Bob Friday Talks - Coding Minds: Crafting Intelligence in Silicon

0:00 hi I'm rahu and welcome to another

0:02 episode of Bob Friday talks AI is often

0:05 regarded as a blackbox or magic

0:08 demystifying AI can be done by

0:10 explaining how AI is built and developed

0:13 today I'm joined by Bob Friday chief AI

0:15 officer here at Juniper today Bob will

0:18 be walking us through how to train your

0:20 AI thanks so much for being here Bob

0:22 thanks for having me let's Dive Right

0:24 into some fundamentals so what is a data

0:27 scientist and what's their role you you

0:30 know what I always tell people data

0:32 scientists are like unicorns no you're

0:34 looking for someone who really

0:36 understands statistics and math and it

0:38 can also write production software these

0:41 are not easy people to find if you look

0:44 what we missed the first data scientist

0:46 I hired was actually a particle

0:48 physicist someone working on physics

0:50 experiments and looking for signals in

0:52 these big physics experiments very

0:54 similar to what we're doing in

0:56 networking trying to find a signal and a

0:58 lot of noise another great example is a

1:01 data scientist who is actually working

1:03 on modeling protein molecules and amino

1:06 acids and he was doing that up to the

1:08 point until Google deep mine actually

1:09 solved that problem with deep learning

1:12 you know so these are people who

1:13 actually understand hardcore math

1:16 statistics but can also write great

1:18 production software these are actually

1:20 very hard people to find so how does how

1:22 does protein amino acids actually relate

1:24 to the building of AI because I've heard

1:26 this come up a few times in my

1:27 conversations with other people and it's

1:29 never been been quite clear to me like

1:31 how does those two things link well you

1:33 think about the protein problem right

1:34 you know they've got all these amino

1:35 acids they're trying to predict the

1:37 properties of proteins based on how

1:40 amino acids join there's billions of

1:42 different combinations here uh before

1:44 deep learning they're actually trying to

1:46 solve this problem with simulations you

1:48 know they're actually going to linear

1:50 accelerators to get x-rays to actually

1:52 look at these protein molecules what

1:54 Google Deep Mind did was actually be

1:56 able to train these deep learning models

1:58 now to predict the prop properties of a

2:00 protein molecule knowing what how the

2:03 amino acids join so that's a great

2:05 example where deep learning is actually

2:06 making a difference in the medical space

2:09 so tell us how do you train your AI yeah

2:12 you know when you look at AI models they

2:14 come in different flavors you know the

2:16 basic ones start with what they call

2:18 supervised training this is where you

2:20 have labeled data where you know the

2:22 answer you know and this is what open Ai

2:24 and chat PT is doing when they train

2:26 chat GPT they're taking tons of data off

2:29 the internet to predict the next word

2:31 and a sequence of words you know what

2:33 we're doing in networking is like with

2:35 zoom data we're taking Zoom data

2:37 combined with network feature datas and

2:39 we're building models now that to can

2:41 predict Zoom latency and user experience

2:44 so that's is an example of supervised

2:46 training self-supervised training is

2:49 where you look back on existing data

2:51 like anomaly detection you know where

2:53 we're looking back for the last week

2:55 months to actually predict the next 10

2:57 minutes or hours Based on data we we've

3:00 seen in the last past and that's what we

3:02 call supervised training and then we get

3:04 into reinforcement learning and this is

3:07 where we're basically training models

3:08 based on a a WS you know where we're

3:11 basically looking to optimize the user

3:13 experience on the wireless network and

3:15 we're creating a symmetrical trying to

3:18 optimize the Symmetry between the user

3:20 and the network so that's an example of

3:22 supervis self-supervised and

3:25 reinforcement learning that we're

3:26 working on how long does it take to

3:28 train the AI yeah you know training

3:30 models varies quite a bit you know you

3:32 look what they're doing at open Ai and

3:34 chat GPT that's millions of dollars

3:37 weeks and months worth of training on

3:38 these big large language models if you

3:41 look what we're doing here at Juniper

3:43 Miss really we're probably training on

3:45 the order of seconds and minutes you

3:47 know whether we're doing a big

3:48 supervised training model which may take

3:50 minutes if we're looking at all the data

3:52 across a universe of networks or we're

3:55 doing some training on a site level

3:58 where we're looking back for the last

3:59 week or months right so yeah training

4:02 can range anywhere from seconds to

4:04 minutes to weeks depending on how big

4:06 these models are your training so the

4:09 term I've heard a lot in conversation

4:10 come up is AI inference can you tell us

4:12 a little bit about what that is yeah

4:14 when they say AI inference what they're

4:16 really talking about is actually running

4:17 a model in production and getting an

4:19 answer you know when we train these

4:22 models for like LM right we may actually

4:26 spend more time and expense training the

4:28 model than actually running an operation

4:30 you know when you look at our llm Marva

4:33 chatbot you know we may have to train it

4:36 on hundreds of thousands of questions

4:39 and answers whereas when we actually put

4:41 it into our actual Network we may be

4:44 dealing with only hundreds of questions

4:46 or thousands of questions per week so

4:48 you think about training and inference

4:51 inference refers to actually running the

4:53 model in a production environment how do

4:55 you measure the effectiveness of an AI

4:57 model yeah I think there's a couple

4:59 answers to this question you know when a

5:01 data scientist trains like a supervised

5:03 model they'll typically take a training

5:06 data set they'll take 70% of the data to

5:10 actually train the model then they may

5:11 save 20 or 30% to actually measure the

5:14 effectiveness and how accurate that

5:16 model is so that's one technique or

5:19 definition of Effectiveness from a data

5:21 science point of view from a customer

5:23 facing point of view here at Juniper we

5:26 actually run our models and our customer

5:28 support team and we do something called

5:30 Marvis efficacy where we basically have

5:33 actual users grade our models and tell

5:35 us how often the model is getting the

5:37 right answer so you can think of

5:39 effectiveness at the data science level

5:41 when you're training your models and

5:43 then there's actually the customer

5:44 facing Effectiveness how effective it is

5:46 with your actual customers or the

5:47 problem you're working on thanks so much

5:49 for your thoughts today Bob we really

5:51 appreciate your time and thank you for

5:52 listening and until next

5:58 time