US Deep learning defined
Deep learning is a form of machine learning in which artificial neural networks with many stacked layers learn to recognize patterns from examples. The word “deep” refers to that depth: a modern network can have dozens or hundreds of layers, each turning the output of the one before into a more abstract representation.
Show a deep network enough labeled photos and it learns, without being told, that edges form shapes, shapes form parts and parts form a bolt, a face or a number plate. The same principle works on sound, text, radar returns and sensor signals.
The ideas are decades old. Backpropagation, the method used to train networks, was popularised in 1986, but deep learning only took off around 2012, when a GPU-trained network called AlexNet won the ImageNet image recognition challenge by a wide margin. Geoffrey Hinton, Yann LeCun and Yoshua Bengio received the 2018 ACM Turing Award for the work behind it.
How deep learning works
A neural network is built from simple artificial neurons, loosely modeled on brain cells and arranged in layers. Each one multiplies its inputs by learned weights, adds them up and passes the result through an activation function. Learning means finding weights that give the right answers, which happens in a loop:
-
//
Forward pass. A batch of training examples flows through the layers and the network makes its predictions.
-
//
Loss. A loss function scores how far those predictions are from the correct labels.
-
//
Backpropagation. The error is traced backwards through the network to work out how much each weight contributed to it.
-
//
Gradient descent. Every weight is nudged slightly in the direction that reduces the error, then the loop repeats, often millions of times.
Training and inference
This loop is training, and for large models it runs for days or weeks on many GPUs. Once trained, the weights are frozen and the model is used for inference: a single forward pass on new data, such as one camera frame, takes milliseconds and far less power. Most organizations train in a data centre or the cloud and run inference on site.
Few industrial teams train a network from scratch. Transfer learning starts from a model already trained on a large public dataset, such as ImageNet or COCO, and fine-tunes it on a few hundred or a few thousand images of your own parts or scenes. That cuts both the labeled data and the GPU time needed, often from weeks to hours.
Types of deep learning model
Different network designs, or architectures, suit different kinds of data. These are the main families in use today.
| Architecture | Typical input | Used for | Edge example |
|---|---|---|---|
| Convolutional neural network (CNN) | Images, video frames | Classification, detection, segmentation | Defect detection on a production line |
| Recurrent network (RNN, LSTM) | Time series, sequences | Forecasting, signal analysis | Predicting a fault from sensor trends |
| Transformer | Text, images, audio | Language models, vision transformers, translation | Voice commands on a robot |
| Autoencoder | Any signal or image | Anomaly detection, compression | Flagging unusual vibration on a pump |
| Diffusion model and GAN | Noise plus a prompt | Generating images and synthetic data | Creating extra training images of rare defects |
Many current systems combine them, for example a CNN backbone for detection feeding an object tracker, or a vision transformer paired with a language model.
Why deep learning runs on GPUs
Almost all the arithmetic in a neural network is matrix multiplication: millions of independent multiply-and-add operations that can run at the same time. A GPU has thousands of cores built for exactly that kind of parallel work, and recent NVIDIA GPUs add Tensor Cores that speed up matrix maths at reduced precision. Frameworks such as PyTorch and TensorFlow call NVIDIA libraries like cuDNN so developers rarely write GPU code directly.
Training favours the largest data centre GPUs with plenty of memory. Inference can use much smaller devices, especially once the model is quantized to 8-bit integers, which is why the same network trained on a server cluster can run on a small embedded module drawing a few tens of watts.
Deep learning at the edge
Running a trained model on site means exporting it from the training framework, usually to the ONNX format, then compiling it for the target hardware: TensorRT for NVIDIA GPUs and Jetson, the Dataflow Compiler for Hailo, or OpenVINO for Intel processors.
Our Jetson Orin computers are the most common choice for multi-camera vision models in vehicles, robots and outdoor cabinets, while Hailo AI accelerators add deep learning inference to x86 and Arm systems at a few watts. Our guide to deploying AI at the edge with NVIDIA Orin covers the full pipeline. Typical edge uses include visual inspection, number plate recognition, people and vehicle counting, robot navigation and acoustic or vibration monitoring.
Quick answers on deep learning
Deep learning: common questions
What is deep learning used for?
Anything that learns from images, sound or text. Machine vision, medical imaging, speech recognition, translation, chatbots, recommendation systems, robotics and autonomous vehicles all rely on deep learning.
Is deep learning AI?
Yes. Deep learning is a branch of machine learning, which is itself the main approach within artificial intelligence.
What is a deep neural network?
A neural network with many hidden layers. Anything beyond a few layers between input and output is usually called deep. Deep neural networks are the models that deep learning trains.
Is generative AI deep learning?
Yes. Generative AI, including large language models and image generators, is built on deep neural networks such as transformers and diffusion models.
Which GPU do you need for deep learning?
It depends on whether you train or run models. Training large models needs data centre GPUs with large memory. Inference often runs on a single workstation GPU, a Jetson module or an NPU.
We build Jetson, GPU and Hailo-based computers for deep learning inference with the operating system, drivers and AI runtime installed as standard, and validate each system with your model before it ships. Send us your model, camera count and frame rates, and we will size the hardware to suit.