BERT vs GPT

#machinelearning #shorts #deeplearning #chatgpt #neuralnetwork #datascience

[@darshantank554]

One is for Natural language understanding and another is for Natural language generation

[@Ram_jagat]

true

[@borntodoit8744]

Remember it as INPUT>MODEL >OUTPUT MODEL INPUT (NLU) - txt recognition, vision recognition/image/movie, sound recognition/voice, MODEL OUTPUT (NLG) -txt generation, image/movie generation, sound/voice generation + tool integration MODEL PROCESSING -Basic (classification summeration extraction) advanced (reasoning, planning, orchestration)

[@VarunTulsian]

this is very useful. Just wanted to add that the gpt decoder doesn't have the cross attention in the transformer block.

[@Tech_kenya]

What is cross attention

[@methylphosphatePOET]

@@Tech_kenya It's when word vectors reference other word vectors as opposed to just referencing themselves.

[@imran7TW]

@@methylphosphatePOET so kinda the opposite of self attention?

[@cs101-qm2ud]

Wonderfully put.

[@CodeEmporium]

Thanks a lot :)

[@maninzn]

Great explanation. For eg, if I have to read all the client emails and understand their requirements and auto create tasks based on that prediction, which model should I go for? BERT or GPT?

[@JillRhoads]

I hadnt known that BERT was an acronym and had been wondering why the Sweden LLM was called Bert. I wonder if this is why. Thanks for the info!

[@contactdi8426]

Can you please explain about their training process?

[@vladislavkorecky618]

What if I stack both encoders and decoders? Do I get some BERTGPT hybrid?

[@davronsherbaev9133]

there is also Whiser model, that has similar facebook BART decoder part, but has audio decoder.

[@Dr_Larken]

Bert also Drives a trans am! Jokes aside I do appreciate your videos!

[@nicholaszustak6299]

So BERT doesn’t have a decoder? Did I misunderstand

[@saimadhaviyalla5682]

transformer models are usually parallelly run right?

[@eugeneku3239]

Not when it's decoding. No.

[@hubgit9556]

good

[@usama57926]

I love you ❤

[@obieda_ananbeh]

Awesome 👏

[@CodeEmporium]

Thanks So much