Much of the current love for AI arguably comes from deep learning on neural networks. These are essentially brute force, pattern recognition machines that – if provided with enough training data – can go where more traditional data science (often based on statistics and mathematics) stops. Deep learning can be combined with other technology-enabled approaches, such as reinforcement learning, in order to provide even more raw, unmatched problem-solving power. Its simplicity is appealing, as it functions as a black box that simply needs lots of training data to become accurate. But as we live in a world of tools, it is now more than ever a matter of finding the right balance between man and machine powers.
- Many of the current breakthroughs in AI (although certainly not all of them) are due to deep learning machine learning models on neural networks; a way of detecting and classifying features through multiple layers in raw input.
- Provided there is abundant training data as input, deep learning neural networks can recognize patterns much more effectively than traditional (typically statistics and algorithm-based) data science approaches, increasingly more effectively than humans.
- Advances in the ability to collect, store and access large amounts of training, together with the emergence of powerful graphical processing units (GPUs) and other hardware accelerators have been instrumental to the current success.
- Deep learning neural networks prove to be useful in cognitive areas such as image, audio and speech recognition, natural language understanding, robotics, and in many complex analytical areas where traditional approaches are not sufficient, including drug discovery, customer behavior analysis, bioinformatics, medical applications, fraud and risk detection, predictive maintenance and notably Cyber Security and IT operations.
- Reinforcement learning uses an action/reward approach to learn from actual interaction (often in a simulated environment with synthetic data) to find optimized strategies and next steps. Combined with deep learning, it creates even more powerful AI applications in areas such as robotics, scheduling and gaming.
- The German retailer, Otto uses an unconventional deep learning algorithm (originally developed by CERN for particle-physics experiments) to predict what customers will order. Finding hidden patterns across 3-billion transactions, it considers over 200 variables – from weather to sales – reducing product returns by 2-million per year.
- Using many data features, including time stamps on transactional data, American Express found deep learning techniques – such as long short-term memory and temporal convolutional networks – can be adapted to enhance fraud detection results.
- UCLA researchers have developed a deep learning, GPU-powered device that can detect cancer cells in a few milliseconds, hundreds of times faster than previous methods.
- Using AWS Rekognition, an AI system was built for retailers to analyze real-time footage of foot fall within a store – to improve customer engagement, thereby increasing sales.
- AWS’s DeepRacer uses reinforcement learning on simulated, 3D virtual tracks to train models for fully autonomous 1:18 scale racecars; they can then compete on a real-life track without having been there before.
- Google’s AlphaGo Zero made the South Korean Go world champion Lee Se-dol retire from professional playing, after he was conclusively beaten by the system. Considered otherworldly complex, the game Go was believed to be beyond the reach of even the most sophisticated analytical systems, with an almost infinite number of configurations.
- Solving problems that were deemed impossible to solve – or insufficiently successful – with more classic data science approaches.
- Creating powerful, complex autonomous systems, even with an occasional lack of sufficient volumes of training data.
- Building next generation predictive and prescriptive analytics that go beyond human (or statistics-based) approaches in their capability to detect patterns in seemingly unmanageable volumes of unrelated data.
- Deep learning / neural networks: TensorFlow, Microsoft Cognitive Toolkit, Theano, MXNet, Keras, Chainer, PyTorch, Gluon, Horovod, AWS Deep Learning, Deepomatic computer vision
- Reinforcement learning: AWS DeepRacer, Facebook Horizon, Gym on OpenAI, Microsoft Project Malmo
- AI infrastructure accelerators: NVIDIA deep learning, AWS Deep Learning AMIs, Google Cloud TPU, Intel AI and Neural Compute Stick, Apple Neural Engine