Hidden Markov Models 11: the Viterbi algorithm

[@kevinigwe3143]

Thank you so much. Following the previous lectures (...08,09, 10) up till 11 cleared all my doubts and confusion. The best I have been waiting for!

[@djp3]

You are most welcome

[@mrlh69]

I'm from Peru. I'm writing a thesis about speech recognition using HMM. This was really helpful, thank you!!

[@taggebagge]

Best of luck with your thesis and hope that you stay safe from the corona man.

[@bengonoobiang6633]

it's just impressive. I'm a PhD student working on a speech recognition system. And understanding the algorithmic foundations and modeling of the underlying problem is not very easy with neurological methods. This HMM course has helped me to understand the problems and solutions of sequence processing problems in detail.

[@djp3]

Thanks! It's a cool algorithm as long as your state space isn't really huge.

[@codediporpal]

This is great stuff. A bit repetitive, but that's a good thing for a complicated subject like this. The "motivations" videos were gold, since too often subject are taught without teaching why it's important.

[@goonerrn]

he looks like the child of Steve Carell and Ryan Reynolds.. and great video

[@djp3]

It's could be worse

[@wandering_star365]

Maybe it's because I'm watching this on my phone lol but I totally see it too! I was like woah, who's Reynolds trying to fool with those glasses🤭

[@RandomTrash-ey5di]

@@djp3 I never thought I'd be learning about HMMs from deadpool

[@theuser969]

Thanks for providing this well explained video on this topic.

[@djp3]

Glad I could help!

[@amitotc]

Very well explained, really enjoyed watching the seires of videos on HMM. Thank you so much :)

[@djibrildassebe5055]

Very nice series of course thank you Prof. Patterson

[@okh6201]

Thank you for the great explanation Sir, but i have one question. What if some states's emission probabilities contain a probability to stay silent will viterbi's algorithm still work?

[@djp3]

It can, but you need to add a place holder emission symbol of “silent” so that the probability of emission still adds up to 1.

[@okh6201]

​@@djp3 Thanks for the reply Sir, so when we calculate Alpha or Beta do we have to always assume there is a silent emission and if so, how do we factor it into the observation sequence?

[@djp3]

@@okh6201 thats a modeling question. It depends on the phenomenon you are trying to represent. If your observation sequence has “silent” observations then explicitly treating that as a fake observation is one way to manage it

[@okh6201]

@@djp3 I see I will try that then, thank you 😁

[@Martin-iy8gd]

Another great video !

[@ryanjones1704]

Why are we seeking to maximize P( Q, O | lambda) and not P(Q | O, lambda)? It's stated at ~10 minutes that the two are equivalent, so I'm sure there's some bayesian principal I'm missing. But even at ~12 minutes, it's stated "that accounts for the first t observations," which seems like "given the first t observations." Can somebody point me in the right direction? In any case: awesome series! Much appreciated!

[@haipingwang7075]

To my understanding , Q is the hidden part, it can be more Q s jointed with one O , the joint probability is good enough for us to solve the backward and forward problems. No need to know how , like to figure out P(Q|O,lambda), it is a trick

[@akashgopal6896]

This was a fantastic explanation and really cleared things up for me, but I have a question. In the scenario where all states have both the probability to emit or remain silent, how do you factor that into the calculations? As the observation sequence does not indicate how many times the states have been silent, would you still be able to use Viterbi's algorithm to find the most probable sequence of states?

[@djp3]

You can add a fake emission which is “silent”.

[@pierrestemmettmusic]

Great video! Very helpful, thank you

[@Oleg_Litvinov]

Awesome lecturer!

[@fgfanta]

10:42 It makes me think that {O_1, ..., O_T} should be on the right side of the |, with Lambda. Because the observations are given.

[@ycdantywong]

Since both alpha_t(i) and beta_t(i) are calculated using the transition matrix which already includes information regarding which transitions are possible/impossible; I don't understand how argmax[gamma_t(i)] could end up with a state i that came from an impossible transition ?

[@youssefdirani]

2:45 it is regrdless

[@djp3]

gamma aggregates over all paths. So it is just saying that it is possible by some path to get to and from i.

[@wafaguendouz8430]

Ryan Reynolds is that you?

[@djp3]

No but I appreciate the vote of confidence!

[@xntumrfo9ivrnwf]

Hello, my head exploded. Please advise.

[@djp3]

band-aids and aspirin

[@jasmeetgujral5665]

Amazing

[@djp3]

Thanks

[@karannchew2534]

3:40 Normalise All possible observation given all possible states.

[@ganeshkshet5631]

thaank you

[@djp3]

You're welcome!

[@beko_61]

thank you ryan reynolds