Stochastic Gradient Descent: Difference between revisions
(Marked this version for translation) Tag: Reverted |
([deploy-bot] Deploy from CI (5358a41)) Tags: ci-deploy Manual revert Reverted |
||
| Line 15: | Line 15: | ||
<translate> | <translate> | ||
<!--T:1--> | <!--T:1--> | ||
'''Stochastic | '''Stochastic gradient descent''' (often abbreviated '''{{Term|SGD|SGD}}''') is an iterative optimisation algorithm used to minimise an {{Term|objective function|objective function}} written as a sum of differentiable sub-functions. It is the workhorse behind modern machine-learning training, powering everything from logistic regression to deep neural networks. | ||
<!--T:2--> | |||
== Motivation == | |||
<!--T:3--> | <!--T:3--> | ||
In classical {{Term|gradient descent}}, the full gradient of the {{Term|loss function}} is computed over the entire training set before each parameter update. When the dataset is large this becomes prohibitively expensive. SGD addresses the problem by estimating the gradient from a single randomly chosen sample (or a small '''{{Term|mini-batch}}''') at each step, trading a noisier estimate for dramatically lower per-iteration cost. | In classical {{Term|gradient descent}}, the full gradient of the {{Term|loss function}} is computed over the entire training set before each parameter update. When the dataset is large this becomes prohibitively expensive. SGD addresses the problem by estimating the gradient from a single randomly chosen sample (or a small '''{{Term|mini-batch}}''') at each step, trading a noisier estimate for dramatically lower per-iteration cost. | ||
<!--T:4--> | |||
== Algorithm == | |||
<!--T:5--> | <!--T:5--> | ||
| Line 37: | Line 39: | ||
<!--T:9--> | <!--T:9--> | ||
where <math>\eta_t</math> is the '''{{Term|learning rate}}''' ( | where <math>\eta_t</math> is the '''{{Term|learning rate}}''' (step size) and <math>i_t</math> is a randomly selected index. | ||
=== Mini-batch variant === | <!--T:10--> | ||
=== Mini-batch variant === | |||
<!--T:11--> | <!--T:11--> | ||
| Line 50: | Line 53: | ||
Common batch sizes range from 32 to 512. Larger batches reduce gradient variance but increase memory usage. | Common batch sizes range from 32 to 512. Larger batches reduce gradient variance but increase memory usage. | ||
<!--T:14--> | |||
=== Pseudocode === | |||
<!--T:15--> | <!--T:15--> | ||
| Line 64: | Line 68: | ||
</pre> | </pre> | ||
<!--T:16--> | |||
== Learning rate schedules == | |||
<!--T:17--> | <!--T:17--> | ||
| Line 71: | Line 76: | ||
<!--T:18--> | <!--T:18--> | ||
* '''Constant''' — simple but may overshoot or stall. | * '''Constant''' — simple but may overshoot or stall. | ||
* '''Step decay''' — multiply <math>\eta</math> by a factor (e.g. 0.1) every <math>k</math> | * '''Step decay''' — multiply <math>\eta</math> by a factor (e.g. 0.1) every <math>k</math> epochs. | ||
* '''Exponential decay''' — <math>\eta_t = \eta_0 \, e^{-\lambda t}</math>. | * '''Exponential decay''' — <math>\eta_t = \eta_0 \, e^{-\lambda t}</math>. | ||
* '''Cosine annealing''' — smoothly reduces the rate following a cosine curve, often with warm restarts. | * '''Cosine annealing''' — smoothly reduces the rate following a cosine curve, often with warm restarts. | ||
* '''Linear warm-up''' — ramp up from a small <math>\eta</math> during the first few iterations to stabilise early training. | * '''Linear warm-up''' — ramp up from a small <math>\eta</math> during the first few iterations to stabilise early training. | ||
<!--T:19--> | |||
== Convergence properties == | |||
<!--T:20--> | <!--T:20--> | ||
| Line 85: | Line 91: | ||
<!--T:22--> | <!--T:22--> | ||
converges almost surely to the global minimum (Robbins–Monro conditions). For non-convex problems — the typical regime for | converges almost surely to the global minimum (Robbins–Monro conditions). For non-convex problems — the typical regime for deep learning — SGD converges to a stationary point, and empirical evidence shows it often finds good local minima. | ||
<!--T:23--> | |||
== Popular variants == | |||
<!--T:24--> | <!--T:24--> | ||
Several extensions reduce the variance of the gradient estimate or adapt the | Several extensions reduce the variance of the gradient estimate or adapt the step size per parameter: | ||
<!--T:25--> | <!--T:25--> | ||
