Note: This article is present in its copy on author’s medium.com account as well.

#ai #artificialintelligence #robotics #machinelearning

This is a review of Chapter 2 of their book, Bartneck et al., on Ethics in AI and Robotics.

Key points in the chapter are as follows:

— AI definitions (Bartneck et al): It involves the study, design, and making of intelligent agents to achieve certain intelligence-based goals.

— It’s not necessary for AI to learn new concepts for example the surgery department in a hospital needs to do repetitive high-precision tasks.

— Even within AI there are various kinds of techniques to formulate intelligent solutions, for example, on one side is an experts system wherein there is a knowledge base on the basis of which problems are solved while there are prediction and classification algorithms which predicts solutions entirely by learning the new tasks.

— How machine is considered intelligent? This relates to Turning test.

— In a Turing test a human expert shall ask questions from both human and a machine. If the human expert is unable to distinguish between man and machine, then the machine is declared to have passed the test and called an intelligent machine. Even if a system pass Turing Test still it is not known to be a necessary or sufficient condition.

— This test makes the rather vague term of intelligence as a testable in terms of metrics.

— John Searle classified two kinds of AI — weak AI and strong AI. Weak AI consists of a specified number of tasks of intelligence that the algorithm can do. Strong intelligence or general intelligence, competes with human intelligence and solves a broad range of problems just like humans do.

— At present AI works in a simulated environment meant for testing and predicting the solutions. Changing the environment to a real-time environment may change some degree of solutions as well.

— Two major kinds of systems talked about by the author in this chapter are expert systems and machine learning models, as described above.

— Machine learning can be of three kinds — supervised learning, unsupervised learning, and reinforcement learning.

— Supervised learning consists of classification and regression tasks.

— Unsupervised learning consists of clustering and PCA-like solutions.

— Reinforcement learning provides feedback in the form of task completion or nearness to a goal, to mention a few. With these inputs, the algorithm learns and re-formulates the tasks to be performed. This can be done through sensors or verbal, or written feedback.

Robots

— Robots have a physical body.

— They can sense and act on basis of that.

— The sensors can be like camera which have their own drawback of light shifts.

— Using the values collected the Robot finds its goal.

— Machines don’t have human-like perception-gathering abilities.

— Robots understand non-symbolic sensors better than symbolic ones, which humans absorb well.

— Robotic system integration is often challenging as it involves combining all sensors and their output generation engines.

— Facing newer and ever-changing inputs can confuse robotic outputs.

References

C. Bartneck et al., Chapter 2, An Introduction to Ethics in Robotics and AI, SpringerBriefs in Ethics, 2021, https://doi.org/10.1007/978-3-030-51110-4_2