Machine learning gives computers the ability to learn without being explicitly programmed. Instead, it learns from annotated historical data. It powers key applications of 2017 including voice assistance, image understanding, and content recommendation.
The goal of supervised learning is to learn patterns from historical data and find similar patterns in new samples. Input data must be annotated often by a human. Supervised learning is the most common machine learning and includes applications like image recognition, object detection and natural language processing.
The goal of unsupervised learning is to find patterns in the set of unlabeled data. The common unsupervised method is cluster analysis and it is used to find hidden patterns or grouping in data. Its Application in computer vision is for example grouping images with similar features or style.
Machine learning consists of simpler algorithms used in data analytics. Its benefit is good mathematical understanding of inner structure and performance. It contains algorithms like Linear Regression, Decision Tree or Random Forest.
Deep learning uses many layers of neurons designed to solve a specific task like image recognition or face recognition . It is very time and data consuming to train a deep network. Deep learning has developed in last few years thanks to the parallel processing using GPUs.
Living room, plant, sofa
spacious, B&W, lamp, living room
Accepted/not accepted; modern style
Wall
Home cinema (x,y,w,h)
40% similar
Most of the machine learning applications are designed to improve in time based on a data feedback loop. Machine learning flow is based on new generated data and feedback for the current model performance. Model is periodically retrained to reflect the most relevant data.
+ Feedback loop and improvements
+ Simple scaling and parallelism
+ Higher accuracy
+ Versioning
- One second latency
- Higher bandwidth
+ Low latency
+ Distributed performance
- Battery draining
- Lower accuracy
- No feedback loop
Single image processing fires 8 billion floating point operations
Each model has 25 million of parameters
Neural network pre-trained on 50 million images
Each GPU provides 5 TFLOPS of processing power
+ Feedback loop and big data
+ Problem overview
+ ML optimization and meta-learning
+ Infrastructure and scaling
+ No development time and cost
+ Covers specific scenarios
+ No data transfer
- High development time and costs