IDL Lect1C Why AI Became a Black Box: From AI History to Deep Learning

追加: 2026-05-25

[00:00:00]math student to this lecture to 1C.

[00:00:05]And in this lecture, what we'll be taking you that we in the previous lecture we know that AI and deep learning became more non-interpretable and black box.

[00:00:18]So, we will see that how the history of AI from the history of AI we'll see that how it became a black box problem and non-interpretable problem. So, now when So, if we talk about AI end of 40s and practically 50s, 1950s it started. But when it AI field came up in reality by the scientists and researchers it was not as a black box. And early AI was more rule-based and it is easier for inspection. Means human can inspect why early AI system was failing or why it is making decision. It is on in the human language. Okay.

[00:01:10]So, it started as a white box system.

[00:01:14]Let's see what made it black box with history. Okay.

[00:01:20]So, you know that in early the in 1950s and 60s where the that was called the summer of AI first time AI came and people started thinking that everything AI will take over. There is no job in the future. This is a first summer of AI. Right now we we are in the fourth summer of AI. So, even now you are hearing that AI will take over your job and that was the case in 1950s as well.

[00:01:51]So, that symbolic AI came and the first time early AI was basically represented knowledge using symbols and logics, mathematical logic and handcrafted rules. So, for example in this in the for all X if PX belongs to QX then there exist a rule X. Okay. So, this is a discrete mathematics logic rules are there. Then P of A if intersected with Q of A then the rule B is going to be applied.

[00:02:28]Okay. So, if P of X then yes, then QX.

[00:02:32]Okay.

[00:02:34]P of X yes, then QX and if not then apply rule RX. So, handcrafted rule were there in the symbolic AI.

[00:02:43]Then comes the AI programming language came up, the Lisp which helps in AI programming and also symbolic reasoning more easily. So, this was when I was studying the bachelor degree in 1990 to 2003 that time in our AI course Lisp programming was being taught.

[00:03:06]So, this was been in 1960s Lisp came up and where it has been like defined a reason X if fact X is exist true and rule is there X then you do call this body kind of function true.

[00:03:23]So, this kind of AI based programming came up and later on the reasoning programming came up, the early reasoning system where facts like PA, QA, RA is there, rules like PX belong to QX QX intersection of RX then belong to SX then SX is going to be TX.

[00:03:42]So, this kind of rule was defined and search and inference engine was there then final answer was coming up. So, 50s and 60s it was practically human understandable and whatever happening human can understand.

[00:03:57]Explicit knowledge, logic, symbolic reasoning was there. So, it was a white box inspectable. Now, after 60s we moved to 70s and 80s. In 60s, 50s and 60s the AI system was not been able to be solve all our problem. So, then came the expert system in 70s. Now, this is a time when uh it is a second summer of uh AI where people started thinking AI is going to solve all the problem and all the jobs will go away. Uh So, in that time what people built uh there are facts and there are rules.

[00:04:38]So, let's say AI based diagnosis system is rules are there.

[00:04:43]Like like I making human logic into the program.

[00:04:48]So, if fever exist and infection is there then consider giving antibiotic. Okay.

[00:04:55]So, if fever exist and infection is there then consider giving antibiotic.

[00:05:00]If high white blood corpuscles are there, that means showing infection and cough is there then it may be a pneumonia.

[00:05:07]If fever is there and skin rashes are there then it may be a viral infection.

[00:05:12]So, this kind of rules was there and that was giving the decision. So, expert system was like handcrafted rule to mimic human kind of decision making.

[00:05:24]Then came Mycin [snorts] kind of rule-based expert system And then the new advancement like support vector machine came up which found optimal separation between two classes.

[00:08:22]Very nice innovation or I could call invention.

[00:08:27]So, SVM and neural network helped in learning from the data.

[00:08:33]Okay. So, we moved towards the performance. Performance became better.

[00:08:38]Models learn pattern instead of following rules and we started advancing and then early neural network came up like neural network system came up and it started learning from the data.

[00:08:52]Now, what happened after that?

[00:08:54]We are in 2010.

[00:08:56]And that changed. Like in my previous lecture I was telling you, right? I was doing PhD in 2008 and 9 I was working with Caltech 101 Caltech 256 data set where 101 classes and 256 classes was there. In Caltech 101 the total images was around 9,000 something.

[00:09:16]And Caltech 256 the total images was 30,000.

