WEBVTT

1
00:00:00.160 --> 00:00:03.040
<v Speaker 1>Welcome to the deep dive. Today, we're going to be

2
00:00:03.080 --> 00:00:07.400
<v Speaker 1>diving into the world of automated machine learning exciting specifically

3
00:00:07.679 --> 00:00:10.759
<v Speaker 1>on Microsoft's Azure platform. You know, think of it as

4
00:00:10.839 --> 00:00:16.879
<v Speaker 1>AI simplified. And to help us navigate this really complex world,

5
00:00:17.280 --> 00:00:20.399
<v Speaker 1>we have Dennis Sawyer's Practical Guide to Auto mL on Azure.

6
00:00:20.600 --> 00:00:22.320
<v Speaker 2>Yes, this book is great.

7
00:00:22.440 --> 00:00:25.679
<v Speaker 1>So who wants to start today is to give you

8
00:00:25.719 --> 00:00:29.480
<v Speaker 1>a solid understanding of what auto mL is, what its

9
00:00:29.559 --> 00:00:31.920
<v Speaker 1>benefits are, and then how you can actually use it

10
00:00:31.920 --> 00:00:32.399
<v Speaker 1>on Ashure.

11
00:00:32.479 --> 00:00:34.000
<v Speaker 2>Awesome, No, that's good.

12
00:00:34.000 --> 00:00:37.399
<v Speaker 1>The book starts off with a pretty startling statistic. Eighty

13
00:00:37.520 --> 00:00:40.359
<v Speaker 1>seven percent of AI projects fail.

14
00:00:41.200 --> 00:00:42.560
<v Speaker 2>Wow, that's a lot.

15
00:00:42.759 --> 00:00:45.799
<v Speaker 1>That's a lot of unrealized potential. Yeah, that is, especially

16
00:00:45.799 --> 00:00:48.240
<v Speaker 1>considering the resources that are being poured into this field.

17
00:00:48.399 --> 00:00:50.359
<v Speaker 2>It really is, and it's interesting.

18
00:00:50.439 --> 00:00:52.520
<v Speaker 1>We'll see even more because the book.

19
00:00:52.320 --> 00:00:55.240
<v Speaker 2>Digs into why interesting projects fail.

20
00:00:55.359 --> 00:00:58.960
<v Speaker 1>Book digs into why it's not always they fail, and

21
00:00:59.000 --> 00:01:02.280
<v Speaker 1>it's not all dramatic implosion.

22
00:01:01.840 --> 00:01:04.640
<v Speaker 2>Right, dramatic inclusions a very slow bird. It's like a

23
00:01:04.680 --> 00:01:08.799
<v Speaker 2>slow leak in a spaceship, just slowly draining resources and.

24
00:01:08.840 --> 00:01:13.200
<v Speaker 1>Hope exactly, and as the book details, like the traditional

25
00:01:13.239 --> 00:01:17.439
<v Speaker 1>machine learning workflow is incredibly time consuming incomplex so much.

26
00:01:17.400 --> 00:01:18.879
<v Speaker 2>More than just building models.

27
00:01:18.920 --> 00:01:22.040
<v Speaker 1>It's not just building the model, it's the data cleaning,

28
00:01:22.400 --> 00:01:27.840
<v Speaker 1>the feature engineering, and then the dreaded hyper parameter tuning. Oh,

29
00:01:27.879 --> 00:01:30.159
<v Speaker 1>tell me about it, which can be a real headache.

30
00:01:30.200 --> 00:01:33.680
<v Speaker 2>It's like solving a Rubik's cube, yes, blindfolded while riding

31
00:01:33.719 --> 00:01:34.359
<v Speaker 2>a unicycle.

32
00:01:34.480 --> 00:01:35.159
<v Speaker 1>Absolutely.

33
00:01:35.359 --> 00:01:37.840
<v Speaker 2>And then, as the book points out, many data scientists

34
00:01:37.840 --> 00:01:40.640
<v Speaker 2>aren't actually trained and deploying these models in the real world.

35
00:01:40.719 --> 00:01:40.879
<v Speaker 1>Oh.

36
00:01:40.959 --> 00:01:44.719
<v Speaker 2>Interesting, So they end up with these really fragile okay,

37
00:01:44.959 --> 00:01:48.760
<v Speaker 2>hobbled together solutions that are just waiting to.

38
00:01:48.760 --> 00:01:50.719
<v Speaker 1>Crumble, right, And this is where and.

39
00:01:50.680 --> 00:01:53.519
<v Speaker 2>So this is where AutoML comes in. Gota mL comes in,

40
00:01:53.719 --> 00:01:57.599
<v Speaker 2>offering a potential solution to this ROI dilemma that exists

41
00:01:57.599 --> 00:01:58.640
<v Speaker 2>within data science.

42
00:01:58.840 --> 00:02:02.359
<v Speaker 1>Yeah that ROI prom Okay, so elevator pitch time. What

43
00:02:02.519 --> 00:02:05.120
<v Speaker 1>is auto mL and how does it solve this problem?

44
00:02:05.439 --> 00:02:08.599
<v Speaker 2>Okay? Imagine having an AI assistant that takes care of

45
00:02:08.680 --> 00:02:11.800
<v Speaker 2>all of the heavy lifting in that machine learning process. Ooh,

46
00:02:11.879 --> 00:02:16.159
<v Speaker 2>I like that AutoML can train multiple models simultaneously using

47
00:02:16.159 --> 00:02:20.319
<v Speaker 2>the latest algorithms. Got it handles feature engineering, nice fine

48
00:02:20.319 --> 00:02:24.400
<v Speaker 2>tunes those pesky hyper parameters and even offers built and

49
00:02:24.479 --> 00:02:29.560
<v Speaker 2>explainability features which are really crucial for building trust and transparency.

50
00:02:29.680 --> 00:02:31.319
<v Speaker 1>Wait, so does that mean data scientists are going to

51
00:02:31.319 --> 00:02:31.879
<v Speaker 1>be out of a job.

52
00:02:32.120 --> 00:02:36.439
<v Speaker 2>Not at all. Auto mL actually empowers data scientists. It

53
00:02:36.520 --> 00:02:39.599
<v Speaker 2>frees them up to focus on the higher level aspects

54
00:02:39.639 --> 00:02:44.479
<v Speaker 2>of a like problem definition, strategy and interpreting those results.

55
00:02:44.599 --> 00:02:48.000
<v Speaker 1>So it's more of a collaboration exactly between human expertise

56
00:02:48.599 --> 00:02:53.599
<v Speaker 1>and AI efficiency. Now let's talk Azure. Why is Azure

57
00:02:54.280 --> 00:02:56.759
<v Speaker 1>so central to this auto mL story?

58
00:02:56.960 --> 00:03:01.960
<v Speaker 2>So the book focuses on Azure because Microsoft's cloud platform

59
00:03:02.479 --> 00:03:06.240
<v Speaker 2>offers a comprehensive suite of tools for auto mL, and

60
00:03:06.240 --> 00:03:10.439
<v Speaker 2>that's all through the Azure Machine Learning Service AMLS as

61
00:03:10.439 --> 00:03:10.879
<v Speaker 2>it's called.

62
00:03:10.960 --> 00:03:13.560
<v Speaker 1>So it's like our auto mL playground exactly.

63
00:03:13.680 --> 00:03:16.479
<v Speaker 2>And the book actually guides us I like it through

64
00:03:16.520 --> 00:03:19.719
<v Speaker 2>setting up an AMLS workspace, and it gets us familiar

65
00:03:19.719 --> 00:03:21.719
<v Speaker 2>with all the features we have at our disposal.

66
00:03:21.960 --> 00:03:22.240
<v Speaker 1>Nice.

67
00:03:22.479 --> 00:03:26.319
<v Speaker 2>One key concept is compute compute, which is essentially the

68
00:03:26.360 --> 00:03:28.560
<v Speaker 2>engine towering your auto mL jobs.

69
00:03:28.599 --> 00:03:32.360
<v Speaker 1>So compute is like the horsepower behind our AI engine.

70
00:03:32.400 --> 00:03:36.159
<v Speaker 2>That's a great analogy, and the book distinguishes between compute

71
00:03:36.159 --> 00:03:40.199
<v Speaker 2>instances which are for simpler tasks, and then compute clusters

72
00:03:40.199 --> 00:03:43.840
<v Speaker 2>for those more demanding god the resource intensive workloads.

73
00:03:43.479 --> 00:03:46.039
<v Speaker 1>So when you need to kick it into high gear exactly.

74
00:03:46.159 --> 00:03:47.479
<v Speaker 2>And here's where it gets interesting.

75
00:03:47.680 --> 00:03:49.360
<v Speaker 1>Okay, these compute.

76
00:03:48.960 --> 00:03:53.599
<v Speaker 2>Clusters can autoscale. Ooh, I like that, meaning you only

77
00:03:53.639 --> 00:03:57.280
<v Speaker 2>pay for what you use budget conscious, which is a

78
00:03:57.360 --> 00:04:00.360
<v Speaker 2>major win for budget conscious projects.

