then, to generate predictions, we will set a threshold,\n",
    "for example, choosing the label with the maximum predicted probabilities.\n",
    "\n",
    "Put formally, we would like any output $\\hat{y}_j$\n",
    "to be interpreted as the probability\n",
    "that a given item belongs to class $j$.\n",
    "Then we can choose the class with the largest output value\n",
    "as our prediction $\\operatorname*{argmax}_j y_j$.\n",
    "For example, if $\\hat{y}_1$, $\\hat{y}_2$, and $\\hat{y}_3$\n",
    "are 0.1, 0.8, and 0.1, respectively,\n",
    "then we predict category 2, which (in our example) represents \"chicken\".\n",
    "\n",
    "You might be tempted to suggest that we interpret\n",
    "the logits $o$ directly as our outputs of interest.\n",
    "However, there are some problems with directly\n",
    "interpreting the output of the linear layer as a probability.\n",
    "On one hand,\n",
    "nothing constrains these numbers to sum to 1.\n",
    "On the other hand, depending on the inputs, they can take negative values.\n",
    "These violate basic axioms of probability presented in :numref:`sec_prob`\n",
    "\n",
    "To interpret our outputs as probabilities,\n",
    "we must guarantee that (even on new data),\n",
    "they will be nonnegative and sum up to 1.\n",
    "Moreover, we need a training objective that encourages\n",
    "the model to estimate faithfully probabilities.\n",
    "Of all instances when a classifier outputs 0.5,\n",
    "we hope that half of those examples\n",
    "will actually belong to the predicted class.\n",
    "This is a property called *calibration*.\n",
    "\n",
    "The *softmax function*, invented in 1959 by the social scientist\n",
    "R. Duncan Luce in the context of *choice models*,\n",
    "does precisely this.\n",
    "To transform our logits such that they become nonnegative and sum to 1,\n",
    "while requiring that the model remains differentiable,\n",
    "we first exponentiate each logit (ensuring non-negativity)\n",
    "and then divide by their sum (ensuring that they sum to 1):\n",
    "\n",
    "$$\\hat{\\mathbf{y}} = \\mathrm{softmax}(\\mathbf{o})\\quad \\text{where}\\quad \\hat{y}_j = \\frac{\\exp(o_j)}{\\sum_k \\exp(o_k)}. $$\n",
    ":eqlabel:`eq_softmax_y_and_o`\n",
    "\n",
    "It is easy to see $\\hat{y}_1 + \\hat{y}_2 + \\hat{y}_3 = 1$\n",
    "with $0 \\leq \\hat{y}_j \\leq 1$ for all $j$.\n",
    "Thus, $\\hat{\\mathbf{y}}$ is a proper probability distribution\n",
    "whose element values can be interpreted accordingly.\n",
    "Note that the softmax operation does not change the ordering among the logits $\\mathbf{o}$,\n",
    "which are simply the pre-softmax values\n",
    "that determine the probabilities assigned to each class.\n",
    "Therefore, during prediction we can still pick out the most likely class by\n",
    "\n",
    "$$\n",
    "\\operatorname*{argmax}_j \\hat y_j = \\operatorname*{argmax}_j o_j.\n",
    "$$\n",
    "\n",
    "Although softmax is a nonlinear function,\n",
    "the outputs of softmax regression are still *determined* by\n",
    "an affine transformation of input features;\n",
    "thus, softmax regression is a linear model.\n",
    "\n",
    "\n",
    "\n",
    "## Vectorization for Minibatches\n",
    ":label:`subsec_softmax_vectorization`\n",
    "\n",
    "To improve computational efficiency and take advantage of GPUs,\n",
    "we typically carry out vector calculations for minibatches of data.\n",
    "Assume that we are given a minibatch $\\mathbf{X}$ of examples\n",
    "with feature dimensionality (number of inputs) $d$ and batch size $n$.\n",
    "Moreover, assume that we have $q$ categories in the output.\n",
    "Then the minibatch features $\\mathbf{X}$ are in $\\mathbb{R}^{n \\times d}$,\n",
    "weights $\\mathbf{W} \\in \\mathbb{R}^{d \\times q}$,\n",
    "and the bias satisfies $\\mathbf{b} \\in \\mathbb{R}^{1\\times q}$.\n",
    "\n",
    "$$ \\begin{aligned} \\mathbf{O} &= \\mathbf{X} \\mathbf{W} + \\mathbf{b}, \\\\ \\hat{\\mathbf{Y}} & = \\mathrm{softmax}(\\mathbf{O}). \\end{aligned} $$\n",
    ":eqlabel:`eq_minibatch_softmax_reg`\n",
    "\n",
    "This accelerates the dominant operation into\n",
    "a matrix-matrix product $\\mathbf{X} \\mathbf{W}$\n",
    "vs. the matrix-vector products we would be executing\n",
