Infosys TechCompass

AI algorithms

Trend 1: Deep neural network architectures help improve generalization and accuracy

Deep-learning algorithms promise higher accuracy and better generalization characteristics than classical algorithms such as SVM, Naive Bayes, and random forest. Enterprise-class problems can be aptly resolved through graphics processing unit (GPU) computing; accessibility of large, labeled data; and fast-paced innovations in deep-learning algorithms. However, the need for a large set of labeled data and the cost of GPU computing are key challenges. Nevertheless, transfer learning-based models, which involve storing knowledge on one domain and then applying to other related problems, have made massive headway in overcoming the limitations of insufficient labeled data and GPU. Also, evolving architecture, such as transformers, has eased certain complex problems in computer vision, NLP, and speech domains.

A large technology company wanted to advance its existing system that worked on certain preconfigured historical rules and policies to moderate user-uploaded content.The company, in partnership with Infosys, developed an AI model for supervised transfer of learning-based deep neural net architecture for vision and text. This helped the company to identify, classify, and isolate any toxic content arriving from user-uploaded forms.

AI algorithms

Trend 2: Transition from system 1 to system 2 deep learning

The current state of deep learning-based AI is referred as System 1 deep learning, and it can be best illustrated with an example of a person driving a car in a known vicinity while talking on the phone or with a passenger, and is able to automatically drive through, without consciously focusing on driving. However, the same person driving through an unknown vicinity will need a lot more focus and will need to use various logical reasoning and connections to reach the destination. These types of problems, which need a combination of reasoning and a sense of on-the-fly decision-making, still can't be solved with current AI discipline maturity and are considered System 2 deep learning.

System 1 deep learning's current state is due to certain current limitations of deep learning's generalization capabilities, where these algorithms

  • are not able to correctly work on (detect) unseen data patterns;
  • need to have balanced distribution of data in training and testing sets;
  • lack the ability to do continuous learning based on changes in enviroments in real time, similar to active agents;
  • lack the logical and reasoning capability to combine high-level semantic concepts, and
  • are unable to deal with out-of-distribution (noise) data.

These are some of the reasons for the current state of AI's inability to deal with the System 2 deep learning state.

System 2 deep learning resolves some of these challenges by leveraging attention-based architectures and models (the general task of dealing with events over time) and multitask learning (multiple tasks solved at the same time), and incorporating principles of consciousness and meta learning, with an emphasis on unsupervised, zero-shot learning techniques. In zero-shot learning techniques, observed classes in the data are associated with non-observed classes through some form of auxiliary information. This speeds up processing times and increases the efficiency of tasks such as object detection and NLP.