79
00:04:00.400 --> 00:04:04.520
<v Speaker 1>Smart. So we have our workspace, our powerful compute engine.

80
00:04:05.080 --> 00:04:07.240
<v Speaker 1>What about the fuel, the data?

81
00:04:07.520 --> 00:04:12.960
<v Speaker 2>Yes, data is the lifeblood of any AI project. MLS

82
00:04:13.280 --> 00:04:17.120
<v Speaker 2>actually works with data sets, data sets which act as

83
00:04:17.160 --> 00:04:20.040
<v Speaker 2>pointers to your data sources, whether they reside in a

84
00:04:20.120 --> 00:04:23.879
<v Speaker 2>storage account, got it or a SQL database. And Azure

85
00:04:24.120 --> 00:04:28.079
<v Speaker 2>even provides open data sets for experimentation.

86
00:04:27.480 --> 00:04:29.040
<v Speaker 1>Free data to play with. I like that.

87
00:04:29.240 --> 00:04:32.439
<v Speaker 2>Yeah, so like the diabetes data set it's mentioned in

88
00:04:32.439 --> 00:04:32.759
<v Speaker 2>the book.

89
00:04:32.839 --> 00:04:35.639
<v Speaker 1>So I give auto mL my data and it magically

90
00:04:35.680 --> 00:04:36.800
<v Speaker 1>spins out a perfect model.

91
00:04:37.120 --> 00:04:40.120
<v Speaker 2>Not quite magic, but AutoML does do a lot of

92
00:04:40.160 --> 00:04:42.720
<v Speaker 2>the heavy lifting kind of scenes. So the book explains

93
00:04:43.120 --> 00:04:45.759
<v Speaker 2>how auto mL starts with what are called data.

94
00:04:45.480 --> 00:04:48.360
<v Speaker 1>Guardrail data guardrails.

95
00:04:47.720 --> 00:04:52.319
<v Speaker 2>Which are automated data quality checks that can identify potential

96
00:04:52.360 --> 00:04:53.319
<v Speaker 2>issues right out.

97
00:04:53.199 --> 00:04:57.240
<v Speaker 1>Of the gate. So bad data equals bad outcomes exactly, garbaging,

98
00:04:57.279 --> 00:04:57.879
<v Speaker 1>garbage out.

99
00:04:58.000 --> 00:05:01.519
<v Speaker 2>So once the data passes inspection, what's next? Auto mL

100
00:05:01.680 --> 00:05:06.199
<v Speaker 2>kicks into high gear with intelligent feature engineering okay, So

101
00:05:06.279 --> 00:05:11.279
<v Speaker 2>it handles tasks like dealing with missing values, transforming categorical

102
00:05:11.360 --> 00:05:15.759
<v Speaker 2>variable for example, using one hot encoding okay, which the

103
00:05:15.800 --> 00:05:20.040
<v Speaker 2>book explains really well, and it even generates new features wow.

104
00:05:20.279 --> 00:05:23.279
<v Speaker 2>And it tailors all of this to the specific algorithms

105
00:05:23.279 --> 00:05:24.079
<v Speaker 2>that it will be using.

106
00:05:24.399 --> 00:05:28.000
<v Speaker 1>So it's not just blindly throwing algorithms at the data,

107
00:05:28.519 --> 00:05:32.399
<v Speaker 1>it's strategically preparing the data for each algorithm exactly. That's

108
00:05:32.399 --> 00:05:36.879
<v Speaker 1>pretty impressive. And speaking of algorithms, seeking of algorithms, AutoML

109
00:05:36.920 --> 00:05:38.680
<v Speaker 1>has a whole arsenal at its disposal.

110
00:05:38.759 --> 00:05:40.160
<v Speaker 2>It has a whole arsenal.

111
00:05:40.040 --> 00:05:44.399
<v Speaker 1>Ranging from classic regression and classification models to more advanced

112
00:05:44.439 --> 00:05:47.600
<v Speaker 1>techniques like gradient boasting and deep learning. So how does

113
00:05:47.639 --> 00:05:50.639
<v Speaker 1>AutoML decide which algorithm to use? Does it just pick

114
00:05:50.680 --> 00:05:51.519
<v Speaker 1>one at random?

115
00:05:51.600 --> 00:05:56.040
<v Speaker 2>It does not. It systematically tests multiple algorithms okay, parallel,

116
00:05:56.079 --> 00:05:58.800
<v Speaker 2>got it. And while it's doing so, it also fine

117
00:05:58.839 --> 00:06:02.240
<v Speaker 2>tunes the hyper parameters for each one, so it's searching

118
00:06:02.319 --> 00:06:03.839
<v Speaker 2>for the best performing combination.

119
00:06:04.199 --> 00:06:08.040
<v Speaker 1>It's like having a team of data scientists working tirelessly

120
00:06:08.120 --> 00:06:12.680
<v Speaker 1>behind the scenes, exact optimizing every step of the process. Yes.

121
00:06:13.160 --> 00:06:15.040
<v Speaker 1>So does this mean I can just sit back and

122
00:06:15.079 --> 00:06:16.759
<v Speaker 1>wait for the perfect model to appear?

123
00:06:17.120 --> 00:06:20.800
<v Speaker 2>Well, not quite. Ok The book emphasizes that while auto

124
00:06:20.920 --> 00:06:23.600
<v Speaker 2>mL is a powerful tool, it's not a magic bullet.

125
00:06:23.639 --> 00:06:27.040
<v Speaker 1>Okay, fair enough. So where does human judgment come in?

126
00:06:27.519 --> 00:06:31.160
<v Speaker 2>You play a crucial role at it in defining the problem,

127
00:06:31.319 --> 00:06:35.360
<v Speaker 2>choosing the right evaluation metrics okay, and most importantly interpreting

128
00:06:35.399 --> 00:06:35.959
<v Speaker 2>the results.

129
00:06:36.079 --> 00:06:39.439
<v Speaker 1>So there's still that need for that human touch, that

130
00:06:39.600 --> 00:06:42.319
<v Speaker 1>understanding of the nuances of the problem that we're trying

131
00:06:42.319 --> 00:06:45.399
<v Speaker 1>to solve. Absolutely, and speaking of getting our hands dirty, Yes,

132
00:06:45.759 --> 00:06:50.240
<v Speaker 1>the book actually dives into building models using Azure Machine

133
00:06:50.279 --> 00:06:53.800
<v Speaker 1>Learning Studio okay, which is a visual interface for working

134
00:06:53.800 --> 00:06:57.120
<v Speaker 1>with mls exactly. It's designed to be user friendly even

135
00:06:57.160 --> 00:06:58.879
<v Speaker 1>if you're not a coding guru. Yes.

136
00:06:59.079 --> 00:07:02.560
<v Speaker 2>And the book walks us through creating our first classification

137
00:07:02.759 --> 00:07:07.079
<v Speaker 2>model using the a classic Titanic passenger data ah.

138
00:07:07.319 --> 00:07:11.560
<v Speaker 1>The Titanic data set a data science right of passage.

139
00:07:11.120 --> 00:07:15.319
<v Speaker 2>Exactly, you know, predicting who survived and who didn't based

140
00:07:15.360 --> 00:07:19.720
<v Speaker 2>on factors like age, gender, ticket class face, and the

141
00:07:19.759 --> 00:07:24.199
<v Speaker 2>book makes it surprisingly straightforward. You upload the data, tell

142
00:07:24.240 --> 00:07:27.079
<v Speaker 2>auto mL you're doing a classification task, and you let

143
00:07:27.079 --> 00:07:28.000
<v Speaker 2>it work its magic.

144
00:07:28.079 --> 00:07:30.639
<v Speaker 1>And while it's working its magic, we can monitor those

145
00:07:30.720 --> 00:07:34.399
<v Speaker 1>data guardrails exactly see if any potential issues are flagged.

146
00:07:34.759 --> 00:07:37.560
<v Speaker 2>Yeah, it's like having a built in data quality watchdog.

147
00:07:37.959 --> 00:07:38.600
<v Speaker 1>I like that.

148
00:07:38.800 --> 00:07:42.279
<v Speaker 2>Now, once automil finishes training the set of models, it's

149
00:07:42.279 --> 00:07:44.000
<v Speaker 2>time to put on our evaluation hats.

150
00:07:44.079 --> 00:07:45.759
<v Speaker 1>So we've got a bunch of models. Yeah, what do

151
00:07:45.839 --> 00:07:46.360
<v Speaker 1>we do with them?

152
00:07:46.480 --> 00:07:51.600
<v Speaker 2>Auto mL gives us a smorgasboard of metrics accuracy, precision, recall,

153
00:07:51.839 --> 00:07:56.199
<v Speaker 2>and it gives us those helpful confusion matrices to visualize

154
00:07:56.199 --> 00:07:57.720
<v Speaker 2>how each model is performing.

