Artificial Intelligence - An Overview

 

In this post, I would like to cover a high level overview of Artificial Intelligence that I found crucial to understand the field of AI and ML from a general perspective.

Important Points to know about AI -

1.  Learning Input A to output B mappings is Supervised Learning. Can work with both structured and non-structured data. Key Applications - Spam/Not Spam Classification, Detecting Fraudulent Transactions, Speech Recognition in smart speakers like Alexa etc. (taking Audio as input and mapping to Text transcript as output), Language Translation, Online Advertising & Recommendation systems (adv and user info as inputs to click or not as Output), Visual Inspection / Image recognition, Self-driving car (images and sensor info as inputs to objects detected and their positions as output),  Text Analytics like Resume screening , Chatbots and Large Language Models (LLM’s) like ChatGPT etc. ( sequence of words as input to the Next word as Output) to generate new text etc. Turns out that the importance of supervised Learning cannot be emphasized enough!

2.  Although AI/ML has been there since long (literally for e.g.: Naïve Bayes was developed in 1760’s and still works great in my opinion, Decision tree was invented in 1963 and similarly many algorithms have been around for long time), AI has recently taken off due to rise in Large Neural Networks and Deep Learning possible due to recent advancements in storage and processing power/chip industry (like Intel, Nvidia etc.) as well as due to rise in Big data enabled through rise of internet and various software appl. capturing enormous data.

3.    Data is critical but needs to be relevant, hence important to check and gather data that adds value.

4.    Primary Output of a Data Science Project is mainly data insights in the form of a Hypothesis or ppt suggesting actions, while the primary output of a ML project is the models that has learned from the data and can make predictions on the unseen data (generalization) without explicitly telling or programming it with any rules/logic. (Rules are learned automatically in the form of the model).

5.   Conducting due diligence and making AI and data strategy relevant to your use-case is critical. Pulling all the data together and rehosting in a unified Data warehouse is crucial for useful data analysis and connecting the dots.

6.   Achieving AI Digital Transformation requires running pilot projects to gain momentum, building inhouse AI team, broad user training, developing AI strategy and internal/external communications

7.    Importance of Network effects cannot be emphasized enough. Large tech co.’s leverage this making them even bigger like Google, Facebook etc.

8.   Don’t expect AI to solve everything - Important to be realistic and identify if a problem is a good use-case.  I find the 1 second rule suggested by Andrew for AI/ML automation very practical and useful

9.   Computer vision is an important AI application area - comprising Image classification (labelling whole image - single or multiclass classification), Object Detection (draw rectangles to detect and locate objects), Image segmentation (takes 1 step further - precisely identifying and separating objects and boundaries at pixel level) and Tracking (detecting direction and trajectory of moving objects)- all essential & useful components for a Self-driving car model.

10.   Natural Language Processing (NLP) is another important application area. Text Classification, Sentiment Analysis, Content Filtering, Information Retrieval/Web search, Name entity recognition (e.g. finding all people names / location names in a sentence) and extracting names/phone no’s, summarization, Speech recognition (Speech to text) used by smart speakers and LLM’s are all examples of this.

11.  Key AI/ML Steps of Smart Speaker (Alexa) are - Trigger word detection -> Speech Recognition -> Intent Recognition -> Command Execution. The first 3 steps are implemented through ML models (supervised classification) and the last needs software code written to execute that command functionality eg: execute a song / joke / quiz etc.

12.  Transfer Learning enables a model learned & trained on a Task A to use and apply its knowledge for task B basically serving as a starting point for another task. Eg: Model trained on detecting cars can be used to detect a golf cart or a bullock cart. Basically, it is very useful when you have limited data for second task and the learnings from one task can be useful and applied for the second task. E.g.: - A drone model trained on identifying planes can be used for identifying birds.

13.   Reinforcement Learning - Another vast and useful application area. This is the 2nd type of AI. Uses Reward signal to provide feedback - good or bad. It automatically learns to maximize its rewards. Eg: Playing chess etc. where the reward or feedback may not be available immediately until after the game is completed. Example of controlling a helicopter through providing reward feedback signals remotely to make the right turns was amazing.

14.  Unsupervised Learning - 3rd type of AI - Extracts meaningful insights and patterns in data and groups it accordingly in clusters. This type of AI does not need or use targets or labelled data. Useful in customer segmentation for marketing etc. Very useful to learn about data and uncover associations or insights never known or imagined.

15.  GAN - Generative Adversarial Network - a class of adversarial neural network model in Deep Learning useful for synthesizing new images and other types of data from scratch. It is a sort of unsupervised learning.

16.   Knowledge Graph - Data that lists / representing real world entities such as objects, people, movies or events etc. and key information about them. Useful in information retrieval and for enhancing web search results. It is also known as semantic network and usually stored in Graph database (ibm.com). Wikipedia and google search use Knowledge graph.

17.  Generative AI - umbrella term referring to collection of AI models - used to synthesize and produce new content - text, images, audio, video etc. GAN’s are the type of Generative AI. LLM’s like Open AI’s ChatGPT and Microsoft Copilot utilizing DALL·E 3 are all examples of Generative AI.

18.  Robotics - refers to Motion Planning & control. Refers to intelligent machines that can learn from data and feedback signals and act autonomously or semi-autonomously. Self-Driving car can be an example of this. Based on perception and figuring out what’s in the world around you. Healthcare and industrial automation can be other application areas.

19.  AI impact on jobs worldwide will mean automation of various tasks (primarily routine and repetitive) through AI and generative AI like Office Support, Production work, Customer Services & Sales etc. and creating different types of new job roles in industries.

20.  Limitations of AI - Prone to Bias and giving biased results. Can be harmful for ethical reasons and company’s reputation. Difficult for regulatory compliance.  Susceptible to Physical attacks to fool the systems and Adversarial attacks due to minor perturbations.  Explainability and interpretability can be a problem and thus trusting on results of AI/ML model can be difficult.

21.  Solutions exist - both technical and non-technical to improve explainability and interpretability and improve bias. E.g.: heatmap etc. This is an open research area. Bias can be improved but not eliminated through feeding unbiased and balanced data.  Another way is to zero out the bias in words by setting them to 0.  Systematic Auditing and regular reviewing are crucial.

22.  Adversarial attacks & defenses - All this comes with a cost and can result in a sort of arms race for breaking the defenses of models and creating more robust defenses against attackers. Can become a zero-sum game. E.g.: Deep Fakes can synthesize realistically looking photos/videos of things/people/objects that never existed. Fake comments & attacks on privacy & democracy can be problematic as well.

23.  Models can be deployed on-premises, on cloud or on-Edge.  Edge deployment refers to running some part of AI/ML models on edge devices such as - IoT devices, smartphones, etc. Smart speakers such as Alexa have edge deployment. This helps to increase performance and makes processing faster avoiding the network and server-side processing.

24.  Lastly, Most of the AI and ML applications we see today such as Chatbots, Large Language Models (LLM’s), or a Self-Driving car are all a form of Narrow AI (which are trained on specific task and outputs or predicts a specialized task) and not General AI i.e. Artificial General Intelligence (AGI) that can match or even out-perform the human capabilities in terms of reasoning and intellect). Fears of being taken over by the AGI (like a Terminator’s Skynet) is unfounded and are a distant reality.

This provides a high level summary of key and noteworthy points on Artificial Intelligence technology at a broader level.