AI is finding it's way in a wide variety of applications pertaining to almost every industry. This AI driven rapidly evolving landscape has created a demand for a unique blend of technical, creative, and interpersonal skills highly sought-after by employers. Listed below are some specialized AI-related skills that are becoming increasingly valuable in the modern times. 1. AI Models Development Understanding how AI and ML work including the underlying algorithms, and learning to develop ML powered apps using tools like TensorFlow or PyTorch is a highly desirable skill to master in the age of AI. Furthermore, the skills in fine-tuning and adapting large pre-trained models (like GPT, BERT, or Vision Transformers) to specific use cases are also useful, allowing you to create specialized applications without starting from scratch. Leveraging pre-trained models and adapting them to new tasks with limited data is particularly useful in NLP and computer vision. 2. AI Models Deployme...
Artificial Intelligence (AI) is actively being pursued for safety-critical applications despite its challenges and shortcomings. Scientists and engineers across the world have been exploring improved AI techniques and models so that the underlying systems remain safe and robust. In this post, I will list few of the amazing applications of AI in safety-critical systems, challenges and possible solutions for their useful implementation.
SAFETY-CRITICAL SYSTEMS
Safety-critical systems are the systems in which a failure may cause damage to life, property or environment. A safety-critical system comprises all the hardware, software or human-machine interaction, any of which performs safety-related functions. Some of the popular examples of safety-critical systems are in the areas of transportation and medicine. Such systems typically require strong guarantees on their correct functionality, robustness, reliability and performance. Such guarantees are provided by conformance to relevant international standards and government regulations. Conformance to strict regulations is achieved by rigorous design and development process. Some of the popular standards in respective domains are: general (IEC 61508), aviation (DO-178B), automotive (ISO 26262), medical (IEC 62304) and nuclear (IEC 61513).
ARTIFICIAL INTELLIGENCE (AI)
Artificial Intelligence is the intelligence demonstrated by machines to achieve complex goals. Popular approaches to AI may be categorized as learning-based AI and rule-based AI. Rule-based AI is the classic form of AI in which knowledge is represented in the form of facts and rules, commonly used in “expert” systems. Learning-based AI is the current and more efficient form of AI employing machine learning (ML) techniques to represent information and knowledge. A neural network is an example of a learning-based AI system.
The ability to learn in learning-based AI simulates adaptive intelligence, which means that prior knowledge may be updated and new knowledge may be acquired. In other words, similar to human intelligence, these systems build rules based on experiences.
Rule-based AI systems employ explicitly defined static models of a domain
Learning-based AI systems create their own models
Learning-based AI systems create their own models
APPLICATIONS
Because AI can achieve narrow goals more effectively than humans, it has huge potential to improve healthcare and wellbeing of the society. AI has been applied to many areas in safety-critical domain, both to complement or replace human ingenuity. One of the amazing applications of AI in safety-critical settings, like transportation and healthcare, is computer vision.
Example 1: A fully autonomous car using deep learning to identify obstacles and objects.
Example 2: A medical device using deep learning to detect lung cancer on X-ray images.
Example 3: A robotic arm to assist the pilot during flying.
CHALLENGES
Although AI technology has huge benefits in safety-critical systems, its incorrect implementation in these systems may have deadly consequences. Fundamental problem lies in the implementation of machine learning algorithms underpinning AI systems, which have been designed to learn and modify their own behavior at runtime. The algorithms may not generalize correctly to new scenarios, resulting in erroneous perception and reasoning. As an example, consider a deep learning algorithm powering a computer vision application in a self-driving car. The algorithm may fail to correctly perceive: obstacles in bad weather, lane markings covered by snow or traffic signs obscured by trees. Consequently, the algorithm may fail to invoke the corresponding rules and actions, such as: apply brakes, control steering or reduce speed.
Another challenge is achieving safety certification for AI based safety-critical systems. Safety certification involves verifying that anything which could happen at runtime has been considered and tested. Learning-based AI implies runtime software evolution which cannot be perfectly predicted and verified in advance, thus complicating the certification. Today, safety certification of learning-based AI systems is an area that is the focus of major research.
Apart from technical challenges, there are various non-technical issues pertaining to the use of AI in safety-critical applications. As the technology advances and becomes widely used in practice, it will be important to address the problems such as legal implications and ethical considerations in its use. For example, when self-driving car crashes into a pedestrian, who is to be prosecuted - the car, the manufacturer or the designer. Similarly, there is plenty of work to do in order to establish the ethical foundation for using AI technology safely and effectively in health care.
ACCIDENTS
Design issues in safety-critical AI systems have resulted in many accidents, thereby reducing consumer trust in these systems. For example, in self-driving cars, problems with computer vision have been concluded as contributing factors in many Tesla crashes. IBM’s Watson AI engine, when applied to medical applications, has been deemed unsafe for treating cancer patients. It is possible that even the seemingly simpler applications of AI may turn out to be a safety problem, if the underlying design is not implemented properly. For example, a house cleaning robot may damage property or life by knocking down objects in its path as a side effect.
SOLUTIONS
To address the known gaps in using AI for safety-critical applications, there are some solutions to minimize the chances of failure. The challenges may be huge, but they do not limit the use of AI in safety-critical settings with human in the loop. Robustness of a system can be greatly improved by using properly curated training data to build expressive and well-specified ML model. Extensive testing and critical evaluation of the system before deployment is particularly important and valuable.
CONCLUSION
AI is being used in a variety of applications such as shopping, social media, home automation, smartphones and websites. Implementation of AI in the domain of safety-critical systems has known gaps demanding some research and study. Due to the promising benefits of AI technology, many organizations across the world are investing heavily in this domain. In order for the technology to be widely accepted, it has to be improved to work to perfection.
Comments
Post a Comment