155
00:07:57.920 --> 00:08:01.480
<v Speaker 1>I'll be honest, sometimes those metrics can feel a little overwhelming,

156
00:08:01.800 --> 00:08:03.199
<v Speaker 1>especially if you're new to machine learning.

157
00:08:03.240 --> 00:08:03.600
<v Speaker 2>I hear you.

158
00:08:03.680 --> 00:08:07.079
<v Speaker 1>There's just the book so many numbers. The book clearly

159
00:08:07.120 --> 00:08:10.720
<v Speaker 1>explains each metric, okay, and helps us to understand which

160
00:08:10.759 --> 00:08:14.480
<v Speaker 1>ones are most important for different types of problems. And

161
00:08:14.519 --> 00:08:17.959
<v Speaker 1>remember those explainability features, Yes, we can use them to

162
00:08:18.040 --> 00:08:23.120
<v Speaker 1>understand why a model is making certain predictions. That's crucial,

163
00:08:23.399 --> 00:08:26.040
<v Speaker 1>which is crucial for building trust and transparency.

164
00:08:26.199 --> 00:08:28.920
<v Speaker 2>So it's not just blindly trusting the numbers. It's about

165
00:08:29.000 --> 00:08:30.800
<v Speaker 2>understanding the reasoning behind them.

166
00:08:30.920 --> 00:08:31.480
<v Speaker 1>Absolutely.

167
00:08:32.080 --> 00:08:34.919
<v Speaker 2>Now, let's say we've found a model that looks pretty

168
00:08:34.919 --> 00:08:38.679
<v Speaker 2>promising based on the metrics and the explainability checks. What

169
00:08:38.759 --> 00:08:40.759
<v Speaker 2>happens next, then.

170
00:08:40.679 --> 00:08:44.279
<v Speaker 1>It's time to deploy deployment. We need to make our

171
00:08:44.360 --> 00:08:48.600
<v Speaker 1>model available for use either for a batch processing of

172
00:08:48.679 --> 00:08:52.000
<v Speaker 1>large data sets or for real time predictions.

173
00:08:52.120 --> 00:08:53.879
<v Speaker 2>So we're taking our model out of the lab and

174
00:08:53.919 --> 00:08:55.039
<v Speaker 2>putting it into the real world.

175
00:08:55.159 --> 00:08:55.679
<v Speaker 1>Exactly.

176
00:08:55.759 --> 00:08:58.639
<v Speaker 2>But we've mainly been talking about classification models.

177
00:08:58.759 --> 00:09:01.559
<v Speaker 1>What about problems where you need to predict a numerical

178
00:09:01.639 --> 00:09:03.360
<v Speaker 1>value instead of a category.

179
00:09:03.480 --> 00:09:07.360
<v Speaker 2>The book covers that took it dives into building regression

180
00:09:07.399 --> 00:09:11.440
<v Speaker 2>models using auto mL on Azure. It uses the diabetes

181
00:09:11.559 --> 00:09:15.000
<v Speaker 2>data set to actually predict the progression of the disease

182
00:09:15.240 --> 00:09:16.480
<v Speaker 2>based on different factors.

183
00:09:16.559 --> 00:09:20.360
<v Speaker 1>So regression is about predicting numbers, right, like stock prices

184
00:09:20.480 --> 00:09:24.679
<v Speaker 1>or sales figures, or in this case, the severity of

185
00:09:24.720 --> 00:09:27.720
<v Speaker 1>a medical condition. Exactly. Okay, And while the process is

186
00:09:27.759 --> 00:09:31.759
<v Speaker 1>similar to classification, obviously the metrics and the algorithms used

187
00:09:31.759 --> 00:09:35.080
<v Speaker 1>are different, but AutoML still handles all that heavy lifting

188
00:09:35.200 --> 00:09:37.440
<v Speaker 1>it does. But we need to use the right tool

189
00:09:37.519 --> 00:09:40.360
<v Speaker 1>for the job. Absolutely, I bet there are some tips

190
00:09:40.399 --> 00:09:42.720
<v Speaker 1>and tricks for getting the best performance out of auto

191
00:09:42.879 --> 00:09:44.879
<v Speaker 1>mL for these regression problems.

192
00:09:45.080 --> 00:09:48.120
<v Speaker 2>There are, and the book offers some great insights.

193
00:09:48.200 --> 00:09:49.200
<v Speaker 1>Okay, lay on me.

194
00:09:49.600 --> 00:09:53.840
<v Speaker 2>Sometimes converting a regression problem into a classification problem can

195
00:09:53.879 --> 00:09:55.279
<v Speaker 2>actually improve performance.

196
00:09:55.600 --> 00:09:57.879
<v Speaker 1>Hold on, how do you convert a problem about predicting

197
00:09:57.919 --> 00:10:01.360
<v Speaker 1>a number into a problem about predicting a category?

198
00:10:01.480 --> 00:10:05.000
<v Speaker 2>It's all about binning binning. You're going to divide the

199
00:10:05.120 --> 00:10:09.279
<v Speaker 2>range of possible numerical values into categories or bins. Okay,

200
00:10:09.360 --> 00:10:12.600
<v Speaker 2>So instead of predicting the exact price of a house, okay,

201
00:10:12.879 --> 00:10:15.559
<v Speaker 2>you might predict whether it falls into a low, medium,

202
00:10:15.639 --> 00:10:16.759
<v Speaker 2>or high price range.

203
00:10:16.960 --> 00:10:20.360
<v Speaker 1>It's like simplifying the problem to help AUTOMML find patterns

204
00:10:20.360 --> 00:10:22.840
<v Speaker 1>more easily. Exactly, Okay? What are their tips does the

205
00:10:22.840 --> 00:10:23.360
<v Speaker 1>book offer?

206
00:10:23.519 --> 00:10:28.440
<v Speaker 2>It emphasizes the importance of experimenting with different primary metrics. Okay, So,

207
00:10:28.600 --> 00:10:31.279
<v Speaker 2>for example, you might find that a metric like mean

208
00:10:31.399 --> 00:10:36.159
<v Speaker 2>absolute error MAE gives you a more meaningful evaluation than

209
00:10:36.200 --> 00:10:39.320
<v Speaker 2>something like root means squared error or RMS.

210
00:10:39.720 --> 00:10:41.559
<v Speaker 1>Right, So it all depends on what you're trying to

211
00:10:41.600 --> 00:10:42.519
<v Speaker 1>achieve with your model.

212
00:10:42.639 --> 00:10:43.039
<v Speaker 2>It does.

213
00:10:43.240 --> 00:10:45.840
<v Speaker 1>Okay, So there's still a lot of room for human

214
00:10:45.919 --> 00:10:49.960
<v Speaker 1>judgment and decision making, even with AUTOMML handling so much

215
00:10:49.960 --> 00:10:53.080
<v Speaker 1>of that complexity. Now I have to ask, what about

216
00:10:53.080 --> 00:10:55.120
<v Speaker 1>those situations where you need to train not just one

217
00:10:55.200 --> 00:10:56.559
<v Speaker 1>or two models, but one hundreds.

218
00:10:57.120 --> 00:10:58.360
<v Speaker 2>The many models.

219
00:10:57.960 --> 00:10:59.600
<v Speaker 1>Probablybe even thousands of models.

220
00:10:59.720 --> 00:11:03.000
<v Speaker 2>That's the book really shines. It introduces a tool called

221
00:11:03.000 --> 00:11:08.200
<v Speaker 2>the Many Models Solution Accelerator or MMSA. MMSA okay, think

222
00:11:08.200 --> 00:11:10.919
<v Speaker 2>of it as an auto mL factory. You feed it

223
00:11:11.080 --> 00:11:14.360
<v Speaker 2>your data I'm listening, and it automatically splits it up

224
00:11:14.399 --> 00:11:18.960
<v Speaker 2>based on certain criteria, different stores, products, or regions, and

225
00:11:19.000 --> 00:11:22.480
<v Speaker 2>then it uses the power of auto mL to train

226
00:11:22.639 --> 00:11:26.320
<v Speaker 2>a separate model for each of those subsets, all running

227
00:11:26.360 --> 00:11:27.039
<v Speaker 2>in parallel.

228
00:11:27.519 --> 00:11:28.720
<v Speaker 1>That's a lot of models.

229
00:11:28.799 --> 00:11:29.639
<v Speaker 2>It is a lot of models.

230
00:11:29.679 --> 00:11:32.320
<v Speaker 1>So why would someone need to create so many models?

231
00:11:32.360 --> 00:11:36.759
<v Speaker 2>It's incredibly useful when you need granular, customized models.

232
00:11:36.799 --> 00:11:37.320
<v Speaker 1>So got it.

233
00:11:37.879 --> 00:11:40.480
<v Speaker 2>The book gives an example of a retail chain that

234
00:11:40.519 --> 00:11:44.120
<v Speaker 2>wants to forecast demand for each product in each store.

235
00:11:44.679 --> 00:11:48.679
<v Speaker 2>By using MMSA, they can create thousands of custom models

236
00:11:48.720 --> 00:11:52.679
<v Speaker 2>that account for all the unique factors that might influence sales.

