Artificial intelligence is the broad goal of building systems that perform tasks associated with human intelligence, and deep learning is one family of methods for reaching it. Deep learning uses neural networks with many layers that learn their own representations from data, as LeCun, Bengio, and Hinton described in Nature in 2015. It is not the best method for every problem: on tabular data, a 2022 benchmark of 45 datasets found tree-based models still ahead at about 10,000 samples.
Key takeaways
- Deep learning is a subset of machine learning, which is a subset of AI; some AI systems use no learning at all.
- Deep Blue beat Garry Kasparov in May 1997 by 3.5 to 2.5 using search and a hand-built evaluation function, not a deep neural network.
- Deep learning has advanced speech recognition, visual object recognition, and object detection, according to the 2015 Nature review.
- On medium-sized tabular datasets, tree-based models such as XGBoost and random forests stayed state of the art in a 2022 benchmark.
- A 2023 study of 176 datasets found the gap between neural nets and boosted trees is often negligible, and tuning can matter more than model family.
How the terms nest
Think of three circles, each inside the last. The outer circle is AI, any technique that lets a machine do work associated with intelligence. Inside it is machine learning, where behavior is fitted to example data instead of written out as rules. Inside that is deep learning, where a network of many layers builds up features step by step, for example edges, then shapes, then objects in an image. LeCun, Bengio, and Hinton describe these methods as learning representations at several levels of abstraction by adjusting each layer's parameters through backpropagation.
The nesting runs one way. Every deep learning system is AI, but many AI systems are not deep learning.
An AI system with no learning: Deep Blue
IBM's Deep Blue won a six-game rematch against world champion Garry Kasparov in May 1997 in New York, 3.5 to 2.5. It evaluated 200 million chess positions per second, and its developers had written a stronger evaluation function for the rematch. Move choice came from exploring possible continuations and scoring positions with that function, which people had designed. Nothing in this description is a deep neural network. The example shows why equating AI with deep learning leaves out decades of working systems built on search, logic, and hand-written rules.
Where deep learning clearly helped
The 2015 review credits deep learning with major advances in speech recognition, visual object recognition, object detection, drug discovery, and genomics. It names convolutional networks as the driver of progress for images, video, speech, and audio, and recurrent networks as effective for sequential data such as text. These tasks share a feature: the raw input is high-dimensional and unstructured, such as pixels or audio samples, and no one knows how to write the right features by hand.
Where it does not win: tabular data
Spreadsheet-style data, with columns of mixed types such as age, income, and category codes, behaves differently. Grinsztajn, Oyallon, and Varoquaux compared deep learning and tree-based models on a standard set of 45 datasets and reported in a July 2022 arXiv preprint that tree-based models remained state of the art on medium-sized data, roughly 10,000 samples, even without counting their speed advantage. They attributed this to differences in inductive bias and named three challenges for tabular neural networks: robustness to uninformative features, preserving data orientation, and learning irregular functions.
A later study by McElfresh and colleagues, published in the NeurIPS 2023 Datasets and Benchmarks track, compared 19 algorithms on 176 datasets. It found that for many datasets the gap between neural nets and gradient-boosted trees is negligible, or that light tuning of a boosted tree matters more than the choice of model family. It also found boosted trees handled skewed or heavy-tailed features better.
Choosing between approaches
The table gives a rough, illustrative comparison. It summarizes the sections above and is a rule of thumb, not a result from a single study.
| Approach | Learns from data? | Typical fit | Cited example |
|---|---|---|---|
| Search with hand-built scoring | No | Well-defined games and planning | Deep Blue, 1997 |
| Boosted or random-forest trees | Yes | Tabular data, mid-sized datasets | 45-dataset benchmark, 2022 |
| Deep neural networks | Yes | Images, audio, text, raw signals | Nature review, 2015 |
FAQ
Is deep learning the same as AI?
No. Deep learning is one method inside machine learning, which sits inside AI. Chess programs built on search and hand-written scoring are AI without any deep learning.
Why is it called deep?
The word refers to the number of layers of learned processing between input and output. More layers allow the system to build higher-level features out of simpler ones.
Does deep learning always need more data than other methods?
Not as a strict rule, but the tabular benchmarks above show it has no automatic edge at around 10,000 samples. Results depend on the data type, tuning effort, and the features available.
Which should a small organization use?
The evidence does not name one answer. Starting with a simple baseline, such as a tree-based model for tabular data, and comparing it with a neural network on held-out data is a sound way to decide.
Related reading
- Deep learning applications and progress
- Artificial versus automated intelligence
- AI compared with computer science
- Deep learning models with over a trillion parameters
More in AI Basics.
Last verified: October 2026
Sources
- Deep learning, LeCun Y, Bengio Y, Hinton G, Nature 521, 2015
- Why do tree-based models still outperform deep learning on tabular data?, Grinsztajn L, Oyallon E, Varoquaux G, arXiv:2207.08815, 2022
- When Do Neural Nets Outperform Boosted Trees on Tabular Data?, McElfresh D et al., arXiv:2305.02997, NeurIPS 2023 Datasets and Benchmarks Track
- Deep Blue, IBM History, International Business Machines Corporation