Deep Learning Overview and Getting Started
Deep-fake videos of well-known politicians, Learning Super Marios, translated websites and automated agriculture: Hardly any other field has experienced such rapid growth in such a short period of time and has found its way into all forms of the world as the topic of deep learning.
Academically interesting as well as directly applicable in practice, today's marketing and reporting seems to have focused on the new all-purpose weapon. But what is it about that thing? What does this modern area of research, which already Turing Prize winners, but hasn't produced a passing Turing test yet? And where do I even start?
With this blog post, I would like to provide a brief overview of what various deep learning models and approaches can do today and how to get started on a technical level as well.
Step by step to the optimum
After the first expectant “AI boom” in the 1950s and the resulting “AI winter” of the 60s and 70s, there are in particular a few Key publications, advances in GPUs (graphics cards) and a sheer flood of data that has made today's deep learning possible. Regardless of the model type and architecture, these form the basis of all neural networks, whether for image processing, text and time series analysis or independent learning via deep reinforcement learning.
Once the basic concepts, which do not go beyond applied school mathematics, are understood, this work is promptly done by extremely powerful programming frameworks such as TensorFlow and Keras. These abstract the design and training of neural networks so much that the whole process appears deceptively simple. The number of adjustments and adjustments, hyperparameters However, it is enormous when it comes to neural networks, especially when you can create an applicable deep learning model in just 10-12 lines of code. The main task of a data scientist is therefore focused on designing models and searching for optimal hyperparameters; a process that is characterized by theoretical knowledge, but above all by empirical values.
A workshop for all cases
There are countless resources online about deep learning, such as medium articles, Reddit discussions and MOOCs. Although many aspects of the topic are covered here, these resources are often insufficient to be prepared to apply deep learning to real data and problems. That is why we have statworx We decided to combine our experience and knowledge in a comprehensive basic workshop: the Deep Learning Bootcamp.
It was extremely important to us to make the complex topic accessible, to have a common thread and yet to cover everything exciting. Not an easy task with the exciting oversupply of deep learning applications. Would I like to determine the content of texts faster than a person could, let a picture of my favorite celebrity grow a beard, or would I rather help a payment service provider identify credit card fraud? All of this is theoretically possible with deep learning. Since many use cases and procedures are very similar, it is crucial for your own understanding and identification of business cases to obtain solid basic knowledge and a good overview. In the workshop, we address a variety of questions so that our participants can get to know the various sub-areas of deep learning. This includes the following questions:
- How do I use 'normal' neural networks (multilayer perceptrons) to predict the price of a house or detect anomalies in my data? Whether in the financial sector, cleaning sensor data or for retail.
- What enable me Convolutional Neural Networks in image recognition? If I can recognize dogs, cats, airplanes, etc. when classifying images, how do I transfer the whole thing to my line of business. For example, could manufacturing errors be identified or the condition of fields monitored?
- How do I turn recurrent neural models to deal with sequential data? How does analyzing financial market or machine data differ from dealing with texts in order to distinguish a good cover letter from a spam message?
- How can I use the same reinforcement learning algorithms that let the computer master a video game using Deep Q-learning Optimize processes in real time, such as in a chemical reactor?
Experience & best practices
From a technical and theoretical point of view, the differences between the respective applications of the individual model types are sometimes shockingly small. Conversely, however, the learning effect and transferability are enormous. The only catch is that there are rarely rules of thumb or guidelines that provide you with the best model or hyperparameters.
This is where the complexity of deep learning models is revealed. Because of the high calculation effort, it is not plausible to try out all or many settings. Personal experience, best practices and empirical findings are thus in the foreground. While these are implicit in some online resources, we (and student feedback) believe that we can add significant value to our course. As data scientists who work on a variety of projects every year, we are also happy to share our own experiences and (painful) lessons learned along the basics.
Interested in Deep Learning Bootcamp?
If we were able to pique your interest, then take a look at our detailed course plan Write to or write us a message if you would like to develop a prototype together with us in a workshop! Otherwise, we wish everyone an exciting and good start into the world of deep learning, see you on Stack Overflow. Several times. Every day...
