Deep learning is a class of machine-learning methods in which data passes through a hierarchy of many layers of a neural network and, at each layer, is transformed into an increasingly abstract representation. The word "deep" refers precisely to the large number of such layers — from a few to thousands.
In the biology of aging, deep learning is used to build epigenetic clocks and other models of biological age from large data sets.
What it is
Deep learning is a subfield of machine learning that relies on artificial neural networks with many hidden layers. Its key feature is representation learning: the network itself finds useful features in the data, whereas in classical methods the features are specified by a human.
The practical breakthrough is associated with 2012 (the victory of the AlexNet network in the ImageNet image-recognition competition), and in 2018 Yoshua Bengio, Geoffrey Hinton and Yann LeCun received the Turing Award for developing these ideas. Today deep learning underpins image and speech recognition, language processing and the analysis of medical images.
How it is applied in aging science
Aging "clocks" are models that predict age and the rate of aging from biological data. Deep networks are trained to predict age from DNA methylation, from routine biochemistry and the cellular composition of blood, and also from MRI scans (estimating "brain age") and from the picture of systemic inflammation ("inflammatory age").
The difference between predicted biological age and chronological age — so-called age acceleration — is regarded as a possible biomarker of aging and is used to assess the effect of interventions.
What is proven and what is being studied
The ability of deep models to accurately predict chronological age from data has been shown many times. However, how well "age acceleration" from such a model predicts individual diseases and lifespan is a question that is still being studied and requires independent validation in large cohorts.
It is important to understand the limitations: deep networks often work as a "black box," are sensitive to data quality and are not a diagnosis. This is a research tool, not a substitute for a clinical examination.
Calculate using this marker
Calculators where Deep learning is used directly:
Frequently asked questions
How does deep learning differ from ordinary machine learning?
It is its subfield, based on multilayer neural networks: the network itself extracts features from the data rather than receiving them ready-made from a human. "Depth" is the number of layers.
Can a neural network tell me my "true" age?
The model gives an estimate of biological age from data, not the exact truth. It is a reference point for tracking changes over time, not a diagnosis.