Artificial intelligence (AI) is omnipresent, it appears. On our phones and at home, we encounter it. If entrepreneurs and business innovators are to be believed, AI will be a part of almost every good or service we use before we know it. Additionally, its use in solving commercial problems is expanding quickly. At the same time, worries about the potential effects of AI are growing. We are concerned about how automation enabled by AI will affect the workplace, employment, and society. 

Opportunities, restrictions, and challenges

Understanding recent developments in deep-learning algorithms is a good place to start. These innovations, which are arguably the most intriguing in AI, are bringing leaps in categorization and prediction accuracy without the typical 'feature engineering' associated with traditional supervised learning. Deep learning uses massive neural networks with layers made up of millions of simulated 'neurons.' Convolutional neural networks (CNNs) and recurrent neural networks are the two most popular types of networks (RNNs). These neural networks pick up new information using back propagation techniques and training data.

Labeling of data

The majority of modern AI models are developed using 'supervised learning.' This implies that the underlying data must be labeled and categorized by humans, which can be a laborious and error-prone task. For instance, businesses creating self-driving car technology are employing hundreds of individuals to manually annotate hours of video feeds from test vehicles to aid in the training of these systems. In-stream supervision, which was demonstrated by Eric Horvitz and his Microsoft Research colleagues, is one of the potential new techniques that is emerging at the same time. With this technique, data may be tagged as it is being used naturally. The demand for large, labeled data sets can be reduced with unsupervised or semi supervised methods. Reinforcement learning and generative adversarial networks are two strategies that show promise. 

Obtaining huge data sets for training

It has previously been demonstrated that basic AI methods based on linear models can sometimes approach the expertise of professionals in the domains of medicine and other professions. However, the most recent wave of machine learning needs training data sets that are suitably large and inclusive in addition to being labeled. When using deep learning techniques, models need millions of data records to perform at the level of humans and thousands of records to become reasonably successful at categorization tasks.

The issue is that for many corporate use cases, it can be challenging to access or develop big data volumes (think: limited clinical-trial data to predict treatment outcomes more accurately).

The issue of explainability

The problem of explainability for AI systems is not new. But it has expanded in tandem with deep learning's popularity and uptake, which has led to both more varied and sophisticated applications and increased opaqueness. It is challenging to describe in human words why a certain decision was made when using larger and more complicated models (and even harder when it was reached in real time). This is one of the reasons why explainability is still relevant or even necessary in application domains where some AI methods are still not widely adopted. Additionally, as the use of AI grows, regulatory restrictions can potentially increase the demand for more comprehensible AI models.