# Why Deep Learning Matters

- Canonical URL: https://amund.blog/archive-why-deep-learning-matters/
- Author: Amund Tveit (https://amund.blog/about/)
- Published: 2016-04-29
- Tags: Deep Learning, AI Systems, Computer Vision
- Original Source: amundtveit.com (https://amundtveit.com/2016/04/29/why-deep-learning-matters/)

---

[Deep Learning](https://en.wikipedia.org/wiki/Deep_learning) — and more specifically a subgroup called [(Deep) Convolutional Neural Networks](https://en.wikipedia.org/wiki/Convolutional_neural_network) — has seen dramatic improvements since Alex Krizhevsky's 2012 publication on what is now called AlexNet. AlexNet won the ImageNet Image Recognition competition with a then jaw-dropping top-5 error rate of **17.0%** *(top-5 error means that if your classifier presents 5 answers, at least one must be correct)*.

Image recognition accuracy has improved many times since then, dropping from **17.00% in 2012** to **3.08% in 2016** *(see publications in the table below)*.

To put this into context: **human beings perform at a 5.10% error rate on this ImageNet recognition task** (see Andrej Karpathy's 2014 experiment below — to be precise, at least one smart, trained, and highly educated human performed at that error rate).

**Computers with Deep Learning can actually see and understand what is in a picture better than humans** (in many, and probably most, classification cases).

| Year | Top-5 Error % | Reference | Author(s) | Organization |
| --- | ---: | --- | --- | --- |
| 2012 | 17.00 | [ImageNet Classification with Deep Convolutional Neural Networks](http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf) (AlexNet) | Alex Krizhevsky et al. | University of Toronto |
| 2014 | 6.66 | [Going Deeper with Convolutions](https://arxiv.org/pdf/1409.4842.pdf) | Christian Szegedy et al. | Google |
| 2014 (Sep) | 5.10 | [What I learned from competing against a ConvNet on ImageNet](http://karpathy.github.io/2014/09/02/what-i-learned-from-competing-against-a-convnet-on-imagenet) | Andrej Karpathy | Stanford University |
| 2015 (Feb) | 4.94 | [Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification](https://arxiv.org/pdf/1502.01852v1.pdf) | Kaiming He et al. | Microsoft Research |
| 2015 (Dec) | 3.57 | [Deep Residual Learning for Image Recognition](https://arxiv.org/pdf/1512.03385v1.pdf) | Kaiming He et al. | Microsoft Research |
| 2016 (Feb) | 3.08 | [Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning](https://arxiv.org/abs/1602.07261) | Christian Szegedy et al. | Google |

## Implications of Superhuman Image Recognition

Deep Convolutional Neural Network research has moved so fast that real-world applications still lag behind. Most robots, drones, and software running on servers, laptops, mobiles, wearables, and medical equipment do not yet take advantage of these results, leaving huge untapped potential.

There are already important clinical and industrial deployments, however — such as the [Samsung Medison RS80A Ultrasound Machine](http://www.samsungmedison.com/public/news_view.jsp?docNo=919), which uses convolutional neural networks for [breast cancer](https://en.wikipedia.org/wiki/Breast_cancer) diagnosis.

![Typical Convolutional Neural Network architecture](./typical-cnn.png)

*Amund Tveit*
