$$ \newcommand{\bone}{\mathbf{1}} \newcommand{\bbeta}{\mathbf{\beta}} \newcommand{\bdelta}{\mathbf{\delta}} \newcommand{\bepsilon}{\mathbf{\epsilon}} \newcommand{\blambda}{\mathbf{\lambda}} \newcommand{\bomega}{\mathbf{\omega}} \newcommand{\bpi}{\mathbf{\pi}} \newcommand{\bphi}{\mathbf{\phi}} \newcommand{\bvphi}{\mathbf{\varphi}} \newcommand{\bpsi}{\mathbf{\psi}} \newcommand{\bsigma}{\mathbf{\sigma}} \newcommand{\btheta}{\mathbf{\theta}} \newcommand{\btau}{\mathbf{\tau}} \newcommand{\ba}{\mathbf{a}} \newcommand{\bb}{\mathbf{b}} \newcommand{\bc}{\mathbf{c}} \newcommand{\bd}{\mathbf{d}} \newcommand{\be}{\mathbf{e}} \newcommand{\boldf}{\mathbf{f}} \newcommand{\bg}{\mathbf{g}} \newcommand{\bh}{\mathbf{h}} \newcommand{\bi}{\mathbf{i}} \newcommand{\bj}{\mathbf{j}} \newcommand{\bk}{\mathbf{k}} \newcommand{\bell}{\mathbf{\ell}} \newcommand{\bm}{\mathbf{m}} \newcommand{\bn}{\mathbf{n}} \newcommand{\bo}{\mathbf{o}} \newcommand{\bp}{\mathbf{p}} \newcommand{\bq}{\mathbf{q}} \newcommand{\br}{\mathbf{r}} \newcommand{\bs}{\mathbf{s}} \newcommand{\bt}{\mathbf{t}} \newcommand{\bu}{\mathbf{u}} \newcommand{\bv}{\mathbf{v}} \newcommand{\bw}{\mathbf{w}} \newcommand{\bx}{\mathbf{x}} \newcommand{\by}{\mathbf{y}} \newcommand{\bz}{\mathbf{z}} \newcommand{\bA}{\mathbf{A}} \newcommand{\bB}{\mathbf{B}} \newcommand{\bC}{\mathbf{C}} \newcommand{\bD}{\mathbf{D}} \newcommand{\bE}{\mathbf{E}} \newcommand{\bF}{\mathbf{F}} \newcommand{\bG}{\mathbf{G}} \newcommand{\bH}{\mathbf{H}} \newcommand{\bI}{\mathbf{I}} \newcommand{\bJ}{\mathbf{J}} \newcommand{\bK}{\mathbf{K}} \newcommand{\bL}{\mathbf{L}} \newcommand{\bM}{\mathbf{M}} \newcommand{\bN}{\mathbf{N}} \newcommand{\bP}{\mathbf{P}} \newcommand{\bQ}{\mathbf{Q}} \newcommand{\bR}{\mathbf{R}} \newcommand{\bS}{\mathbf{S}} \newcommand{\bT}{\mathbf{T}} \newcommand{\bU}{\mathbf{U}} \newcommand{\bV}{\mathbf{V}} \newcommand{\bW}{\mathbf{W}} \newcommand{\bX}{\mathbf{X}} \newcommand{\bY}{\mathbf{Y}} \newcommand{\bZ}{\mathbf{Z}} \newcommand{\bsa}{\boldsymbol{a}} \newcommand{\bsb}{\boldsymbol{b}} \newcommand{\bsc}{\boldsymbol{c}} \newcommand{\bsd}{\boldsymbol{d}} \newcommand{\bse}{\boldsymbol{e}} \newcommand{\bsoldf}{\boldsymbol{f}} \newcommand{\bsg}{\boldsymbol{g}} \newcommand{\bsh}{\boldsymbol{h}} \newcommand{\bsi}{\boldsymbol{i}} \newcommand{\bsj}{\boldsymbol{j}} \newcommand{\bsk}{\boldsymbol{k}} \newcommand{\bsell}{\boldsymbol{\ell}} \newcommand{\bsm}{\boldsymbol{m}} \newcommand{\bsn