237
00:11:53.000 --> 00:11:55.240
<v Speaker 1>Wow. So it's like hyper personalized AI.

238
00:11:55.799 --> 00:11:56.240
<v Speaker 2>It is.

239
00:11:56.399 --> 00:11:57.200
<v Speaker 1>That's incredible.

240
00:11:57.440 --> 00:11:58.759
<v Speaker 2>Imagine there are things to keep in.

241
00:11:58.639 --> 00:12:01.039
<v Speaker 1>Mind, right, there's to be some catches.

242
00:12:01.159 --> 00:12:03.679
<v Speaker 2>One of the key points the book highlights is choosing

243
00:12:03.720 --> 00:12:05.200
<v Speaker 2>the right partition columns.

244
00:12:05.759 --> 00:12:06.200
<v Speaker 1>Columns.

245
00:12:06.200 --> 00:12:08.039
<v Speaker 2>These are the criteria that you use to divide your

246
00:12:08.080 --> 00:12:11.320
<v Speaker 2>data into those subsets. You need to think carefully about

247
00:12:11.360 --> 00:12:15.039
<v Speaker 2>which factors are most likely to influence the target variable

248
00:12:15.080 --> 00:12:16.200
<v Speaker 2>that you're trying to predict.

249
00:12:16.639 --> 00:12:19.000
<v Speaker 1>So it's like making sure you're slicing and dicing your

250
00:12:19.080 --> 00:12:22.039
<v Speaker 1>data along the right lines so that the models you

251
00:12:22.120 --> 00:12:25.799
<v Speaker 1>create are actually meaningful and relevant precisely. Now, once you've

252
00:12:25.799 --> 00:12:29.039
<v Speaker 1>trained all these models using the MMSA, you still need

253
00:12:29.080 --> 00:12:31.480
<v Speaker 1>to deploy them. You do, and that can get even

254
00:12:31.519 --> 00:12:35.480
<v Speaker 1>more complex when you're dealing with thousands of models.

255
00:12:35.639 --> 00:12:37.840
<v Speaker 2>It can instead is just one it does.

256
00:12:38.320 --> 00:12:39.759
<v Speaker 1>So how do you manage all of that?

257
00:12:40.360 --> 00:12:43.440
<v Speaker 2>The book really stresses the importance of automation here. Okay,

258
00:12:43.519 --> 00:12:45.759
<v Speaker 2>It introduces a tool called Azure Data.

259
00:12:45.559 --> 00:12:47.480
<v Speaker 1>Factory as your data factor, which can.

260
00:12:47.360 --> 00:12:52.240
<v Speaker 2>Help orchestrate complex data flows, connect to various data sources,

261
00:12:52.519 --> 00:12:55.840
<v Speaker 2>transform data, and even trigger those mL pipelines that we

262
00:12:55.840 --> 00:12:56.720
<v Speaker 2>talked about earlier.

263
00:12:56.759 --> 00:12:59.240
<v Speaker 1>So if those mL pipelines are like assembly lines for

264
00:12:59.279 --> 00:13:02.519
<v Speaker 1>building our model. Yeah, then Azure Data Factory is like

265
00:13:02.639 --> 00:13:05.360
<v Speaker 1>the logistics manager, making sure all the raw materials and

266
00:13:05.399 --> 00:13:08.600
<v Speaker 1>finished products are flowing smoothly, perfect. And in addition to

267
00:13:08.600 --> 00:13:11.159
<v Speaker 1>all this, the book is packed with helpful tips and

268
00:13:11.159 --> 00:13:13.799
<v Speaker 1>tricks for getting the most out of AUTOMML. It is

269
00:13:14.120 --> 00:13:16.600
<v Speaker 1>regardless of the type of problem that you're tackling. Yes,

270
00:13:16.720 --> 00:13:19.960
<v Speaker 1>I love that practical advice. It's like having a seasoned

271
00:13:19.960 --> 00:13:23.399
<v Speaker 1>auto mL expert looking over your shoulder, absolutely guiding you

272
00:13:23.440 --> 00:13:24.600
<v Speaker 1>along the way exactly.

273
00:13:24.799 --> 00:13:28.799
<v Speaker 2>And the book really encourages a spirit of experimentation. Okay,

274
00:13:28.879 --> 00:13:31.200
<v Speaker 2>it wants you to be a data detective. Oh I

275
00:13:31.279 --> 00:13:33.720
<v Speaker 2>like that, to try different approaches and see what works

276
00:13:33.720 --> 00:13:34.919
<v Speaker 2>best for your situation.

277
00:13:35.720 --> 00:13:38.480
<v Speaker 1>That's what makes data science so exciting. It's not about

278
00:13:38.480 --> 00:13:44.000
<v Speaker 1>blindly following rules. It's about exploring, discovering, and finding creative solutions.

279
00:13:44.120 --> 00:13:44.879
<v Speaker 2>Absolutely.

280
00:13:45.120 --> 00:13:47.279
<v Speaker 1>We've talked a lot about az your machine Learning studio,

281
00:13:47.320 --> 00:13:51.519
<v Speaker 1>which has that visual, user friendly interface, But for those

282
00:13:51.559 --> 00:13:55.000
<v Speaker 1>who prefer to work with code, the book also covers

283
00:13:55.000 --> 00:13:58.559
<v Speaker 1>the Azure Machine Learning SDK for Python. It does so

284
00:13:58.879 --> 00:14:01.039
<v Speaker 1>for those who are comfortable covading, there's a way to

285
00:14:01.039 --> 00:14:04.320
<v Speaker 1>get even more control and flexibility over that auto mL

286
00:14:04.440 --> 00:14:07.399
<v Speaker 1>process there is, okay, and the book walks us through

287
00:14:07.639 --> 00:14:11.080
<v Speaker 1>using Jupiter notebooks within Asure machine Learning, which is a

288
00:14:11.240 --> 00:14:15.320
<v Speaker 1>very popular way to write and execute Python code for

289
00:14:15.440 --> 00:14:18.320
<v Speaker 1>data science tasks. It is, and you can actually use

290
00:14:18.440 --> 00:14:22.240
<v Speaker 1>those powerful compute clusters that we talked about earlier to

291
00:14:22.399 --> 00:14:25.639
<v Speaker 1>run your code in the cloud, yes, giving you access

292
00:14:25.679 --> 00:14:27.159
<v Speaker 1>to tons of processing power.

293
00:14:27.240 --> 00:14:27.879
<v Speaker 2>Absolutely.

294
00:14:28.240 --> 00:14:31.360
<v Speaker 1>It really is amazing how cloud computing has made these

295
00:14:31.399 --> 00:14:35.519
<v Speaker 1>really complex tasks so much more accessible. It has. Now

296
00:14:35.559 --> 00:14:38.200
<v Speaker 1>I'm curious about the different ways that you can actually

297
00:14:38.240 --> 00:14:41.879
<v Speaker 1>deploy models once they're trained. We talked about batch scoring

298
00:14:41.919 --> 00:14:44.559
<v Speaker 1>and real time scoring, yes, but the book also mentioned

299
00:14:44.559 --> 00:14:46.159
<v Speaker 1>something called mL pipelines.

300
00:14:46.440 --> 00:14:50.120
<v Speaker 2>mL pipelines are a fantastic way to automate your entire

301
00:14:50.240 --> 00:14:54.639
<v Speaker 2>machine learning workflow, okay, from data preprocessing and feature engineering

302
00:14:54.919 --> 00:14:56.919
<v Speaker 2>to model training and deployment.

303
00:14:57.240 --> 00:14:59.440
<v Speaker 1>So it's like having an assembly line for your AI,

304
00:15:00.279 --> 00:15:02.799
<v Speaker 1>ensuring that each step is executed in the right order,

305
00:15:03.080 --> 00:15:07.440
<v Speaker 1>with the right settings. Absolutely, automation, efficiency, consistency. That's it

306
00:15:07.440 --> 00:15:09.559
<v Speaker 1>sounds like a well oiled machine exactly.

307
00:15:09.600 --> 00:15:12.159
<v Speaker 2>And to take it a step further, the book introduces

308
00:15:12.240 --> 00:15:13.759
<v Speaker 2>us to Azure Data.

309
00:15:13.559 --> 00:15:15.440
<v Speaker 1>Factory Azure Data Factory.

310
00:15:15.519 --> 00:15:19.159
<v Speaker 2>This is a cloud based data integration service okay that

311
00:15:19.200 --> 00:15:23.159
<v Speaker 2>can handle even more complex data flows, connecting to different

312
00:15:23.240 --> 00:15:26.679
<v Speaker 2>data sources, got it, transforming your data okay, and even

313
00:15:26.720 --> 00:15:28.519
<v Speaker 2>triggering your mL pipelines.

314
00:15:28.720 --> 00:15:31.279
<v Speaker 1>So if mL pipelines are the assembly lines, that Azure

315
00:15:31.320 --> 00:15:34.679
<v Speaker 1>Data Factory is like the logistics manager making sure that

316
00:15:34.759 --> 00:15:38.639
<v Speaker 1>all the raw materials and the finished products are flowing smoothly.

