What is the typical deep learning architecture

What is deep learning?

robotics

Many of the recent developments in robotics have been driven by advances in AI and deep learning. For example, AI enables robots to perceive their surroundings and to react to them. This extends their area of ‚Äč‚Äčapplication - the robots can, for example, navigate through warehouses and handle objects that are irregularly shaped, fragile or even disordered. An act like picking a strawberry is easy for humans, but robots have always had remarkable difficulties with it. Advances in artificial intelligence will continue to expand the capabilities of robots.

In view of the developments in AI, it is to be expected that robots will increasingly function as human assistants in the future. You will no longer just understand and answer questions - as in some cases today - but will be able to react to voice commands or gestures and even foresee the next work steps. Collaborative robots are already working directly with people; both sides carry out the tasks that best suit their respective strengths.

Agriculture

Artificial intelligence has the potential to revolutionize agricultural cultivation. Deep learning currently enables farmers to use devices that distinguish between cultivated plants and weeds through optical detection. Herbicides can be selectively applied to foreign plants while the crops remain untouched. Agricultural machines even optimize the cultivation of certain individual plants in the field using computer vision based on deep learning by spraying herbicides, fungicides, insecticides, fertilizers and biological substances only at the right places. The use of pesticides is reduced and the yield increased - but deep learning can also be applied to other areas of agriculture, such as irrigation and harvesting.

Medical image processing and healthcare in general

Deep learning can be used particularly effectively in medical image processing because high-quality data is available and convolutional neural networks can classify images. The classification of skin cancers is achieved by deep learning systems at least as effectively as dermatologists. Several providers have already received FDA approval for deep learning algorithms for diagnostic purposes, for example for image evaluations in connection with cancer and retinal diseases. Deep learning is also well on its way to improving the quality of health care by forecasting medical events from electronic health records.

Deep learning of the future

Various neural network architectures are currently optimized for certain types of input data and tasks. Convolutional neural networks are very useful for classifying images. Another form of deep learning architecture uses recurrent neural networks to process sequential data. In both convolutional and recurrent neural networks, monitored learning takes place, for which large amounts of data are required. In the future, more sophisticated types of AI will use unsupervised learning. There is a lot of research going on into technologies for unsupervised and partially supervised learning.

Reinforcement learning is a slightly different approach to deep learning where an agent learns through trial and error in a simulated environment based solely on rewards and punishments. Deep learning extensions in this area are known as Deep Reinforcement Learning (DRL). Great strides have been made in this regard, as demonstrated by DRL programs that defeat people in the very old game of GO.

Designing neural network architectures to solve problems is incredibly difficult. Numerous hyperparameters to be adjusted and many loss functions that can be selected for optimization make the matter even more complex. Extensive research activities target neural network architectures for autonomous learning. There is steady progress in learning what is known as metal learning or AutoML.

The current artificial neural networks are based on the knowledge that prevailed in the 1950s about information processing in the human brain. Since then, neuroscientists have made considerable new discoveries, and deep learning architectures have become so complex that they appear to have grid cell structures such as those found in biological neural brains for navigation. Neuroscience and deep learning can positively influence each other through the exchange of ideas - both areas will very likely merge at some point.

Mechanical computers are no longer used today, and similarly there will come a time when digital computers will be a relic of the past. A new generation of quantum computers will take their place. Quantum computing has made some leaps in development in recent years, and learning algorithms can certainly benefit from the gigantic computing capacity of quantum computers. It is also possible to use learning algorithms to evaluate the output data of the probabilistic quantum computer. Quantum machine learning is a very vibrant sub-category of machine learning, and 2018 is slated for the first international quantum machine learning conference - a good start!