    "if we processed one example at a time.\n",
    "Since each row in $\\mathbf{X}$ represents a data example,\n",
    "the softmax operation itself can be computed *rowwise*:\n",
    "for each row of $\\mathbf{O}$, exponentiate all entries and then normalize them by the sum.\n",
    "Triggering broadcasting during the summation $\\mathbf{X} \\mathbf{W} + \\mathbf{b}$ in :eqref:`eq_minibatch_softmax_reg`,\n",
    "both the minibatch logits $\\mathbf{O}$ and output probabilities $\\hat{\\mathbf{Y}}$\n",
    "are $n \\times q$ matrices.\n",
    "\n",
    "## Loss Function\n",
    "\n",
    "Next, we need a loss function to measure\n",
    "the quality of our predicted probabilities.\n",
    "We will rely on maximum likelihood estimation,\n",
    "the very same concept that we encountered\n",
    "when providing a probabilistic justification\n",
    "for the mean squared error objective in linear regression\n",
    "(:numref:`subsec_normal_distribution_and_squared_loss`).\n",
    "\n",
    "\n",
    "### Log-Likelihood\n",
    "\n",
    "The softmax function gives us a vector $\\hat{\\mathbf{y}}$,\n",
    "which we can interpret as estimated conditional probabilities\n",
    "of each class given any input $\\mathbf{x}$, e.g.,\n",
    "$\\hat{y}_1$ = $P(y=\\text{cat} \\mid \\mathbf{x})$.\n",
    "Suppose that the entire dataset $\\{\\mathbf{X}, \\mathbf{Y}\\}$ has $n$ examples,\n",
    "where the example indexed by $i$\n",
    "consists of a feature vector $\\mathbf{x}^{(i)}$ and a one-hot label vector $\\mathbf{y}^{(i)}$.\n",
    "We can compare the estimates with reality\n",
    "by checking how probable the actual classes are\n",
    "according to our model, given the features:\n",
    "\n",
    "$$\n",
    "P(\\mathbf{Y} \\mid \\mathbf{X}) = \\prod_{i=1}^n P(\\mathbf{y}^{(i)} \\mid \\mathbf{x}^{(i)}).\n",
    "$$\n",
    "\n",
    "According to maximum likelihood estimation,\n",
    "we maximize $P(\\mathbf{Y} \\mid \\mathbf{X})$,\n",
    "which is\n",
    "equivalent to minimizing the negative log-likelihood:\n",
    "\n",
    "$$\n",
    "-\\log P(\\mathbf{Y} \\mid \\mathbf{X}) = \\sum_{i=1}^n -\\log P(\\mathbf{y}^{(i)} \\mid \\mathbf{x}^{(i)})\n",
    "= \\sum_{i=1}^n l(\\mathbf{y}^{(i)}, \\hat{\\mathbf{y}}^{(i)}),\n",
    "$$\n",
    "\n",
    "where for any pair of label $\\mathbf{y}$ and model prediction $\\hat{\\mathbf{y}}$ over $q$ classes,\n",
    "the loss function $l$ is\n",
    "\n",
    "$$ l(\\mathbf{y}, \\hat{\\mathbf{y}}) = - \\sum_{j=1}^q y_j \\log \\hat{y}_j. $$\n",
    ":eqlabel:`eq_l_cross_entropy`\n",
    "\n",
    "For reasons explained later on, the loss function in :eqref:`eq_l_cross_entropy`\n",
    "is commonly called the *cross-entropy loss*.\n",
    "Since $\\mathbf{y}$ is a one-hot vector of length $q$,\n",
    "the sum over all its coordinates $j$ vanishes for all but one term.\n",
    "Since all $\\hat{y}_j$ are predicted probabilities,\n",
    "their logarithm is never larger than $0$.\n",
    "Consequently, the loss function cannot be minimized any further\n",
    "if we correctly predict the actual label with *certainty*,\n",
    "i.e., if the predicted probability $P(\\mathbf{y} \\mid \\mathbf{x}) = 1$ for the actual label $\\mathbf{y}$.\n",
    "Note that this is often impossible.\n",
    "For example, there might be label noise in the dataset\n",
    "(some examples may be mislabeled).\n",
    "It may also not be possible when the input features\n",
    "are not sufficiently informative\n",
    "to classify every example perfectly.\n",
    "\n",
    "### Softmax and Derivatives\n",
    ":label:`subsec_softmax_and_derivatives`\n",
    "\n",
    "Since the softmax and the corresponding loss are so common,\n",
    "it is worth understanding a bit better how it is computed.\n",
    "Plugging :eqref:`eq_softmax_y_and_o` into the definition of the loss\n",
    "in :eqref:`eq_l_cross_entropy`\n",
    "and using the definition of the softmax we obtain:\n",
    "\n",
    "$$\n",
    "\\begin{aligned}\n",
    "l(\\mathbf{y}, \\hat{\\mathbf{y}}) &=  - \\sum_{j=1}^q y_j \\log \\frac{\\exp(o_j)}{\\sum_{k=1}^q \\exp(o_k)} \\\\\n",
    "&= \\sum_{j=1}^q y_j \\log \\sum_{k=1}^q \\exp(o_k) - \\sum_{j=1}^q y_j o_j\\\\\n",
    "&= \\log \\sum_{k=1}^q \\exp(o_k) - \\sum_{j=1}^q y_j o_j.\n",