317
00:15:38.799 --> 00:15:39.879
<v Speaker 2>A great way to put it.

318
00:15:40.000 --> 00:15:42.559
<v Speaker 1>Now, we've covered a lot of ground here, from the

319
00:15:42.679 --> 00:15:46.320
<v Speaker 1>really medi gritty details of data preparation and feature engineering

320
00:15:46.679 --> 00:15:49.960
<v Speaker 1>to the broader concepts of model deployment and automation. We have,

321
00:15:50.159 --> 00:15:53.840
<v Speaker 1>but let's not forget about one really crucial aspect, the

322
00:15:54.080 --> 00:15:58.120
<v Speaker 1>human element. Building a model is only part of the story. Yeah,

323
00:15:58.159 --> 00:16:02.320
<v Speaker 1>what about gaining the trust and buy in of the

324
00:16:02.360 --> 00:16:04.320
<v Speaker 1>people who will actually be using these models.

325
00:16:04.440 --> 00:16:06.279
<v Speaker 2>That's a great point. You know, we can build the

326
00:16:06.320 --> 00:16:09.320
<v Speaker 2>most sophisticated AI in the world, but if people don't

327
00:16:09.360 --> 00:16:11.279
<v Speaker 2>trust it or understand how.

328
00:16:11.120 --> 00:16:12.480
<v Speaker 1>It works, what good is it.

329
00:16:12.200 --> 00:16:14.799
<v Speaker 2>It's not going to be very useful, right, And that's

330
00:16:14.840 --> 00:16:19.200
<v Speaker 2>why those explainability features we keep talking about are so important.

331
00:16:19.799 --> 00:16:24.799
<v Speaker 2>Extremely the book emphasizes the need to clearly articulate how

332
00:16:24.840 --> 00:16:28.320
<v Speaker 2>a model is making decisions, especially in industries that have

333
00:16:28.399 --> 00:16:31.159
<v Speaker 2>strict regulations, got it, or where the stakes.

334
00:16:30.840 --> 00:16:34.240
<v Speaker 1>Are high, right, like healthcare of finance exactly. So it's

335
00:16:34.279 --> 00:16:37.039
<v Speaker 1>not enough to just say the computer says this is

336
00:16:37.080 --> 00:16:39.399
<v Speaker 1>the best course of action. We need to be able

337
00:16:39.399 --> 00:16:42.000
<v Speaker 1>to back that up with insights and evidence.

338
00:16:41.639 --> 00:16:44.519
<v Speaker 2>Absolutely, and auto mL gives us the tools to do

339
00:16:44.759 --> 00:16:48.200
<v Speaker 2>just that. You can use feature importance scores to see

340
00:16:48.200 --> 00:16:51.480
<v Speaker 2>which factors are driving the model's predictions. You can even

341
00:16:51.559 --> 00:16:55.039
<v Speaker 2>drill down into individual predictions to understand why the model

342
00:16:55.080 --> 00:16:56.519
<v Speaker 2>made a specific decision.

343
00:16:56.840 --> 00:16:58.879
<v Speaker 1>So it's like having a transparent AI where we can

344
00:16:58.919 --> 00:17:02.159
<v Speaker 1>peek under the hood and see what's going on exactly. Now,

345
00:17:02.200 --> 00:17:04.200
<v Speaker 1>before we wrap up this part of our deep dive,

346
00:17:05.079 --> 00:17:07.119
<v Speaker 1>I want to highlight something that really stood out to

347
00:17:07.160 --> 00:17:09.720
<v Speaker 1>me while reading the book. It mentions that auto mL

348
00:17:09.799 --> 00:17:13.279
<v Speaker 1>can actually be used in other Microsoft products besides Azure

349
00:17:13.359 --> 00:17:16.359
<v Speaker 1>Machine Learning Studio. Yes, oh, so it's not just limited

350
00:17:16.359 --> 00:17:17.759
<v Speaker 1>to this one platform, not at all.

351
00:17:18.119 --> 00:17:21.079
<v Speaker 2>The book talks about integrating auto mL with tools like

352
00:17:21.240 --> 00:17:26.279
<v Speaker 2>Powerbi Powerbi which is a powerful data visualization and business

353
00:17:26.319 --> 00:17:27.960
<v Speaker 2>intelligence platform, So you.

354
00:17:27.920 --> 00:17:32.079
<v Speaker 1>Could create these interactive dashboards that not only display the data,

355
00:17:32.279 --> 00:17:35.960
<v Speaker 1>but use auto mL to generate predictions and insights. Yes,

356
00:17:36.359 --> 00:17:40.079
<v Speaker 1>that's next level. That takes data storytelling to a whole

357
00:17:40.279 --> 00:17:42.960
<v Speaker 1>new level. It does. And the book also touches on

358
00:17:43.359 --> 00:17:47.359
<v Speaker 1>using auto mL with Azure Synapse Analytics, which is a

359
00:17:47.480 --> 00:17:50.680
<v Speaker 1>cloud based data warehousing and analytics.

360
00:17:50.240 --> 00:17:53.880
<v Speaker 2>Service exactly, and this opens up even more possibilities. Wow,

361
00:17:53.920 --> 00:17:57.559
<v Speaker 2>for working with massive data sets and building these enterprise

362
00:17:57.599 --> 00:17:59.160
<v Speaker 2>scale AI solutions.

363
00:17:59.200 --> 00:18:01.799
<v Speaker 1>It sounds like the possible are pretty much endless. They are.

364
00:18:02.079 --> 00:18:04.640
<v Speaker 1>We've covered so much ground, but it feels like we've

365
00:18:04.640 --> 00:18:06.759
<v Speaker 1>only just begun to scratch the surface of what auto

366
00:18:06.839 --> 00:18:07.880
<v Speaker 1>mL on Azure can do.

367
00:18:08.039 --> 00:18:09.599
<v Speaker 2>We've just scratched the surface.

368
00:18:09.799 --> 00:18:11.480
<v Speaker 1>So as we move into the next part of our

369
00:18:11.480 --> 00:18:15.880
<v Speaker 1>deep dive, I'm curious where should someone start or if

370
00:18:15.880 --> 00:18:18.680
<v Speaker 1>they want to dive in and explore this world of

371
00:18:18.759 --> 00:18:20.559
<v Speaker 1>auto mL on Azure.

372
00:18:20.799 --> 00:18:24.359
<v Speaker 2>Well, this book we've been discussing is an excellent starting point. Okay,

373
00:18:24.519 --> 00:18:27.319
<v Speaker 2>Dennis Sawyer's has done a really great job of creating

374
00:18:27.359 --> 00:18:30.319
<v Speaker 2>a practical and easy to follow guide. I agree, pact

375
00:18:30.359 --> 00:18:35.240
<v Speaker 2>with real world examples, code snippets, and tons of helpful advice.

376
00:18:35.119 --> 00:18:38.799
<v Speaker 1>And there are tons of online resources, including Microsoft's own

377
00:18:38.880 --> 00:18:43.920
<v Speaker 1>documentation and tutorials exactly. And plus the Azure community is

378
00:18:44.000 --> 00:18:45.640
<v Speaker 1>incredibly active and supportive.

379
00:18:45.720 --> 00:18:46.039
<v Speaker 2>It is.

380
00:18:46.119 --> 00:18:48.440
<v Speaker 1>It's great. So if you have questions or get stuck,

381
00:18:48.880 --> 00:18:51.519
<v Speaker 1>you'll find plenty of people willing to help you will.

382
00:18:51.799 --> 00:18:54.640
<v Speaker 1>I think the key takeaway here is that AutoML is

383
00:18:54.680 --> 00:18:57.079
<v Speaker 1>not some futuristic concept.

384
00:18:57.279 --> 00:18:57.839
<v Speaker 2>No it's not.

385
00:18:58.240 --> 00:18:59.920
<v Speaker 1>It's here here, it's here now.

386
00:19:00.079 --> 00:19:01.599
<v Speaker 2>It's now a powerful tool.

387
00:19:01.720 --> 00:19:03.920
<v Speaker 1>It's a powerful tool that's available right now.

388
00:19:03.799 --> 00:19:05.960
<v Speaker 2>And it's making AI more accessible.

389
00:19:05.559 --> 00:19:08.039
<v Speaker 1>Than ever before. And with the power of auto mL

390
00:19:08.160 --> 00:19:13.039
<v Speaker 1>on Azure, anyone can become an AI innovator. So to

391
00:19:13.200 --> 00:19:17.720
<v Speaker 1>our listener, we challenge you, what problem will you solve

392
00:19:18.039 --> 00:19:20.960
<v Speaker 1>with auto mL. That's a great question, that's something to

393
00:19:21.000 --> 00:19:23.319
<v Speaker 1>think about. Think about it. We'll be back in just

394
00:19:23.359 --> 00:19:25.599
<v Speaker 1>a moment with part two of our deep dive into

395
00:19:25.599 --> 00:19:29.720
<v Speaker 1>AutoML on Azure. Welcome back to our deep dive into

396
00:19:29.759 --> 00:19:33.319
<v Speaker 1>auto mL on Azure. Now, before the break, we were

397
00:19:33.319 --> 00:19:35.599
<v Speaker 1>talking about how auto mL can be used in other

398
00:19:35.680 --> 00:19:39.519
<v Speaker 1>Microsoft products. Yes, besides just as your machine learning studio.