[00:09:20]And in new kind of data set came up called ImageNet and which has 20,000 categories and 14 million images. Means 14 million images. You can understand 1. 4 crores or 14 million images instead of 9,000, 30,000. If if us a machine learning was taking 28 days to learn from 30,000 images imagine on 14 million images what it will do. And that was a best paper award, I think.

[00:09:54]So, ImageNet data set came up and in 2012 AlexNet came. AlexNet was implemented on GPU and it shows the error from 26% to 15% 141% improvement and then all the focus went to using the GPUs and that you can see the rise of Nvidia. Nvidia was not even a hundred billion dollar company. Now it is the four trillion dollar company.

[00:10:23]Because we are able to have a deep learning implemented on a GPU and it can be trained on large dataset and it's practically changed the modern world.

[00:10:36]Currently what you see AI advancement in everything it is this is the you know Eureka moment or the acceleration moment for modern technology and this massive parallel computers in this possible and after that people have started combining multiple GPUs together. Initially this was done on one single GPU but now you can train using 40,000, 50,000, 100,000 GPUs.

[00:11:07]So not one GPU with 3,000, 4,000 processors GPU cores you have 100,000 GPUs with 4,000 each so 100,000 into 4,000 you can imagine you have millions of core running together.

[00:11:26]So it AlexNet given the whole focus and since 2012 what currently we see everything is AI and AI and AI and industry started adopting it. We started having very good system where machine learning etc. can come up. So then came the 2020 and in 2017-18 this transformer came up and that revolutionized the training process and then the in 2014 the generative GAN came up and then GAN was being integrated with transformer and then we see the GPT started giving amazing performance and we can see the large model.

[00:12:14]So in AlexNet it was still in millions of parameters but now we have moved to the trillions of parameter. Now modern GPT models are having trillions of parameter. They are being trained on trillion terabytes of data.

[00:12:33]Not megabyte of data. Not a So megabyte to gigabyte to terabyte and now petabytes of data they are getting trained.

[00:12:42]And since it has started impacting industry wider adoption and it has a very high capability and now society started getting impacted.

[00:12:52]Then and this as you know that this models are all becoming black box and then industries are playing uh people are having proprietary data, proprietary knowledge so it's becoming more opaque. So need of governance came up and then AI regulation came up uh last year. So I have a separate course on AI regulation and how to make your system more uh closer to AI act. So since 2020 the you know the it's becoming more accessible and easy to make a system. Like nowadays over a weekend a developer can make an AI system. It became so quick.

[00:13:33]So that also adding that need uh trust, transparency and oversight is needed.

[00:13:39]So in my book uh I have given more detail, more first wave, second wave and we are currently in the fourth wave of AI and more visual uh things are there. But as you saw that AI started as a white box system but now because of the AI became more capable and other thing we moved from rule based system to learn from the data system and large dataset and large model it became a black box model and that's how where we are at the black box problem. We don't know uh why AI is giving the a system and as you know that mixing deep learning more uh black box like because it has a many layers.

[00:14:26]So many million billions of parameters, trillions of parameter, there are many non-linear operations are there.

[00:14:32]There are high dimension of data.

[00:14:34]Earlier we were working with the time series data but now we have a uh more than millions dimension of data means not XYZ. There are XYZ and millions more dimensions of data AI model can work with.

[00:14:50]Learned feature was there but now it is learning itself with human cannot understand.

[00:14:58]And that is making uh deep learning more black box like and since it is a black box kind of thing so it is making risky like it has been used in a health care.

[00:15:10]So the mistake in diagnosis can bring more more problem.

[00:15:15]Autonomous car if it does mistake in some kind of driving mistake then it is a risky because we don't know why it is doing certain kind of You you applied for loan, your loan application is rejected. You applied for governmental any kind of thing you have been rejected and that is removing rights from you so that is risky.

[00:15:37]You applied for job application earlier humans was looking through the CV. Now AI is getting used and AI has bias and is rejecting qualified candidate also.

[00:15:49]The black box uh law enforcement needed. It is getting used in education.

[00:15:56]So that makes uh having a black box kind of model making this AI risky and that's where we are learning this course that how to make it more interpretable.

[00:16:08]So in next lecture we will learn what are the challenges of interpretability and how uh how we can manage between performance and interpretability aspect of it.

[00:16:20]So I hope you learn this from the AI history in a very briefly that we started with a white box. Now in midway we went to the gray box. Now we we are in a black box. Now we have a task to make it interpretable more closer to be a