}{\boldsymbol{n}} \newcommand{\bso}{\boldsymbol{o}} \newcommand{\bsp}{\boldsymbol{p}} \newcommand{\bsq}{\boldsymbol{q}} \newcommand{\bsr}{\boldsymbol{r}} \newcommand{\bss}{\boldsymbol{s}} \newcommand{\bst}{\boldsymbol{t}} \newcommand{\bsu}{\boldsymbol{u}} \newcommand{\bsv}{\boldsymbol{v}} \newcommand{\bsw}{\boldsymbol{w}} \newcommand{\bsx}{\boldsymbol{x}} \newcommand{\bsy}{\boldsymbol{y}} \newcommand{\bsz}{\boldsymbol{z}} \newcommand{\bsA}{\boldsymbol{A}} \newcommand{\bsB}{\boldsymbol{B}} \newcommand{\bsC}{\boldsymbol{C}} \newcommand{\bsD}{\boldsymbol{D}} \newcommand{\bsE}{\boldsymbol{E}} \newcommand{\bsF}{\boldsymbol{F}} \newcommand{\bsG}{\boldsymbol{G}} \newcommand{\bsH}{\boldsymbol{H}} \newcommand{\bsI}{\boldsymbol{I}} \newcommand{\bsJ}{\boldsymbol{J}} \newcommand{\bsK}{\boldsymbol{K}} \newcommand{\bsL}{\boldsymbol{L}} \newcommand{\bsM}{\boldsymbol{M}} \newcommand{\bsN}{\boldsymbol{N}} \newcommand{\bsP}{\boldsymbol{P}} \newcommand{\bsQ}{\boldsymbol{Q}} \newcommand{\bsR}{\boldsymbol{R}} \newcommand{\bsS}{\boldsymbol{S}} \newcommand{\bsT}{\boldsymbol{T}} \newcommand{\bsU}{\boldsymbol{U}} \newcommand{\bsV}{\boldsymbol{V}} \newcommand{\bsW}{\boldsymbol{W}} \newcommand{\bsX}{\boldsymbol{X}} \newcommand{\bsY}{\boldsymbol{Y}} \newcommand{\bsZ}{\boldsymbol{Z}} \newcommand{\calA}{\mathcal{A}} \newcommand{\calB}{\mathcal{B}} \newcommand{\calC}{\mathcal{C}} \newcommand{\calD}{\mathcal{D}} \newcommand{\calE}{\mathcal{E}} \newcommand{\calF}{\mathcal{F}} \newcommand{\calG}{\mathcal{G}} \newcommand{\calH}{\mathcal{H}} \newcommand{\calI}{\mathcal{I}} \newcommand{\calJ}{\mathcal{J}} \newcommand{\calK}{\mathcal{K}} \newcommand{\calL}{\mathcal{L}} \newcommand{\calM}{\mathcal{M}} \newcommand{\calN}{\mathcal{N}} \newcommand{\calO}{\mathcal{O}} \newcommand{\calP}{\mathcal{P}} \newcommand{\calQ}{\mathcal{Q}} \newcommand{\calR}{\mathcal{R}} \newcommand{\calS}{\mathcal{S}} \newcommand{\calT}{\mathcal{T}} \newcommand{\calU}{\mathcal{U}} \newcommand{\calV}{\mathcal{V}} \newcommand{\calW}{\mathcal{W}} \newcommand{\calX}{\mathcal{X}} \newcommand{\calY}{\mathcal{Y}} \newcommand{\calZ}{\mathcal{Z}} \newcommand{\R}{\mathbb{R}} \newcommand{\C}{\mathbb{C}} \newcommand{\N}{\mathbb{N}} \newcommand{\Z}{\mathbb{Z}} \newcommand{\F}{\mathbb{F}} \newcommand{\Q}{\mathbb{Q}} \DeclareMathOperator*{\argmax}{arg\,max} \DeclareMathOperator*{\argmin}{arg\,min} \newcommand{\nnz}[1]{\mbox{nnz}(#1)} \newcommand{\dotprod}[2]{\langle #1, #2 \rangle} \newcommand{\ignore}[1]{} \let\Pr\relax \DeclareMathOperator*{\Pr}{\mathbf{Pr}} \newcommand{\E}{\mathbb{E}} \DeclareMathOperator*{\Ex}{\mathbf{E}} \DeclareMathOperator*{\Var}{\mathbf{Var}} \DeclareMathOperator*{\Cov}{\mathbf{Cov}} \DeclareMathOperator*{\stddev}{\mathbf{stddev}} \DeclareMathOperator*{\avg}{avg} \DeclareMathOperator{\poly}{poly} \DeclareMathOperator{\polylog}{polylog} \DeclareMathOperator{\size}{size} \DeclareMathOperator{\sgn}{sgn} \DeclareMathOperator{\dist}{dist} \DeclareMathOperator{\vol}{vol} \DeclareMathOperator{\spn}{span} \DeclareMathOperator{\supp}{supp} \DeclareMathOperator{\tr}{tr} \DeclareMathOperator{\Tr}{Tr} \DeclareMathOperator{\codim}{codim} \DeclareMathOperator{\diag}{diag} \newcommand{\PTIME}{\mathsf{P}} \newcommand{\LOGSPACE}{\mathsf{L}} \newcommand{\ZPP}{\mathsf{ZPP}} \newcommand{\RP}{\mathsf{RP}} \newcommand{\BPP}{\mathsf{BPP}} \newcommand{\P}{\mathsf{P}} \newcommand{\NP}{\mathsf{NP}} \newcommand{\TC}{\mathsf{TC}} \newcommand{\AC}{\mathsf{AC}} \newcommand{\SC}{\mathsf{SC}} \newcommand{\SZK}{\mathsf{SZK}} \newcommand{\AM}{\mathsf{AM}} \newcommand{\IP}{\mathsf{IP}} \newcommand{\PSPACE}{\mathsf{PSPACE}} \newcommand{\EXP}{\mathsf{EXP}} \newcommand{\MIP}{\mathsf{MIP}} \newcommand{\NEXP}{\mathsf{NEXP}} \newcommand{\BQP}{\mathsf{BQP}} \newcommand{\distP}{\mathsf{dist\textbf{P}}} \newcommand{\distNP}{\mathsf{dist\textbf{NP}}} \newcommand{\eps}{\epsilon} \newcommand{\lam}{\lambda} \newcommand{\dleta}{\delta} \newcommand{\simga}{\sigma} \newcommand{\vphi}{\varphi} \newcommand{\la}{\langle} \newcommand{\ra}{\rangle} \newcommand{\wt}[1]{\widetilde{#1}} \newcommand{\wh}[1]{\widehat{#1}} \newcommand{\ol}[1]{\overline{#1}} \newcommand{\ul}[1]{\underline{#1}} \newcommand{\ot}{\otimes} \newcommand{\zo}{\{0,1\}} \newcommand{\co}{:} %\newcommand{\co}{\colon} \newcommand{\bdry}{\partial} \newcommand{\grad}{\nabla} \newcommand{\transp}{^\intercal} \newcommand{\inv}{^{-1}} \newcommand{\symmdiff}{\triangle} \newcommand{\symdiff}{\symmdiff} \newcommand{\half}{\tfrac{1}{2}} \newcommand{\mathbbm}{\Bbb} \newcommand{\bbone}{\mathbbm 1} \newcommand{\Id}{\bbone} \newcommand{\SAT}{\mathsf{SAT}} \newcommand{\bcalG}{\boldsymbol{\calG}} \newcommand{\calbG}{\bcalG} \newcommand{\bcalX}{\boldsymbol{\calX}} \newcommand{\calbX}{\bcalX} \newcommand{\bcalY}{\boldsymbol{\calY}} \newcommand{\calbY}{\bcalY} \newcommand{\bcalZ}{\boldsymbol{\calZ}} \newcommand{\calbZ}{\bcalZ} $$