399
00:19:40.079 --> 00:19:42.680
<v Speaker 1>So it's not just limited to that one platform, not

400
00:19:42.799 --> 00:19:44.079
<v Speaker 1>at all, Okay, So tell me more.

401
00:19:44.240 --> 00:19:47.119
<v Speaker 2>The book talks about integrating auto mL with tools like

402
00:19:47.200 --> 00:19:51.400
<v Speaker 2>powerbi powerbi okay, which is a powerful data visualization and

403
00:19:51.559 --> 00:19:53.680
<v Speaker 2>business intelligence platform.

404
00:19:53.799 --> 00:19:57.240
<v Speaker 1>So we can use AutoML to create these really interactive

405
00:19:57.319 --> 00:20:00.240
<v Speaker 1>dashboards you got it that not only display d data

406
00:20:00.359 --> 00:20:02.799
<v Speaker 1>but also generate predictions and insights.

407
00:20:03.160 --> 00:20:03.720
<v Speaker 2>Exactly.

408
00:20:03.880 --> 00:20:04.400
<v Speaker 1>It's awesome.

409
00:20:04.440 --> 00:20:06.680
<v Speaker 2>It takes data storytelling to a whole new level.

410
00:20:06.720 --> 00:20:09.279
<v Speaker 1>It does, Yeah, it does. And the book even touches

411
00:20:09.319 --> 00:20:13.039
<v Speaker 1>on using AutoML with Azure SYNEPS Analytics, which is a

412
00:20:13.160 --> 00:20:17.039
<v Speaker 1>cloud based data warehousing and analytics service. Wow, so many options.

413
00:20:17.480 --> 00:20:20.559
<v Speaker 1>So this opens up even more possibilities for working with

414
00:20:20.680 --> 00:20:25.240
<v Speaker 1>massive data sets, yes, and building these enterprise scale AI solutions.

415
00:20:25.759 --> 00:20:29.119
<v Speaker 1>It sounds like AutoML is becoming this really integral part

416
00:20:29.240 --> 00:20:32.039
<v Speaker 1>of the entire Microsoft ecosystem. It is, so it's not

417
00:20:32.119 --> 00:20:35.440
<v Speaker 1>just this standalone tool, it's really integrated into all these

418
00:20:35.480 --> 00:20:39.400
<v Speaker 1>different products and services exactly. Yeah, okay. Now, speaking of

419
00:20:39.400 --> 00:20:42.880
<v Speaker 1>getting hands on with AutoML, the book goes beyond just

420
00:20:42.960 --> 00:20:47.160
<v Speaker 1>talking about the concepts. It actually dives into using Azure

421
00:20:47.200 --> 00:20:51.599
<v Speaker 1>Machine Learning Studio Yes, which, as we discussed earlier, is

422
00:20:51.640 --> 00:20:55.839
<v Speaker 1>that more visual, user friendly interface for working with AMLS.

423
00:20:56.039 --> 00:20:59.960
<v Speaker 1>It is, and it uses a classic example, the Titan

424
00:21:00.279 --> 00:21:01.200
<v Speaker 1>Passenger Data.

425
00:21:01.279 --> 00:21:02.920
<v Speaker 2>Oh, the Titanic data set.

426
00:21:02.759 --> 00:21:07.079
<v Speaker 1>Predicting who survived and who didn't based on factors like age, gender,

427
00:21:07.359 --> 00:21:08.920
<v Speaker 1>ticket class.

428
00:21:08.640 --> 00:21:10.000
<v Speaker 2>Right, a classic problem.

429
00:21:10.119 --> 00:21:13.319
<v Speaker 1>It's like a rite of passage for anyone learning data science. Exactly.

430
00:21:13.359 --> 00:21:15.359
<v Speaker 1>I remember working with that data set in my early day.

431
00:21:15.480 --> 00:21:18.359
<v Speaker 1>So the book actually walks you through the entire process,

432
00:21:18.440 --> 00:21:20.119
<v Speaker 1>making it surprisingly straightforward.

433
00:21:20.279 --> 00:21:21.640
<v Speaker 2>It does. It makes it really easy.

434
00:21:22.200 --> 00:21:24.559
<v Speaker 1>So so you upload your data, tell AutoML it's a

435
00:21:24.559 --> 00:21:26.839
<v Speaker 1>classification task, and let it do its thing.

436
00:21:27.000 --> 00:21:29.079
<v Speaker 2>Yeah, let it work its magic, exactly.

437
00:21:29.599 --> 00:21:32.480
<v Speaker 1>And while it's working its magic, we can keep an

438
00:21:32.480 --> 00:21:35.599
<v Speaker 1>eye on those data guard rails to make sure nothing

439
00:21:35.640 --> 00:21:38.279
<v Speaker 1>fishy is going on with our data exactly, right, We've

440
00:21:38.319 --> 00:21:42.240
<v Speaker 1>got our data quality watchdog. We do. Now, once auto

441
00:21:42.400 --> 00:21:45.880
<v Speaker 1>mL has finished training a whole bunch of models, it's

442
00:21:45.880 --> 00:21:49.519
<v Speaker 1>evaluation time, right. Okay, So we have all these models

443
00:21:49.559 --> 00:21:52.440
<v Speaker 1>all trained and ready to go. How do we pick

444
00:21:52.480 --> 00:21:53.319
<v Speaker 1>the best one?

445
00:21:53.599 --> 00:21:56.680
<v Speaker 2>So auto mL gives us a whole Smorgus board of

446
00:21:56.880 --> 00:22:01.519
<v Speaker 2>metrics to consider. Okay, you know, accuracy, decision recall.

447
00:22:01.279 --> 00:22:02.079
<v Speaker 1>So many metrics.

448
00:22:02.160 --> 00:22:02.799
<v Speaker 2>There are a lot.

449
00:22:03.039 --> 00:22:05.240
<v Speaker 1>Sometimes I feel a little overwhelmed, especially if you're new

450
00:22:05.240 --> 00:22:05.920
<v Speaker 1>to machine learning.

451
00:22:06.079 --> 00:22:09.279
<v Speaker 2>Sure, I understand that there's just so many numbers, right,

452
00:22:09.400 --> 00:22:12.759
<v Speaker 2>but the book breaks down each metric, explaining what it

453
00:22:12.839 --> 00:22:14.359
<v Speaker 2>means and when it's most relevant.

454
00:22:14.559 --> 00:22:15.519
<v Speaker 1>Okay, that's helpful.

455
00:22:15.599 --> 00:22:19.519
<v Speaker 2>Yeah, and it also reminds us to use those explainability features, oh,

456
00:22:19.680 --> 00:22:24.160
<v Speaker 2>to really understand why a model is making certain predictions.

457
00:22:24.279 --> 00:22:27.279
<v Speaker 1>Right. It's not enough to just blindly trust the numbers exactly.

458
00:22:27.359 --> 00:22:27.960
<v Speaker 2>We need more.

459
00:22:28.279 --> 00:22:32.480
<v Speaker 1>It's about understanding the logic behind those numbers. Yes, making

460
00:22:32.519 --> 00:22:34.920
<v Speaker 1>sure the model is making decisions for the right reasons.

461
00:22:35.039 --> 00:22:38.279
<v Speaker 1>That's right. Okay. So let's say we've found a model

462
00:22:38.319 --> 00:22:41.119
<v Speaker 1>that looks good both in terms of its performance and

463
00:22:41.200 --> 00:22:42.160
<v Speaker 1>its explainability.

464
00:22:42.279 --> 00:22:46.839
<v Speaker 2>Okay, what's next deployment? It's time to take that model

465
00:22:46.880 --> 00:22:48.720
<v Speaker 2>out of the training ground and put it to work

466
00:22:48.759 --> 00:22:49.559
<v Speaker 2>in the real world.

467
00:22:50.000 --> 00:22:52.960
<v Speaker 1>Okay, so we're moving from theory to practice. Yes, but

468
00:22:53.079 --> 00:22:54.799
<v Speaker 1>how do we actually deploy a model?

469
00:22:55.359 --> 00:22:57.480
<v Speaker 2>So the book gives us a couple of options. Okay,

470
00:22:57.480 --> 00:23:01.880
<v Speaker 2>there's batch scoring, which is great for processing large amounts

471
00:23:01.880 --> 00:23:05.319
<v Speaker 2>of data at scheduled intervals, and then there's real time scoring,

472
00:23:05.359 --> 00:23:08.279
<v Speaker 2>which is ideal for when you need those instant predictions,

473
00:23:08.319 --> 00:23:11.160
<v Speaker 2>like in fraud detection or personalized recommendation.

