Evaluate the use of Autonomous Intelligence technologies
Machine Learning (ML) and Deep Learning (DL) have reached human parity level (they perform as good as human on the task) for many applications, including object recognition, speech recognition, reading comprehension, machine translation and general language processing; and near human parity for speech synthesis. However, ML and DL algorithms suffer from the limitations of data, i.e., the results will only be as good as the quality of the input data and bias is going to be amplified. They're also black boxes that have none or reduced explainability, and therefore aren't trusted, especially in industrial settings. They also have errors (false positives and false negatives) that aren't human-like errors and are unexpected (every inference can be an error).
Deep Reinforcement Learning (DRL) algorithms have many “superpowers”: they can change behavior in response to changing conditions; can learn complex, non-linear relationships that require decisions to be made across fuzzy, non-linear relationships between variables, transits of the environment state; act on complex perception; respond to environment changes that aren't directly measured; and the most important one, it learns strategy.
However, DRLs also have limitations: it has no previous knowledge of the task at hand - which means it has to learn everything by a very large number of trial-and-error iterations. This process is inefficient if it has to navigate a very large state space. Fortunately, it's possible to reduce the complexity of DRLs by introducing domain knowledge to reduce the search space.
In the following table we summarize the strengths and limitations of the two main technologies included in Autonomous Intelligence: ML and DRL.
| What is it? | When should you use it? | What are the limitations? | |
|---|---|---|---|
| - Machine Learning (ML) - Neural Networks (NN) - Deep Machine Learning (DML) |
Network of interconnected nodes that learns from data to recognize patterns or replicate behavior | - When you need advance perception: computer vision, natural language processing, sound, etc. - Classification - Prediction |
- Difficult to train - Limited / non-intuitive explainability (black box) - Easily hackable - Every inference can be a false positive or negative - Low reliability, non-deterministic - Not actionable, difficult to integrate into an industrial process - When you need advance perception: computer vision, natural language processing, sound, etc. - Classification - Low adoptability |
| - Reinforcement Learning (RL) - Deep Reinforcement Learning (DRL) |
An ML that practices in simulation / real world | - Difficult state assessment, non-linear, dynamic environments. - When you need advance perception: computer vision, natural language processing, sound, etc. - Classification - When we need to learn strategy at human level |
- Difficult to train. - Limited / non-intuitive explainability (black box). - Difficult to integrate into an industrial process. - When you need advance perception: computer vision, natural language processing, sound, etc. - Classification - No previous knowledge of the task at hand. |