| Line 101: | Line 108: | ||
| '''Nesterov accelerated gradient''' || Evaluates the gradient at a "look-ahead" position || Nesterov, 1983 | | '''Nesterov accelerated gradient''' || Evaluates the gradient at a "look-ahead" position || Nesterov, 1983 | ||
|- | |- | ||
| ''' | | '''Adagrad''' || Per-parameter rates that shrink for frequently updated features || Duchi et al., 2011 | ||
|- | |- | ||
| '''RMSProp''' || Fixes | | '''RMSProp''' || Fixes Adagrad's diminishing rates using a moving average of squared gradients || Hinton (lecture notes), 2012 | ||
|- | |- | ||
| '''{{Term|Adam}}''' || Combines {{Term|momentum}} with RMSProp-style adaptive rates || Kingma & Ba, 2015 | | '''{{Term|Adam}}''' || Combines {{Term|momentum}} with RMSProp-style adaptive rates || Kingma & Ba, 2015 | ||
|- | |- | ||
| '''AdamW''' || Decouples | | '''AdamW''' || Decouples weight decay from the adaptive gradient step || Loshchilov & Hutter, 2019 | ||
|} | |} | ||
<!--T:26--> | |||
== Practical considerations == | |||
<!--T:27--> | <!--T:27--> | ||
* '''Data shuffling''' — Re-shuffle the dataset each | * '''Data shuffling''' — Re-shuffle the dataset each epoch to avoid cyclic patterns. | ||
* '''{{Term|gradient clipping|Gradient clipping}}''' — Cap the gradient norm to prevent exploding updates, especially in recurrent networks. | * '''{{Term|gradient clipping|Gradient clipping}}''' — Cap the gradient norm to prevent exploding updates, especially in recurrent networks. | ||
* '''{{Term|batch normalization|Batch normalisation}}''' — Normalising layer inputs reduces sensitivity to the {{Term|learning rate}}. | * '''{{Term|batch normalization|Batch normalisation}}''' — Normalising layer inputs reduces sensitivity to the {{Term|learning rate}}. | ||
* '''Mixed-precision training''' — Using half-precision floats accelerates SGD on modern GPUs with minimal accuracy loss. | * '''Mixed-precision training''' — Using half-precision floats accelerates SGD on modern GPUs with minimal accuracy loss. | ||
<!--T:28--> | |||
== Applications == | |||
<!--T:29--> | <!--T:29--> | ||
| Line 125: | Line 134: | ||
<!--T:30--> | <!--T:30--> | ||
* Training deep neural networks (computer vision, NLP, speech recognition) | * Training deep neural networks (computer vision, NLP, speech recognition) | ||
* Large-scale linear models ( | * Large-scale linear models (logistic regression, SVMs via SGD) | ||
* Reinforcement learning policy optimisation | * Reinforcement learning policy optimisation | ||
* | * Recommendation systems and collaborative filtering | ||
* Online learning settings where data arrives in a stream | * Online learning settings where data arrives in a stream | ||
<!--T:31--> | |||
== See also == | |||
<!--T:32--> | <!--T:32--> | ||
| Line 139: | Line 149: | ||
* [[Convex optimisation]] | * [[Convex optimisation]] | ||
<!--T:33--> | |||
== References == | |||
<!--T:34--> | <!--T:34--> | ||
* Robbins, H. and Monro, S. (1951). "A Stochastic Approximation Method". ''Annals of Mathematical Statistics''. | * Robbins, H. and Monro, S. (1951). "A Stochastic Approximation Method". ''Annals of Mathematical Statistics''. | ||
* Bottou, L. (2010). "Large-Scale Machine Learning with Stochastic Gradient Descent". ''COMPSTAT''. | * Bottou, L. (2010). "Large-Scale Machine Learning with Stochastic Gradient Descent". ''COMPSTAT''. | ||
* Kingma, D. P. and Ba, J. (2015). " | * Kingma, D. P. and Ba, J. (2015). "Adam: A Method for Stochastic Optimization". ''ICLR''. | ||
* Ruder, S. (2016). "An overview of | * Ruder, S. (2016). "An overview of gradient descent optimization algorithms". ''arXiv:1609.04747''. | ||
<!--T:35--> | <!--T:35--> | ||
Revision as of 19:43, 27 April 2026
- {{#arraymap:Estimates gradients from random mini-batches instead of the full dataset; Learning rate schedule is critical for convergence; Variants like Adam and AdamW add adaptive per-parameter rates; Converges to global minimum for convex problems under Robbins–Monro conditions|;|@@item@@|
- @@item@@ |}}
Stochastic gradient descent (often abbreviated SGD) is an iterative optimisation algorithm used to minimise an objective function written as a sum of differentiable sub-functions. It is the workhorse behind modern machine-learning training, powering everything from logistic regression to deep neural networks.