2015

  1. Distilling the Knowledge in a Neural Network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean
    Mar 2015
    arXiv Bib PDF

    Paper Abstract

    A very simple way to improve the performance of almost any machine learning algorithm is to train many different models on the same data and then to average their predictions. Unfortunately, making predictions using a whole ensemble of models is cumbersome and may be too computationally expensive to allow deployment to a large number of users, especially if the individual models are large neural nets. Caruana and his collaborators have shown that it is possible to compress the knowledge in an ensemble into a single model which is much easier to deploy and we develop this approach further using a different compression technique. We achieve some surprising results on MNIST and we show that we can significantly improve the acoustic model of a heavily used commercial system by distilling the knowledge in an ensemble of models into a single model. We also introduce a new type of ensemble composed of one or more full models and many specialist models which learn to distinguish fine-grained classes that the full models confuse. Unlike a mixture of experts, these specialist models can be trained rapidly and in parallel.

Three Important Things

1. Motivation

The paper opens with a really interesting analogy: insects begin with a larval form optimized for extracting energy and nutrients from the environment, before metamorphosing into a completely adult form optimized for traveling and reproduction.

In a similar vein, for machine learning models we have access to plentiful compute and time during the training, but become constrained by different considerations like latency during inference. Given the different requirements and affordances of both phases, why should we expect to simply use the same model for both of them?

To address this, the paper introduces the idea of distillation, to transfer the knowledge of a very large/cumbersome model (possibly an ensemble of models) to a smaller one.

2. Distillation

There are several ways to approach distillation.

  1. A natural first approach would be to make the student model learn to predict the same class probabilities as the teacher model. These are known as “soft targets” (in contrast, “hard targets” are the ground-truth labels). Soft targets are “an effective way of communicating the regularities discovered by a model trained on all of the data to another model”.
  2. In practice, the correct label usually dominates the others by orders of magnitude. But a lot of information about the learned function can be gleaned from the ratios of the other labels, which can provide information on how one label might be confused for another and allow the model to draw more conclusions.

If we just use cross-entropy loss, these vital set of information will be lost as their relative values are tiny relative to the most confident prediction and don’t contribute much to the loss.

To address this, the authors added a temperature parameter before the softmax, and increased the temperature to a suitably high level until the targets are “soft” enough. The same temperature is used when training the small model. The temperature is reset to 1 afterwards at test time.

Note that changing temperature also requires the gradient update of the cross-entropy loss of the softmax predictions to be scaled appropriately by \((1/T^2)\) (via the chain rule).

The paper shows that the gradient of the cross-entropy loss is approximately \(\frac{1}{NT^2}(z_i - v_i)\) using a first-order Taylor approximation, which holds when temperature is high. By integrating this, the loss is proportionate to \(\frac{(z_i - v_i)^2}{2}\), showing that the loss function is essentially matching the logits.

3. Ensembles of Specialists

One way they applied the idea of matching soft targets with temperature is as follows.

Train time:

  1. Train a big generalist model the normal way
  2. Use the generalist model to identify clusters of labels by using the covariance matrix of its predictions
  3. For each of these clusters, train a specialist model. The specialist model only learns to classify the things in its cluster, and all other labels are lumped into a “dustbin” class

Test time:

  1. Run generalist model on test case sample. Take the top \(n\) most probable classes, call this set \(k\)
  2. Take all the specialist classes that can handle anything in \(k\). So each specialist class outputs its own logits. The goal is to optimize for the logits \(\mathbf{q}\) that minimizes over the KL between the logits of the generalist model and \(\mathbf{q}\) , as well as the logits between each of the specialist classes and \(\mathbf{q}\), i.e: \(K L\left(\mathbf{p}^g, \mathbf{q}\right)+\sum_{m \in A_k} K L\left(\mathbf{p}^m, \mathbf{q}\right)\)
  3. Take the most probable class in \(\mathbf{q}\) as the prediction

Most Glaring Deficiency

The paper did not (and acknowledged it did not) report on the results of actually distilling the outputs of the “cumbersome” ensemble of specialist models back into a small model, which felt like one of the main selling points that they were making.

Conclusions for Future Work

Distillation is now a pretty popular way to transfer the knowledge of larger/more cumbersome models into smaller ones, allowing for wins in latency and cost during live inference.

Written July 25, 2024