474
00:23:11.319 --> 00:23:14.640
<v Speaker 1>So batch scoring is like sending out a monthly newsletter. Yes,

475
00:23:14.839 --> 00:23:17.200
<v Speaker 1>while real time scoring is like having a live chat

476
00:23:17.200 --> 00:23:18.000
<v Speaker 1>with your customers.

477
00:23:18.279 --> 00:23:19.920
<v Speaker 2>That's a great analogy.

478
00:23:19.680 --> 00:23:23.400
<v Speaker 1>Right, Different tools for different needs. Absolutely. Up until now

479
00:23:23.480 --> 00:23:27.559
<v Speaker 1>we've mostly focused on classification models, right, where the goal

480
00:23:27.680 --> 00:23:31.440
<v Speaker 1>is to predict a category. But the book doesn't stop there, No,

481
00:23:31.640 --> 00:23:35.279
<v Speaker 1>it doesn't. It also dives into regression models, okay, where

482
00:23:35.319 --> 00:23:37.920
<v Speaker 1>the goal is to predict a numerical value.

483
00:23:37.720 --> 00:23:40.160
<v Speaker 2>Right, like predicting the price of a house exactly, or

484
00:23:40.200 --> 00:23:42.480
<v Speaker 2>the number of sales you'll make next month, right.

485
00:23:42.480 --> 00:23:44.319
<v Speaker 1>Or in the case of the diabetes data set that

486
00:23:44.319 --> 00:23:47.920
<v Speaker 1>we mentioned earlier, predicting the progression of the disease exact.

487
00:23:47.960 --> 00:23:50.880
<v Speaker 1>So the process is very similar, but obviously the metrics

488
00:23:50.880 --> 00:23:52.359
<v Speaker 1>and the algorithms we use are different.

489
00:23:52.519 --> 00:23:54.160
<v Speaker 2>Right, Different tools for different jobs.

490
00:23:54.279 --> 00:23:58.079
<v Speaker 1>But auto mL still handles all of that heavy lifting

491
00:23:58.160 --> 00:24:00.880
<v Speaker 1>for us. It does, which is great, but we still

492
00:24:00.920 --> 00:24:04.039
<v Speaker 1>need to understand which tools are the best for the

493
00:24:04.079 --> 00:24:04.799
<v Speaker 1>task at hand.

494
00:24:05.200 --> 00:24:05.839
<v Speaker 2>Absolutely.

495
00:24:06.359 --> 00:24:08.559
<v Speaker 1>Now, I bet there are some special tips and tricks

496
00:24:08.599 --> 00:24:10.799
<v Speaker 1>for getting the most out of auto mL. There are,

497
00:24:11.359 --> 00:24:15.319
<v Speaker 1>specifically for regression problems. Yes, so spill the beans. What

498
00:24:15.440 --> 00:24:16.200
<v Speaker 1>are they? Well?

499
00:24:16.279 --> 00:24:20.319
<v Speaker 2>The book shares some really interesting insights, Okay. For instance,

500
00:24:20.640 --> 00:24:23.640
<v Speaker 2>sometimes it can actually improve performance to.

501
00:24:23.640 --> 00:24:26.880
<v Speaker 1>Convert a regression problem into a classification problem.

502
00:24:26.920 --> 00:24:28.480
<v Speaker 2>Hold on, how do you do that? How do you

503
00:24:28.519 --> 00:24:31.799
<v Speaker 2>convert a problem about predicting a number into a problem

504
00:24:31.839 --> 00:24:33.000
<v Speaker 2>about predicting a category?

505
00:24:33.039 --> 00:24:35.880
<v Speaker 1>It's all about binning dah. Okay, you divide the range

506
00:24:35.920 --> 00:24:41.039
<v Speaker 1>of possible numerical values into categories or vins. Okay, So

507
00:24:41.319 --> 00:24:44.519
<v Speaker 1>instead of predicting the exact price of a house, you

508
00:24:44.599 --> 00:24:47.480
<v Speaker 1>might predict whether it falls into a low, medium, or

509
00:24:47.559 --> 00:24:48.759
<v Speaker 1>high price range.

510
00:24:48.960 --> 00:24:51.160
<v Speaker 2>So it's like simplifying the problem a bit to make

511
00:24:51.200 --> 00:24:54.240
<v Speaker 2>it easier for auto and mlifying those patterns exactly. Okay,

512
00:24:54.279 --> 00:24:55.440
<v Speaker 2>I see any other tips.

513
00:24:55.720 --> 00:24:59.440
<v Speaker 1>The book also emphasizes the importance of experimenting with different

514
00:24:59.519 --> 00:25:00.599
<v Speaker 1>primary metrics.

515
00:25:00.680 --> 00:25:01.400
<v Speaker 2>Okay, right.

516
00:25:01.440 --> 00:25:03.960
<v Speaker 1>For example, you might find that a metric like mean

517
00:25:04.079 --> 00:25:08.880
<v Speaker 1>absolute error or MAE gives you a more meaningful evaluation

518
00:25:09.559 --> 00:25:14.160
<v Speaker 1>than something like root means squared error or URMSK. It

519
00:25:14.200 --> 00:25:16.680
<v Speaker 1>really depends on what you're trying to achieve with your model.

520
00:25:17.039 --> 00:25:19.880
<v Speaker 2>So it's not just about blindly applying the same metrics

521
00:25:19.920 --> 00:25:20.359
<v Speaker 2>every time.

522
00:25:20.640 --> 00:25:21.359
<v Speaker 1>No, not at all.

523
00:25:21.559 --> 00:25:23.920
<v Speaker 2>You need to really think about what's most relevant for

524
00:25:23.960 --> 00:25:25.200
<v Speaker 2>your specific problem.

525
00:25:25.839 --> 00:25:27.839
<v Speaker 1>Okay, So there's still a lot of room for human

526
00:25:27.920 --> 00:25:30.960
<v Speaker 1>judgment and decision making, even with auto mL doing so

527
00:25:31.039 --> 00:25:33.519
<v Speaker 1>much of the work. Absolutely, Now I have to ask,

528
00:25:34.000 --> 00:25:36.319
<v Speaker 1>what about those situations where you need to train not

529
00:25:36.400 --> 00:25:39.759
<v Speaker 1>just one or two models, but hundreds, maybe even thousands.

530
00:25:39.880 --> 00:25:42.039
<v Speaker 2>Ah, the many models problem?

531
00:25:42.240 --> 00:25:44.920
<v Speaker 1>Right. It sounds daunting, it can.

532
00:25:44.799 --> 00:25:46.759
<v Speaker 2>Be, but that's where the book really shines.

533
00:25:46.839 --> 00:25:47.119
<v Speaker 1>Okay.

534
00:25:47.160 --> 00:25:50.960
<v Speaker 2>It introduces this tool called the Many Models Solution Accelerator

535
00:25:51.079 --> 00:25:54.039
<v Speaker 2>or MMSA for sure, MMSA. Think of it as an

536
00:25:54.039 --> 00:25:55.160
<v Speaker 2>auto mL factory.

537
00:25:55.240 --> 00:25:55.839
<v Speaker 1>Okay. I like that.

538
00:25:56.160 --> 00:25:59.039
<v Speaker 2>You feed it your data and it automatically splits it

539
00:25:59.119 --> 00:26:03.319
<v Speaker 2>up based on certain criteria like different stores, products, or regions.

540
00:26:03.480 --> 00:26:06.359
<v Speaker 1>Right, so you can create really customized models for each

541
00:26:06.440 --> 00:26:10.759
<v Speaker 1>specific situation exactly. Okay. So when would someone actually need

542
00:26:10.799 --> 00:26:12.079
<v Speaker 1>to create so many models?

543
00:26:12.119 --> 00:26:16.039
<v Speaker 2>It's ideal for when you need those really granular customized models.

544
00:26:16.119 --> 00:26:16.440
<v Speaker 1>Okay.

545
00:26:16.559 --> 00:26:19.240
<v Speaker 2>The book gives an example of a retail chain okay

546
00:26:19.319 --> 00:26:22.759
<v Speaker 2>that wants to forecast demand for each product in each store.

547
00:26:23.119 --> 00:26:25.480
<v Speaker 1>Ah, So they need a separate model for each product

548
00:26:25.519 --> 00:26:26.079
<v Speaker 1>in each store.

549
00:26:26.480 --> 00:26:31.720
<v Speaker 2>Yes, And by using the MMSA, they can create thousands

550
00:26:31.799 --> 00:26:35.000
<v Speaker 2>of custom models that take into account all of the

551
00:26:35.079 --> 00:26:38.440
<v Speaker 2>unique factors that might influence sales at each location.

552
00:26:38.920 --> 00:26:42.440
<v Speaker 1>That's incredibly powerful. It is, but I imagine there are

553
00:26:42.440 --> 00:26:46.319
<v Speaker 1>some considerations when working with this MMSA.

554
00:26:45.480 --> 00:26:48.480
<v Speaker 2>Definitely, and one of the key points the book highlights

555
00:26:49.039 --> 00:26:51.240
<v Speaker 2>is choosing the right partition.

