What are the latest AI trends to need to know?
Artificial Intelligence has recently seen a lot of attention due to advancements in deep learning and the increased availability of open-source datasets. Learning algorithms, like Deep Neural Networks (DNNs), can now use large amounts of data generated by social media platforms, image processing services, scientific research papers, and video game engines. This generated dataset is much larger than previous datasets that were used for training these types of models.
This increased availability has made publishing research papers that would not have been accepted 5 years ago. Now artificial intelligence researchers are publishing more academic articles than ever before:
“The number of research papers published in the field went up from 150 between 2010–2011 to 368 in 2015–2016, an increase of about 128%. An example of a company that is using deep learning for its product or service is Google. They are using AI to classify images via Deep Learning. This technology has been trained on over 1 billion images and can recognize cars, trees, people, and animals, among other things:
Surge in AI
A surge in artificial intelligence startups has also raised concern for investors due to a large amount of money being poured into these companies. The total value in funding in 2016 was around $5B+. Investors are concerned because only 10 companies account for approximately 70% of all funding. Also, fewer companies receive large amounts of funding compared to 2015, when 30 companies received at least $100 million in investment. Due to the recent increase in deep learning models and the increased availability of open-source datasets it has led to a surge in AI startups receiving large sums of funding. With more developments happening in this space, I wonder how the role of data scientists will be affected? And if you’re looking for a job, what skills should you focus on?
The world of AI often invokes images of fully sentient robot overlords, advanced technology indistinguishable from magic and compelling AIs that quickly end the lives of those who oppose them. In truth, much as is with many areas within computer science, most artificial intelligence applications are far more mundane – performing tedious tasks such as closing captchas, making recommendations for products or services, speech recognition, etc.
Generally, all entities in a natural language text can be described as a combination of a class (e.g., Person or Location) and a collection of properties (e.g., occupation: ‘Doctor’ or location: ‘Manchester’). It results in an entity-property matrix where each row represents an entity and columns represent properties.
If we now want to discover the association between entities, it would be possible to look at every pair of the entities in our dataset and record if they are associated or not by creating an entry in a table where each row represents one entity pairing. Each column represents whether the two entities have been linked together within any documents (or rows) in your data set. For example, you might find that Person A is often mentioned for Location B. However, Person A is never mentioned for Person B. This type of matrix is often called “the co-occurrence matrix.”
Evolving face of AI
However, just as the world of AI has moved on from dark and gothic alleys, modern techniques for discovering such relationships have evolved. One such technique involves using statistical reasoning to work out if two entities are likely to be associated together or not. This idea forms the basis of semantic role labeling, which uses words occurring close to another word (neighbors) to indicate their relationship with each other (e.g., ‘woman’ knows about ‘shoe’). Recent developments also include the use of networks for this task. So instead of creating a table where rows represent entities, columns represent properties, and entries indicate if two entities are linked together, an interactive network is built with arcs representing the relationships between entities.
The most important advantage of using networks is that it can be far easier to extract very general information according to RemoteDBA.com. For example, if we search for the entity ‘red’ in a network created for this purpose, we don’t need to know which ‘red’ refers to (apple? cardigan? Rose?). Instead, all red nodes will be returned (and their neighbors) regardless of what they specifically refer to.
Another helpful feature of these techniques is that traditional supervised machine learning methods can also be used at this stage with great success – often resulting in an accuracy score on the data set itself (whereas previous techniques may only result in high accuracy on unseen data). A second technique, word2vec, allows us to create a dense vector representation for words in our text data (the closer two vectors are, the more similar they are). By training word2vec on large volumes of text, we can determine which words often occur in groups, and this forms an alternative way to discover relationships between entities.
This approach is helpful because it allows us to discover associations we already know about (such as that ‘apple’ and ‘computer’ can be associated with each other) and that less obvious links may exist within the dataset. One such example might be discovering that ‘cheese’ is frequently found near ‘bread,’ however, rarely occurring together with other items. This knowledge could then indicate that these two entities form some relationship within the dataset.
Modern AI Techniques
Modern AI techniques are now merging to provide us with a coherent approach to discovering both known and unknown associations between entities within text data – often resulting in an accuracy score on the data itself. Perhaps more importantly, this also allows us to understand better what our data is talking about.
AI trends can be described as a combination of a class (e.g., Person or Location) and a collection of properties (e.g., occupation: ‘Doctor’ or location: ‘Manchester’). This result in an entity-property matrix where each row represents an entity and columns represent properties. If we now want to discover the association between entities, it would be possible to look at every pair of the entities in our data set and see how often they co-occur.
For example, if we want to discover the association between ‘apple’ and ‘pie,’ ‘we could look at all pairs of entities (such as apple or apple pie) in our dataset (e.g., find every row where apple is mentioned near pie). Each time this occurs, we can create a 1; otherwise, it will be 0. Once complete, each row will now represent an entity and each column represents a property that exists for that particular entity (e.g., occupation: ‘Doctor’ or location: ‘Manchester’). This type of matrix is often called “the co-occurrence matrix”.
Karen Anthony is a Business Tech Analyst. She loves to share her tips with friends. She is passionate about new trendy gadgets.