The potential for Artificial Intelligence (AI) is enormous and the applications seemingly unlimited. One subset of AI, deep learning, offers the promise of efficiently solving a large range of challenges involving unstructured data by harnessing neural networks to save time and money, and even perform seemingly impossible tasks.
Deep learning has revolutionized the fields of artificial intelligence, computer vision, speech recognition, and more in recent years. Deep learning can draw information from unstructured data such as images or text in a way that was unthinkable a decade ago. In industries like Pharma and Biopharma, deep learning can help all the way from understanding how cells work using live cell imaging to monitoring manufacturing using audio.
The unlimited applications of deep learning
An article in Forbes in 2017 stated that ‘More progress has been achieved on artificial intelligence in the past five years than in the past five decades’, as illustrated by the rise of spectacular applications such as autonomous cars.
This has been made possible by deep learning, a subset of AI, that has a wide range of other applications such as computer vision, speech recognition, natural language processing, social network filtering, machine translation, bioinformatics, drug design, and medical image analysis.
Deep learning has helped to improve the control of solar- and wind power systems with predictions on output, and helped to optimize the control of cooling data centers, achieving 30% reductions in costs.
What exactly is deep learning?
Deep learning is a machine learning technique that enables computers to learn by example. The ‘deep’ refers to the number of layers through which the data is transformed, with multiple layers used to extract features at progressively higher levels from raw input. A typical example is image processing, where lower levels could identify edges, while higher levels, or layers, identify more complex items such as letters, or biological cells viewed with a microscope. Importantly, a deep learning process can learn which features to optimally place in which level on its own.
Four key factors have made deep learning possible in the last few years.
- There is now sufficient computational power to handle the enormous amount of data manipulations needed.
- The era of Big Data means that there is sufficient data to work from.
- New algorithms and architectures have been developed to handle the data in an effective way.
- Open source and open access publishing has meant a free exchange of ideas to accelerate development.
How does deep learning work?
Deep learning is based on artificial neural networks, which were inspired by information processing and distributed communication nodes in biological systems. Deep learning involves the development of stacked non-linear models. We start with basic linear regression, where a model of responses is created from the sum of weights and input variables, which can put into an input data matrix and a weight matrix. The second component is an activation function, or non-linear function.
A commonly used method is the rectifier linear unit – that transforms all values less than zero to zero to give a non-linear response. This process is performed repeatedly, feeding hidden layer activations into the next level to develop a neural network. Learning the model weights is critical, and this is typically done using stochastic gradient descent and a back-propagation algorithm that modify earlier layers in the neural network based on the errors detected in the model predictions.
Deep learning involves building successive layers that describe the data through a series of transforms to form a neural network that predicts the response to the original variables.
An efficient process
Deep learning enables feature learning in an optimal way by connecting the raw data all the way to the response and optimizing the transforms all the way through. This process is far more efficient than laborious feature engineering, which is also almost impossible to apply to the analysis of complex unstructured data such as images where deep learning has been so successful.
Classical fully connected neural networks are rarely used since they are inefficient and prone to overfitting. Instead other network architectures are used that take advantage of the type of data being analyzed. In image analysis, for instance, Convolutional Neural Networks (CNNs) are commonly used, which were inspired by the visual cortex and have enabled major breakthroughs in computer vision.
CNNs use three basic operations that are repeated in blocks, namely:
- Convolutions that create feature maps that show different levels of variations between data points.
- Activations that introduce non-linearity.
- Down-sampling, using for instance max pooling that takes the biggest value of two neighboring data points.
Repeating this process means that each data level represents larger and larger sections of the original data to build the neural network.
How can I get started?
Firstly, make sure that you really need to use deep learning. If your data is structured or relatively simple, you have a small data set, or if you are in a hurry then you should choose other options, for example MVDA methods such as OPLS. Having said that, if deep learning looks like a solution for you then get started with the highly popular open source environment Python, which is high level and easy to learn, and open source frameworks such as Tensorflow and PyTorch.
How can we help?
We at Sartorius are continuously looking at unstructured data and deep learning enables us to create a structured representation that can be combined with Multivariate Data Analysis to understand and control the processes involved. We are combining this into a framework called hybrid modeling that includes mechanistic modeling, multivariate linear modeling, and deep learning.
You can find out more by viewing the recorded webinar, ‘Deep learning: what is it and how does it work?’