Explainable AI Technical Whiteboard Series: SHAP

0:06 today we look at explainable Ai and a

0:08 technique called shap which is short for

0:11 shly additive

0:13 explanations AI models don't natively

0:15 explain themselves so data science teams

0:18 must rely on tools and techniques to

0:20 help users understand how the AI Works

0:23 in this video we explore shap one such

0:26 technique used to build trust and

0:28 confidence shap makes machine learning

0:30 models transparent by providing detailed

0:33 visual explanation of the model's

0:35 predictions and an equal initial waiting

0:37 for all model parameters with

0:39 interpretable insights of machine

0:41 learning model output shap helps

0:43 operators to understand and Trust what

0:45 they're using this leads to better

0:48 acceptance and integration of machine

0:49 learning Solutions in various

0:52 applications shapley values were

0:54 introduced in 1951 by Lloyd chapley

0:58 using Concepts from Cooperative game

1:00 Theory it assigns an importance value

1:02 for features or players for a particular

1:04 prediction all features are treated

1:06 equally as input to the

1:08 calculation in a game of network latency

1:10 troubleshooting the players are features

1:12 like bandwidth Wi-Fi radio utilization

1:15 number of clients on an access point use

1:17 of vpns Etc shap is used to assign a

1:21 value to each feature that represents

1:23 how much that feature contributed to the

1:25 final prediction based on every

1:27 combination of possible features by

1:29 looking at all combinations all features

1:31 are fairly represented and features that

1:33 are dependent or synergistic are modeled

1:36 for a given prediction shap values

1:38 explain the impact of each feature

1:40 included in the model which provides

1:42 transparency waiting and The elusive

1:44 answers to the Y

1:46 questions shap values can be visualized

1:49 in various ways such as bar charts for

1:51 individual predictions or summary plots

1:54 for overall feature importance shap

1:56 provides a detailed quantifiable and

1:59 understandable way to explain how each

2:01 feature in a data set influences a

2:03 model's predictions shap values are

2:05 based on data and are unbiased we can

2:08 validate and gain confidence about shap

2:10 values when it reinforces Network domain

2:13 knowledge for example a high Wii radio

2:16 utilization is a clear sign of a Wi-Fi

2:19 network reaching capacity if the shap

2:21 value is large for the radio utilization

2:23 feature when the AP has high radio usage

2:26 we know the model has correctly learned

2:28 the system this makes it a powerful tool

2:31 for enhancing transparency and

2:32 explainability in ai explainable ai is

2:36 closely related to

2:38 interpretability interpretability gives

2:40 context to the meaning of an AI model's

2:42 output many platforms already interpret

2:45 outputs but few of them provide a

2:47 measure of explanation behind those

2:49 interpretations at Juniper Networks shap

2:52 helps to explain video conferencing

2:54 latency predictions for zoom and

2:56 Microsoft teams Juniper employs deep

2:58 learning models that are trained on

3:00 incredibly large data sets and shap

3:02 helps to explain the model's predictions

3:04 so operators can rapidly troubleshoot

3:06 and prioritize their efforts this

3:09 ensures not just exceptional user

3:11 experiences but also the best operator

3:13 experience at Juniper our AI native

3:16 design philosophy focuses on making

3:18 every connection count