2:53, 7:54 These small fun moments are what is missing in other NPTEL courses. Thank you sir for taking the extra effort to make us feel like we are part of a learning programme too !
5:15 If he used the notation of -1 for False and +1 for True in the very beginning, why did he create the truth table with 0s and 1s? It only added to the confusion!
At 6:26, the first condition for the XOR should be w0 + w1 w1 < -w0. Why has he written w1 < +w0? I understand that in all previous examples, we have taken the bias term to be negative. But here we are writing the variable and in summation, we don't write negative. So, the summation is Σw_i x_i where i is from 0 to 2. x0 is 1 and w0 has no value assigned to it. Does he want to assign w0 = -w0? Only then it would be -w0 + w1 < 0 => w1 < w0. Some weird shit!
For scenario 1, x1 = 0, x2=0, w1 = -1, w2=-1, w0=2. I don't understand how h1 will fire. Cause -2+(-1*0)+(-1*0) = -2 which is not greater than zero. So, it should not fire. In fact, none of the 4 middle layer perceptrons should fire. Yet, the outcome I see is that h1 is firing. How? What am I missing?
I'm assuming you are confused at the truth table shown at 5:15. The truth table only has input values as false or true depicted as 0, 1 respectively. But when we pass these inputs through the perceptron we consider false input as -1 and true input as +1 (for which hidden layer neurons fire). We can also build the perceptron to take 0,1 as inputs just by changing the bias. It's just a matter of choice. You can also build a perceptron by taking -3 for false and -1 for true by adjusting the biases appropriately. Though I agree, that the instructor should have explained his transition of input values.
You are missing the fact that the professor had mentioned earlier that in this model, we will consider -1 as false input and +1 as true input. So -1*-1 + -1*-1 + -2 = 0 and hence h1 will fire.
At 6:40 condition is w2>=w0, we have w0=-2, right? then if I take -1 means according to the condition it is right. So 1*-1=-1 then neuron will not fire and answer will -1 however we have output 1 in XOR case. I am confused here
Now if we use the perceptron learning algorithm on the output perceptron, the weights are first randomly initialized and updating operations would be W = W+H or W = W-H where H is the vector of hidden layer outputs right? But as per the network only one of the hidden layer perceptron fire for every input so isn't H having all 0s except for one 1. Thus during the update (W = W +/- H) only one of the weights get changed, and this weight corresponds to that firing hidden layer perceptron. So for unseen inputs the perceptron just hasn't updated the weight and they remain their initial random value??
its not always the case, if you did very well on train data and poor on test data then its called overfitting. there might be cases where you do well on both train and test data. you always need validation/test data to tell if you overfit or not.
It is the input values for h1,h2,h3,h4 when they fire (predicts positive). It is slightly confusing since input "0" is considered as False or "-1" and 1 is considered as Positive or +1.
2:53, 7:54 These small fun moments are what is missing in other NPTEL courses. Thank you sir for taking the extra effort to make us feel like we are part of a learning programme too !
This series has been exceptional!!
You deliver high quality content in such an elegant manner
You made it so easy and effortless. Thank you so much
Myself Mechanical Student Can Learnt Well Just Wow Teacher
5:15 If he used the notation of -1 for False and +1 for True in the very beginning, why did he create the truth table with 0s and 1s? It only added to the confusion!
yes i am also confused but later i realised
truth table is made for only 0 and 1
Excellent explanation..Thanks sir!
At 6:26, the first condition for the XOR should be w0 + w1 w1 < -w0. Why has he written w1 < +w0? I understand that in all previous examples, we have taken the bias term to be negative. But here we are writing the variable and in summation, we don't write negative. So, the summation is Σw_i x_i where i is from 0 to 2. x0 is 1 and w0 has no value assigned to it. Does he want to assign w0 = -w0? Only then it would be -w0 + w1 < 0 => w1 < w0. Some weird shit!
Bahut mast smjhate ho sir🙏.
For scenario 1, x1 = 0, x2=0, w1 = -1, w2=-1, w0=2. I don't understand how h1 will fire. Cause -2+(-1*0)+(-1*0) = -2 which is not greater than zero. So, it should not fire. In fact, none of the 4 middle layer perceptrons should fire. Yet, the outcome I see is that h1 is firing. How? What am I missing?
I'm assuming you are confused at the truth table shown at 5:15. The truth table only has input values as false or true depicted as 0, 1 respectively. But when we pass these inputs through the perceptron we consider false input as -1 and true input as +1 (for which hidden layer neurons fire). We can also build the perceptron to take 0,1 as inputs just by changing the bias. It's just a matter of choice. You can also build a perceptron by taking -3 for false and -1 for true by adjusting the biases appropriately. Though I agree, that the instructor should have explained his transition of input values.
You are missing the fact that the professor had mentioned earlier that in this model, we will consider -1 as false input and +1 as true input. So -1*-1 + -1*-1 + -2 = 0 and hence h1 will fire.
sir, I think for 2 input perceptron network bias would be 0,1,1,2 instead of -2 for each perceptron
No carefully hear what he is saying if the input is -1,-1 etc.Then it will make sense.
True=+1 and False=-1 as stated in 1:09. The rest should then make sense.
at 5.33 how is h1 = 1? pls help me out
When he said why do we care about Boolean function my first response was to watch my favourite movie 😂
At 6:40 condition is w2>=w0, we have w0=-2, right? then if I take -1 means according to the condition it is right. So 1*-1=-1 then neuron will not fire and answer will -1 however we have output 1 in XOR case. I am confused here
Because he has taken the wrong convention! The condition should be w0 + w2 >= 0 => w2 >= -w0. Put w0=-2 and you will get w2 >= 2
@@desiquant yes
Sir Salute
what's going on around @6:29
What if the movie goer is irrational? We all are irrational to some extent. How to model that?
Now if we use the perceptron learning algorithm on the output perceptron, the weights are first randomly initialized and updating operations would be
W = W+H or W = W-H where H is the vector of hidden layer outputs right? But as per the network only one of the hidden layer perceptron fire for every input so isn't H having all 0s except for one 1. Thus during the update (W = W +/- H) only one of the weights get changed, and this weight corresponds to that firing hidden layer perceptron. So for unseen inputs the perceptron just hasn't updated the weight and they remain their initial random value??
Sir, but if model correctly classifies all datapoints, then it is an overfit model, how to deal with that, is it explained in later videos???
its not always the case, if you did very well on train data and poor on test data then its called overfitting. there might be cases where you do well on both train and test data. you always need validation/test data to tell if you overfit or not.
While this lecture is very good, it lacks the generalized mathematical equation of an MLP. I was expecting a generalized equation to represent an MLP.
Can anyone please explain me why h1,h2h3 having value (-1,-1),(-1,1),(1,-1) has taken
So that only one of the perceptron triggers for each of the 4 possible input..
It is the input values for h1,h2,h3,h4 when they fire (predicts positive). It is slightly confusing since input "0" is considered as False or "-1" and 1 is considered as Positive or +1.
@ Sudhir Kumar those are the weights of the inputs x1 & x2 coming to the corresponding neuron
at 5.33 how is h1 = 1? pls help me out
@prachi shree thanks :)