Following is a list of problems that can be solved by using AI technology:
Recognizing Patterns (Classification)
Object recognition and detection as in the case of face recognition (including emotions), classifying loan applications into eligible and not eligible, classifying customers who are likely to churn or continue with the service, classifying hand written digits into 0 to 9, sentiment analysis (positive, neutral, negative) are some examples.
Sales/demand forecasting, predicting future stock price and real estate prices, inventory forecasting, product/movie recommendations, etc.
Though optimization is used widely in Machine Learning (ML) as a tool to reduce the error between prediction and ground-truth (actual) during the training, it’s applications are not considered as part of AI or ML. This field has emerged as part of operations research, and has well known applications such as a travelling salesman’s problem, choice of location for distribution center, airline hubs, retail store locations, marketing spend optimization etc. It is a vast field and has many sub-domains such as linear/non-linear programming, integer programming, dynamic programming etc.
This is a new application area derived from mostly the Deep Learning field. It is essentially creating something new from few inputs. e.g. generation of an art work for a given style and content image, generation of music for given pattern(s), grammar correction in the text, filling faded portions in an image or removing unwanted portions from an image, generating captions for images or generating images for a given text description, writing the summary of an article, etc.
There are many situations where we don’t know what the desired output is, e.g. which credit card transactions are fraudulent, which state of IT server causes it to fail, which state of nuclear power plant causes it to shut down or leak, when does ECG show malfunctioning of the heart etc. In all such cases, we don’t have enough examples of failures to model as in the case of many other problems explained above. Anomaly Detection techniques come to the rescue in such cases. These techniques model what the normal behavior is and raise alarms when an event deviates from this normal state.
This is similar to classification of transactions/objects into multiple groups. However, in case of classification, the types of classes into which objects need to be classified are known, whereas it is not known upfront in the case of Clustering. The objective of clustering is to figure out various types of groups within a given set of data. For example, clustering techniques such as K, means clustering can be used to understand various types of customer groups in a customer database, or to group all news items based on topics that they belong to such as sports, culture, national, politics, etc.
In many real-life problems, a large number of attributes influence target output, e.g. house prices are influenced by area, number of rooms, region, proximity to schools, highways, malls, age of the house, materials/fittings used for construction etc. In one of the studies, more than 150 such attributes that could influence the price of the house were identified. It is humanly impossible to comprehend, which of these attributes really influence the price and to what extent, and this makes the decision complicated or uncertain.
In such cases, Dimensionality Reduction techniques can be used to reduce the number of attributes, without any information loss from original attributes, so that the problem is simplified both from computational and comprehension perspectives. Another example is image processing, where attributes are individual pixels, which run into few thousands. This is again computationally intensive and hard to comprehend. PCA, t-SNE are some of the techniques available in this space.