Motivation
In classical gradient descent, the full gradient of the loss function is computed over the entire training set before each parameter update. When the dataset is large this becomes prohibitively expensive. SGD addresses the problem by estimating the gradient from a single randomly chosen sample (or a small mini-batch) at each step, trading a noisier estimate for dramatically lower per-iteration cost.
Algorithm
Given a parameterised loss function
- $ L(\theta) = \frac{1}{N}\sum_{i=1}^{N} \ell(\theta;\, x_i,\, y_i) $
the SGD update rule at step $ t $ is:
- $ \theta_{t+1} = \theta_t - \eta_t \,\nabla_\theta \ell(\theta_t;\, x_{i_t},\, y_{i_t}) $
where $ \eta_t $ is the learning rate (step size) and $ i_t $ is a randomly selected index.
Mini-batch variant
In practice a mini-batch of $ B $ samples is used:
- $ \theta_{t+1} = \theta_t - \frac{\eta_t}{B}\sum_{j=1}^{B} \nabla_\theta \ell(\theta_t;\, x_{i_j},\, y_{i_j}) $
Common batch sizes range from 32 to 512. Larger batches reduce gradient variance but increase memory usage.
Pseudocode
initialise parameters θ
for epoch = 1, 2, … do
shuffle training set
for each mini-batch B ⊂ training set do
g ← (1/|B|) Σ ∇ℓ(θ; xᵢ, yᵢ) # estimate gradient
θ ← θ − η · g # update parameters
end for
end for
Learning rate schedules
The learning rate $ \eta_t $ strongly influences convergence. Common strategies include:
- Constant — simple but may overshoot or stall.
- Step decay — multiply $ \eta $ by a factor (e.g. 0.1) every $ k $ epochs.
- Exponential decay — $ \eta_t = \eta_0 \, e^{-\lambda t} $.
- Cosine annealing — smoothly reduces the rate following a cosine curve, often with warm restarts.
- Linear warm-up — ramp up from a small $ \eta $ during the first few iterations to stabilise early training.
Convergence properties
For convex objectives with Lipschitz-continuous gradients, SGD with a decaying learning rate satisfying
- $ \sum_{t=1}^{\infty} \eta_t = \infty, \qquad \sum_{t=1}^{\infty} \eta_t^2 < \infty $
converges almost surely to the global minimum (Robbins–Monro conditions). For non-convex problems — the typical regime for deep learning — SGD converges to a stationary point, and empirical evidence shows it often finds good local minima.
Popular variants
Several extensions reduce the variance of the gradient estimate or adapt the step size per parameter:
| Method | Key idea | Reference |
|---|---|---|
| Momentum | Accumulates an exponentially decaying moving average of past gradients | Polyak, 1964 |
| Nesterov accelerated gradient | Evaluates the gradient at a "look-ahead" position | Nesterov, 1983 |
| Adagrad | Per-parameter rates that shrink for frequently updated features | Duchi et al., 2011 |
| RMSProp | Fixes Adagrad's diminishing rates using a moving average of squared gradients | Hinton (lecture notes), 2012 |
| Adam | Combines momentum with RMSProp-style adaptive rates | Kingma & Ba, 2015 |
| AdamW | Decouples weight decay from the adaptive gradient step | Loshchilov & Hutter, 2019 |
Practical considerations
- Data shuffling — Re-shuffle the dataset each epoch to avoid cyclic patterns.
- Gradient clipping — Cap the gradient norm to prevent exploding updates, especially in recurrent networks.
- Batch normalisation — Normalising layer inputs reduces sensitivity to the learning rate.
- Mixed-precision training — Using half-precision floats accelerates SGD on modern GPUs with minimal accuracy loss.
Applications
SGD and its variants are used across virtually all areas of machine learning:
- Training deep neural networks (computer vision, NLP, speech recognition)
- Large-scale linear models (logistic regression, SVMs via SGD)
- Reinforcement learning policy optimisation
- Recommendation systems and collaborative filtering
- Online learning settings where data arrives in a stream
See also
References
- Robbins, H. and Monro, S. (1951). "A Stochastic Approximation Method". Annals of Mathematical Statistics.
- Bottou, L. (2010). "Large-Scale Machine Learning with Stochastic Gradient Descent". COMPSTAT.
- Kingma, D. P. and Ba, J. (2015). "Adam: A Method for Stochastic Optimization". ICLR.
- Ruder, S. (2016). "An overview of gradient descent optimization algorithms". arXiv:1609.04747.