The current developments in Artificial Intelligence have the potential to transform our society. Recent scientific research could endow the computer with superhuman abilities. And since these studies are still in early stages much more advancements are expected in a near future.
AI is migrating from solutions built over sets of rules, or symbolic AI, to solutions that extract information over sets of data, or distributed representations (DARPA AI perspective). The current availability of data and computational resources allowed scientists to create and train machine learning algorithms of several abstraction levels, thus the name deep learning. These models are able to extract meaningful information from large amounts of data and currently present the best results in image, audio and text processing (deep learning survey).
Deep learning algorithms already present results at the same level of humans or even better (superhuman). Although the results are good there are several problems to implement AI solutions like ethical dilemmas and discrimination. These models can also be easily tricked (or hacked), see Adversarial examples.
Although the excitement is justified we are still far from skynet and the terminator. An artificial general intelligence (AGI) that represents machines with consciousness is far from happening. However, there are several other problems that emerge from the current technological power, see the Overview section for a discussion about malicious use of AI.
This website presents a series of explanations over basic concepts related to deep learning. The main idea is to present the concepts and simple code in Python to clarify how these models are implemented. The appendix presents discussions on specific or advanced topics.
Deep Learning
- Overview
- Perceptron & Gradient Descent
- Multilayer Perceptron & Backpropagation
- Softmax
- Convolution and Pooling
APPENDIX