    "\\end{aligned}\n",
    "$$\n",
    "\n",
    "To understand a bit better what is going on,\n",
    "consider the derivative with respect to any logit $o_j$. We get\n",
    "\n",
    "$$\n",
    "\\partial_{o_j} l(\\mathbf{y}, \\hat{\\mathbf{y}}) = \\frac{\\exp(o_j)}{\\sum_{k=1}^q \\exp(o_k)} - y_j = \\mathrm{softmax}(\\mathbf{o})_j - y_j.\n",
    "$$\n",
    "\n",
    "In other words, the derivative is the difference\n",
    "between the probability assigned by our model,\n",
    "as expressed by the softmax operation,\n",
    "and what actually happened, as expressed by elements in the one-hot label vector.\n",
    "In this sense, it is very similar to what we saw in regression,\n",
    "where the gradient was the difference\n",
    "between the observation $y$ and estimate $\\hat{y}$.\n",
    "This is not coincidence.\n",
    "In any exponential family (see the\n",
    "[online appendix on distributions](https://d2l.ai/chapter_appendix-mathematics-for-deep-learning/distributions.html)) model,\n",
    "the gradients of the log-likelihood are given by precisely this term.\n",
    "This fact makes computing gradients easy in practice.\n",
    "\n",
    "### Cross-Entropy Loss\n",
    "\n",
    "Now consider the case where we observe not just a single outcome\n",
    "but an entire distribution over outcomes.\n",
    "We can use the same representation as before for the label $\\mathbf{y}$.\n",
    "The only difference is that rather than a vector containing only binary entries,\n",
    "say $(0, 0, 1)$, we now have a generic probability vector, say $(0.1, 0.2, 0.7)$.\n",
    "The math that we used previously to define the loss $l$\n",
    "in :eqref:`eq_l_cross_entropy`\n",
    "still works out fine,\n",
    "just that the interpretation is slightly more general.\n",
    "It is the expected value of the loss for a distribution over labels.\n",
    "This loss is called the *cross-entropy loss* and it is\n",
    "one of the most commonly used losses for classification problems.\n",
    "We can demystify the name by introducing just the basics of information theory.\n",
    "If you wish to understand more details of information theory,\n",
    "you may further refer to the [online appendix on information theory](https://d2l.ai/chapter_appendix-mathematics-for-deep-learning/information-theory.html).\n",
    "\n",
    "\n",
    "\n",
    "## Information Theory Basics\n",
    ":label:`subsec_info_theory_basics`\n",
    "\n",
    "*Information theory* deals with the problem of encoding, decoding, transmitting,\n",
    "and manipulating information (also known as data) in as concise form as possible.\n",
    "\n",
    "\n",
    "### Entropy\n",
    "\n",
    "The central idea in information theory is to quantify the information content in data.\n",
    "This quantity places a hard limit on our ability to compress the data.\n",
    "In information theory, this quantity is called the *entropy* of a distribution $P$,\n",
    "and it is captured by the following equation:\n",
    "\n",
    "$$H[P] = \\sum_j - P(j) \\log P(j).$$\n",
    ":eqlabel:`eq_softmax_reg_entropy`\n",
    "\n",
    "One of the fundamental theorems of information theory states\n",
    "that in order to encode data drawn randomly from the distribution $P$,\n",
    "we need at least $H[P]$ \"nats\" to encode it.\n",
    "If you wonder what a \"nat\" is, it is the equivalent of bit\n",
    "but when using a code with base $e$ rather than one with base 2.\n",
    "Thus, one nat is $\\frac{1}{\\log(2)} \\approx 1.44$ bit.\n",
    "\n",
    "\n",
    "### Surprisal\n",
    "\n",
    "You might be wondering what compression has to do with prediction.\n",
    "Imagine that we have a stream of data that we want to compress.\n",
    "If it is always easy for us to predict the next token,\n",
    "then this data is easy to compress!\n",
    "Take the extreme example where every token in the stream always takes the same value.\n",
    "That is a very boring data stream!\n",
    "And not only it is boring, but it is also easy to predict.\n",
    "Because they are always the same, we do not have to transmit any information\n",
    "to communicate the contents of the stream.\n",
    "Easy to predict, easy to compress.\n",
    "\n",
    "However if we cannot perfectly predict every event,\n",
