Deep Learning


Definition

Artificial intelligence (AI) is the field of computer science that focuses on the effort to automate intellectual tasks normally performed by humans. The area within it which studies the possibility that computers can go beyond what is already known and find out how to complete an “untaught” task on their own is called machine learning. Nevertheless, AI does include other sections that do not involve any learning at all.

Deep learning is a subfield of machine learning concerned with algorithms inspired by the structure and function of the brain called artificial neural networks. A neural network consists of a certain number of successive layers to extract as much information as possible from a particular type of input. The structure within which said layers are created is named model and how many layers contribute to a model is defined as the depth of the model.

Applications

  • Drug discovery and toxicology - Machine learning approaches within the pharmaceutical industry since they provide a set of tools that can accelerate and streamline the drug development pipeline in preclinical studies as well as in subsequent clinical trials.
  • Customer relationship management - Another strength of deep learning is the ability to effectively make predictions based on data analysis. This includes the approximation of direct marketing actions, which is valuable for a company’s interaction with current and potential clients.
  • Recommendation systems - Deep learning have been used to extract meaningful features for a factor model for content-based music recommendations, this type of model is called Content Based Filtering. However, AI has also influenced another recommender, known as Knowledge Based Filtering. Said recommender asks users about what they are looking for, and finds selections based on the user’s requirements.
  • Bioinformatics - Once more, deep learning is implemented looking for predictions, but this time to foresee gene ontology annotations and predictions of health complications from electronic health record data.
  • Mobile advertising - Due to the increased mobile ad fraud rates, people have been requesting for deep learning to detect them by constant monitoring and big data analysis. In addition, AI Solutions can predict what category of users is the most appropriate for a specific marketing goal, allowing companies to save money and improve the ad performance.
  • Financial fraud detection - Nowadays rules alone are not enough to block fraudulent payments since this system involves attempting to update rules as fraud methods evolve, which is significantly burdensome. Deep learning is able to constantly analyse individual customer behaviour and block or flag a payment as soon as it spots an anomaly. Besides, algorithms can easily be adapted to a greater scale as larger datasets give more examples to rely on.
  • Robotic control - Artificial intelligence alongside robotics is opening the door to entirely new automation possibilities. Currently, AI and machine learning are impacting many areas of robotic processes (vision, grasping, motion control, etc.), making applications more efficient and profitable. Even though progress is being made in a limited way, current applications are promising.

Challenges

  • Quality data - The larger the data available for deep learning, the better its performance. However, if there is not enough high-quality data (whether it correctly represents the real-world construct to which it refers), then any deep learning system is likely to fail.
  • Expectations - There has always been a discrepancy between the expectations of AI technologies and what will likely be possible. This is the reason why there have been two AI winters (dearth of funding as a result of a cycle of intense optimism followed by disappointment and skepticism). Ultimately, the way this challenge is being addressed is by trying to understand that AI is indeed a tool which enhances productivity but cannot replace the ideas of a human brain.
  • Becoming Production-Ready - Since many companies are currently investing in AI, there will be growing pressure on organizations and their developers to transition from modelling to releasing production-grade AI solutions. After all, the significant investments in AI need to translate to value in solving real-life problems if they are to be considered worthwhile.
  • Context - Deep learning algorithms lack of a satisfactory level of understanding when it comes to analysing context. For example, an algorithm might become highly proficient at mastering a video game to the point it can easily defeat human players. However, change the game, and the neural network needs to be trained all over again because it does not understand context.
  • Security - Considering the propensity for outputs from neural networks to be altered after input modifications, these systems are vulnerable to venomous attacks. For instance, self-driving vehicles are partly powered by deep learning. If there were any input alterations in the learning model, the vehicle behaviour could potentially be controlled in a malicious manner.