wshine ai hw4 - TheEvergreenStateCollege/upper-division-cs-23-24 GitHub Wiki
cost function to find the global minimum
it's stochastic because there is randomness is implemented and is desireable.
in order to ensure we explore all possible avenues to reach a global
minimum
as opposed to a local
minimum.
why does SGD converge?
weights and bias are the parameters to the cost function that we control.
Question 0:
a. sigmoid b. weights from the previous layer c. activations from a previous layer d. bias/threshold
Question 1:
3.321
Question 2:
increase the bias for the digit 2 and decrease the biases for other neurons.
Question 3:
neurons that are active, should have the most increase in weight towards the expected output. if neurons a & c are active, and the expected output is 2. the weights between neurons (a, 2) and (c, 2 ) should be increased the most.
Quesiton 4:
changes to the weights requested by each training data
Question 5:
a drunk stumbling down a hill but taking quick steps
Question 6:
12
Question 1:
784 * 100 + 100 * 10
3 for the smaller network
Question 2:
A: activation from the previous layer B: Sigmoid C: Bias of the current layer D: Weight from the previous layer to this neuron E: activations of the current layer
Question 3:
2, the bottom half is the same as the first 3, the second diagram extends back into the network showing a layer l - 2 4, it can be extended further, but do we call the first layer L in this case? 6, I think this is correct as well but not entirely clear what a partial derivative is
I think these are all similar cases where technology is being used to impersonate human beings. I dont even know what kind of safeguards could be used to protect against this. but I think voice cloning in general shouldn't be ok unless given consent, and the intended use is not widely available. Even if the intent is not malicious cloning a public figures voice and putting it out there could be misinterpreted or misunderstood.
I think people should have the right to clone their own voice sure. Not sure what the benefit would be. Maybe adding voice over to a long video recording and only having to record yourself for a fraction of it's length.
I'm not sure what it would look like to simulate someones personality. But even if it could be done, I don't think it should ok unless given written consent from the pserons before they passed.
Historical re-enactments or fictional movies depicting real people provides a re-telling of an event, but not an interaction with the people from the past. I think recreating a personality from a person of the past in order to interact with it has totally different implications.