    "then we might sometimes be surprised.\n",
    "Our surprise is greater when we assigned an event lower probability.\n",
    "Claude Shannon settled on $\\log \\frac{1}{P(j)} = -\\log P(j)$\n",
    "to quantify one's *surprisal* at observing an event $j$\n",
    "having assigned it a (subjective) probability $P(j)$.\n",
    "The entropy defined in :eqref:`eq_softmax_reg_entropy` is then the *expected surprisal*\n",
    "when one assigned the correct probabilities\n",
    "that truly match the data-generating process.\n",
    "\n",
    "\n",
    "### Cross-Entropy Revisited\n",
    "\n",
    "So if entropy is level of surprise experienced\n",
    "by someone who knows the true probability,\n",
    "then you might be wondering, what is cross-entropy?\n",
    "The cross-entropy *from* $P$ *to* $Q$, denoted $H(P, Q)$,\n",
    "is the expected surprisal of an observer with subjective probabilities $Q$\n",
    "upon seeing data that were actually generated according to probabilities $P$.\n",
    "The lowest possible cross-entropy is achieved when $P=Q$.\n",
    "In this case, the cross-entropy from $P$ to $Q$ is $H(P, P)= H(P)$.\n",
    "\n",
    "In short, we can think of the cross-entropy classification objective\n",
    "in two ways: (i) as maximizing the likelihood of the observed data;\n",
    "and (ii) as minimizing our surprisal (and thus the number of bits)\n",
    "required to communicate the labels.\n",
    "\n",
    "\n",
    "## Model Prediction and Evaluation\n",
    "\n",
    "After training the softmax regression model, given any example features,\n",
    "we can predict the probability of each output class.\n",
    "Normally, we use the class with the highest predicted probability as the output class.\n",
    "The prediction is correct if it is consistent with the actual class (label).\n",
    "In the next part of the experiment,\n",
    "we will use *accuracy* to evaluate the model's performance.\n",
    "This is equal to the ratio between the number of correct predictions and the total number of predictions.\n",
    "\n",
    "\n",
    "## Summary\n",
    "\n",
    "* The softmax operation takes a vector and maps it into probabilities.\n",
    "* Softmax regression applies to classification problems. It uses the probability distribution of the output class in the softmax operation.\n",
    "* Cross-entropy is a good measure of the difference between two probability distributions. It measures the number of bits needed to encode the data given our model.\n",
    "\n",
    "## Exercises\n",
    "\n",
    "1. We can explore the connection between exponential families and the softmax in some more depth.\n",
    "    1. Compute the second derivative of the cross-entropy loss $l(\\mathbf{y},\\hat{\\mathbf{y}})$ for the softmax.\n",
    "    1. Compute the variance of the distribution given by $\\mathrm{softmax}(\\mathbf{o})$ and show that it matches the second derivative computed above.\n",
    "1. Assume that we have three classes which occur with equal probability, i.e., the probability vector is $(\\frac{1}{3}, \\frac{1}{3}, \\frac{1}{3})$.\n",
    "    1. What is the problem if we try to design a binary code for it?\n",
    "    1. Can you design a better code? Hint: what happens if we try to encode two independent observations? What if we encode $n$ observations jointly?\n",
    "1. Softmax is a misnomer for the mapping introduced above (but everyone in deep learning uses it). The real softmax is defined as $\\mathrm{RealSoftMax}(a, b) = \\log (\\exp(a) + \\exp(b))$.\n",
    "    1. Prove that $\\mathrm{RealSoftMax}(a, b) > \\mathrm{max}(a, b)$.\n",
    "    1. Prove that this holds for $\\lambda^{-1} \\mathrm{RealSoftMax}(\\lambda a, \\lambda b)$, provided that $\\lambda > 0$.\n",
    "    1. Show that for $\\lambda \\to \\infty$ we have $\\lambda^{-1} \\mathrm{RealSoftMax}(\\lambda a, \\lambda b) \\to \\mathrm{max}(a, b)$.\n",
    "    1. What does the soft-min look like?\n",
    "    1. Extend this to more than two numbers.\n",
    "\n",
    "[Discussions](https://discuss.d2l.ai/t/46)\n"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "conda_pytorch_p36",
   "name": "conda_pytorch_p36"
  },
  "language_info": {
   "name": "python"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}