What is deep learning? Explained

What is a neural network

A neural network is software built from many small units called neurons. Each one takes in numbers, weighs them and passes a result along.

IBM describes deep learning as machine learning driven by multilayered neural networks, with a design inspired by the human brain.

The weights between neurons are what the system learns. Training adjusts them until the outputs match the examples it was shown.

Deep means many layers. Early layers pick up simple features, and later layers combine them into richer ones, such as a face, a word or an invoice total.

What deep learning powers

• Language

  Writing, summarizing and answering questions.

• Images

  Reading photos, scans and video.

• Speech

  Turning calls and voice into text.

• Documents

  Pulling fields from invoices and forms.

Deep learning works well on messy data that has no tidy columns. That is why it sits behind most image, speech and language tools you meet at work.

Language models such as chat assistants are deep networks trained on huge text collections. The large language model entry goes further.

Scale matters here. According to IBM, GPT-3 (2020) had 175 billion parameters, the connection weights a network learns during training.

The same family of methods reads scanned documents and photos. See computer vision for that side.

When you need deep learning

Most business problems do not need it. If your data fits in rows and columns, simpler machine learning is often faster, cheaper and easier to explain.

Deep learning earns its cost when the input is unstructured: free text, voice, images or long documents. It also needs lots of data and computing power.

For a side-by-side view, read machine learning vs deep learning.

One limit matters for managers: deep networks can be hard to explain. See black box AI for why that matters.