MODEL TRANSPARENCY: Is Artificial Intelligence & Machine Learning A Blackbox?

In this week's?#HEXATalk?session we have some interesting people with us who?are going to speak about Artificial Intelligence and Machine Learning, particularly around MODEL TRANSPARENCY.??

Listen below?to the?interesting?discussion?and?why it is so exciting.

#HEXATalk Highlight?

1. Explain What Does AI?&?ML Really Mean??

People have a misconception about the fact that?AI?and?ML?are the same things?but it's not. AI can best be defined as the ability of the computer to replicate?human behavior as closely as possible and this is done through a computer program and the process?by which the computer?program is trained to exhibit?these traits?is called machine learning.??

Machine Learning, we can essentially say is broadly classified into supervised, unsupervised & reinforcement learning.?In supervised learning, the program knows what outcome it wants. Unsupervised learning the program figures out what outcome it needs on its own?& Reinforcement learning is a combination of both supervised & unsupervised?wherein a feedback loop tells?how well?the program is performing and it changes its parameters accordingly to make the?output as accurate as possible.?Methods of Machine Learning include regression, clustering, support vector machines, neural networks, and so on.??

Artificial Intelligence is classified into cognitive and natural language processing.?Cognitive is intelligence regarding computations?while Natural language processing?(NLP)?is more about comprehending human language, example of it would be google translate being able to detect the language that we are typing and?then?asking for a translation.??

AI?dates?way back?to?the 1960s, the technology available around us in the past few years is cheaply available and is making AI more and more useful and applicable.?

Fun Facts:?

• Citadel has only 40 employees working on the trading floor to oversee computers striking trades for around 900Mn shares per day.?
• Goldman Sachs employs more Engineers and Programmers than Facebook.??

2. ?How scared we should be with such examples where there could be bias and no transparency in the model???

There has always been a feeling of uncertainty surrounding?the implementation of machine learning models and this fear of the unknown arises?because we mainly don't really understand how a model works but we can actually do so...??

Machine Learning models are built using codes like there is an extensive amount of code that runs in the background.?A?model consists of?different algorithms, to build an algorithm we have to understand all the coding intricacies very carefully that are involved there. However, to use an algorithm?we can do it without even going into the depths of the architecture.??

To understand better here's a simple analogy, consider the model to?be an F1 car and we here at HEXANIKA are a group of people some of whom know exactly how to design and build this car and some know only how to drive this car. It is never the case that the drivers who don't know how to build this car cannot operate the car, in fact, the more experienced drivers?are actually very well acquainted with how the changes in different parts will affect the working of the car. So, here's an example,?there?is?a?bachelor's?student?and?he?has?no experience in actually building ML models however?he knows how to use them and perhaps is not scared of using them.??

3. Explain?the process with a project or use case?that you worked on in HEXANIKA.?

Here is an example of a very simple algorithm that was used in one of our projects recently. Often in real life, we have to deal with unbalanced data sets?and before fitting them on the model we have to ensure that we balance them so that the model can observe all the classes present equally and be able to differentiate between them.??

In this project, we were given a variety of attributes and our main aim was to predict based on these attributes whether a particular loan application would be approved or denied. So, the data sets that we used for training had quite a lot of data present?there. The?no. of?approvals were around 80% of the entire data set whereas the?no. of denials were only 20%?and this is where the bias comes in. If we fit such a type of data into our model, the model would learn the approvals very well however it would not learn the denials as expected and we want to predict that correctly.?Hence,?we?should?implement?a technique?to balance this data set before we actually fit it into our model. The technique that we used for balancing is SMOTE (Synthetic Minority Oversampling Technique)??

4. Give some additional examples that can shed some light on the topic??

Before discussing how it is useful for financial operations, neural networks?have evolved a lot from the past. So, this neural network is generally BERT (Bidirectional Encoder Representations from Transformers)?model. These are the things which can be used for finance even though they are trained?particularly?in?English?to be the?main language.?Masked language modeling and next sentence prediction is a?way whereby?using?a sentence if we try to cover the sentence and make it invisible, the computer would try to predict what sentence could best fit in.?This helps us to get the sentiment analysis of the statement,?the summarization of the whole text, and also?helps us?analyze whether the statement is right or wrong.??

Recently we did a project on Auto rules creation, where we were given 180 business rules, we were trying to convert?the unstructured data to structured data. We were able to make a rule out of every article, we used keywords to test if there?is a rule or not. By this, we were able to create a rule which was almost similar to the actual rule. The rule that we created was more elaborated and simpler to understand.?So, in simple words there was guidance, we?created rules?and lastly?summarized them.??

---------------------------------------------------------------------------------------------------------------

Have thoughts on this week’s trends or questions for me or the Guests? Post your thoughts in the comment section or send a note to?[email protected].?Please include the hashtag #HEXATalk?and mention me, Yogesh Pandit! Until next week.????

About the speakers:

Advait?Varma has completed a bachelor's degree in electronics and telecommunication and is pursuing his master's in?Operations Research in Finance and Management from Columbia University.?When?it comes to his financial?ability,?he is focused on problem-solving and?has a desire to pursue a career in fintech.??

Saswata Mukherjee is pursuing his undergraduate as a student?in Mechanical Engineering?in?BITS Pilani, Goa,?and is very interested in Artificial Intelligence and related fields.??

Sindhu?Juloori is pursuing Mechanical Engineering from Bits Pilani and is keen?to learn about Artificial Intelligence and its applications in the industry.

#MachineLearning?#ArtificialIntelligence?#Analytics?#DataAnalytics?#DeepLearning?#BigData?#NeuralNetworks?#Team?#People?#Creative #Compliance #Innovationve#Regulations?#Automation?#ModelTransparency?#HEXATalk #Transparency?