556
00:26:50.920 --> 00:26:52.920
<v Speaker 1>Columns, partision calls. What are those?

557
00:26:53.000 --> 00:26:55.640
<v Speaker 2>These are the criteria you use to divide your data

558
00:26:55.640 --> 00:26:58.880
<v Speaker 2>into those subsets. Okay, you need to think carefully about

559
00:26:58.920 --> 00:27:02.640
<v Speaker 2>which factors are most likely to influence the target variable

560
00:27:02.680 --> 00:27:03.599
<v Speaker 2>you're trying to predict.

561
00:27:03.880 --> 00:27:06.359
<v Speaker 1>So it's like making sure you're slicing your data along

562
00:27:06.440 --> 00:27:09.000
<v Speaker 1>the right lines exactly, so the models you create are

563
00:27:09.039 --> 00:27:12.559
<v Speaker 1>actually meaningful and relevant. That's it. Now, once you've trained

564
00:27:12.559 --> 00:27:15.480
<v Speaker 1>all these models using the MMSA, you still need to

565
00:27:15.480 --> 00:27:18.039
<v Speaker 1>deploy them right you do, and that can get pretty

566
00:27:18.079 --> 00:27:22.079
<v Speaker 1>complex when you're dealing with thousands of models. It's just one, yeah,

567
00:27:22.119 --> 00:27:24.079
<v Speaker 1>so how do you manage all of that?

568
00:27:24.799 --> 00:27:28.599
<v Speaker 2>The book really stresses the importance of automation here. It

569
00:27:28.680 --> 00:27:31.680
<v Speaker 2>introduces us to a tool called Azure Data.

570
00:27:31.480 --> 00:27:33.799
<v Speaker 1>Factory as your data Factory, which.

571
00:27:33.640 --> 00:27:37.160
<v Speaker 2>Can help orchestrate complex data flows, connect to different data sources,

572
00:27:37.200 --> 00:27:40.599
<v Speaker 2>transform data, and even trigger those mL pipelines that we

573
00:27:40.640 --> 00:27:41.599
<v Speaker 2>talked about earlier.

574
00:27:41.799 --> 00:27:45.000
<v Speaker 1>So if the mL pipelines are like the assembly lines

575
00:27:45.000 --> 00:27:48.359
<v Speaker 1>for our models. Yeah, then data Factory is like the

576
00:27:48.400 --> 00:27:49.359
<v Speaker 1>logistics manager.

577
00:27:50.160 --> 00:27:51.319
<v Speaker 2>That's a great way to put it.

578
00:27:51.279 --> 00:27:55.079
<v Speaker 1>Making sure everything runs smoothly, yes, okay. And in addition

579
00:27:55.079 --> 00:27:58.240
<v Speaker 1>to all this, the book is packed with helpful tips

580
00:27:58.279 --> 00:28:00.759
<v Speaker 1>and tricks for getting the most out of AutoML no

581
00:28:00.799 --> 00:28:02.119
<v Speaker 1>matter what kind of problem you're working on.

582
00:28:02.279 --> 00:28:04.440
<v Speaker 2>It is. It's full of practical advice.

583
00:28:04.559 --> 00:28:07.200
<v Speaker 1>It's like having a seasoned auto mL expert looking over

584
00:28:07.200 --> 00:28:08.759
<v Speaker 1>your shoulder guiding you along the way.

585
00:28:09.319 --> 00:28:13.960
<v Speaker 2>Absolutely, and the book really encourages that spirit of experimentation, right.

586
00:28:13.960 --> 00:28:15.680
<v Speaker 1>It's all about trying new things.

587
00:28:15.480 --> 00:28:16.960
<v Speaker 2>Exactly, being a data detective.

588
00:28:17.079 --> 00:28:19.079
<v Speaker 1>That's what I love about data science. It's not just

589
00:28:19.119 --> 00:28:21.880
<v Speaker 1>about following rules, it's about exploring and discovering.

590
00:28:21.960 --> 00:28:22.519
<v Speaker 2>Absolutely.

591
00:28:22.559 --> 00:28:25.440
<v Speaker 1>Now we've talked a lot about Azure Machine Learning Studio,

592
00:28:25.960 --> 00:28:30.839
<v Speaker 1>which has that more visual, user friendly interface, but for

593
00:28:30.920 --> 00:28:34.000
<v Speaker 1>those who are comfortable with coding, the book also covers

594
00:28:34.000 --> 00:28:37.160
<v Speaker 1>the Azure Machine Learning SDK for Python. It does, yes,

595
00:28:37.440 --> 00:28:39.960
<v Speaker 1>so for those who are comfortable coding, there's a way

596
00:28:40.000 --> 00:28:43.720
<v Speaker 1>to get even more control and flexibility over the AutoML process.

597
00:28:43.880 --> 00:28:47.680
<v Speaker 1>Exactly okay, And the book walks us through using Jupiter

598
00:28:47.759 --> 00:28:50.440
<v Speaker 1>Notebooks within Azure machine Learning Yes, which is a very

599
00:28:50.480 --> 00:28:54.680
<v Speaker 1>popular way to write and execute Python code for data

600
00:28:54.680 --> 00:28:55.440
<v Speaker 1>science tasks.

601
00:28:55.519 --> 00:28:57.160
<v Speaker 2>It is very popular.

602
00:28:56.799 --> 00:29:00.240
<v Speaker 1>And you can actually use those powerful compute clusters that

603
00:29:00.240 --> 00:29:03.559
<v Speaker 1>we talked about earlier to run your code in the cloud, Yes,

604
00:29:03.599 --> 00:29:06.839
<v Speaker 1>which gives you access to a ton of processing power. Absolutely,

605
00:29:07.039 --> 00:29:09.680
<v Speaker 1>it really is amazing how cloud computing has made these

606
00:29:09.720 --> 00:29:12.559
<v Speaker 1>complex tasks so much more accessible.

607
00:29:12.599 --> 00:29:13.119
<v Speaker 2>It really is.

608
00:29:13.240 --> 00:29:15.440
<v Speaker 1>Now I'm curious about the different ways you can actually

609
00:29:15.480 --> 00:29:18.960
<v Speaker 1>deploy models once they're trained. Okay, we've talked about batch

610
00:29:18.960 --> 00:29:22.000
<v Speaker 1>scoring and real time scoring, but the book also mentions

611
00:29:22.000 --> 00:29:23.640
<v Speaker 1>something called mL pipelines.

612
00:29:24.240 --> 00:29:28.160
<v Speaker 2>mL pipelines are a fantastic way to automate your entire

613
00:29:28.279 --> 00:29:32.680
<v Speaker 2>machine learning workflow, from data preprofitsing and feature engineering to

614
00:29:32.799 --> 00:29:34.440
<v Speaker 2>model training and deployment.

615
00:29:34.519 --> 00:29:36.720
<v Speaker 1>So it's like building an assembly line for your AI,

616
00:29:36.880 --> 00:29:40.319
<v Speaker 1>making sure each step is done in the right order, exactly,

617
00:29:40.480 --> 00:29:45.000
<v Speaker 1>the right settings, automation, efficiency, consistency. It sounds like a

618
00:29:45.000 --> 00:29:46.680
<v Speaker 1>well oiled machine exactly.

619
00:29:46.799 --> 00:29:49.519
<v Speaker 2>And to take it even further, the book introduces another

620
00:29:49.559 --> 00:29:52.480
<v Speaker 2>powerful tool as your Data Factory.

621
00:29:52.720 --> 00:29:54.240
<v Speaker 1>As your data Factory okay.

622
00:29:54.039 --> 00:29:57.559
<v Speaker 2>This cloud based data integration service can handle even more

623
00:29:57.640 --> 00:30:02.119
<v Speaker 2>complex data flows. Okay, connected different data sources transform your

624
00:30:02.160 --> 00:30:04.920
<v Speaker 2>data and even trigger your mL pipelines.

625
00:30:05.319 --> 00:30:09.119
<v Speaker 1>So if mL pipelines are the assembly lines, then as

626
00:30:09.119 --> 00:30:12.039
<v Speaker 1>your data factory is like the logistics manager making sure

627
00:30:12.079 --> 00:30:14.799
<v Speaker 1>all the raw materials and finished products are flowing smoothly.

628
00:30:15.000 --> 00:30:16.519
<v Speaker 2>Perfect analogy, right.

629
00:30:16.400 --> 00:30:18.759
<v Speaker 1>It's all about streamlining and automating as much of the

630
00:30:18.799 --> 00:30:21.400
<v Speaker 1>process as possible. It is now, we've covered a lot

631
00:30:21.440 --> 00:30:24.200
<v Speaker 1>of ground here, from the nitty gritty details of data

632
00:30:24.200 --> 00:30:28.000
<v Speaker 1>preparation and feature engineering to the broader concepts of Muddel

633
00:30:28.039 --> 00:30:30.759
<v Speaker 1>deployment and automation. We have, but let's not forget about

634
00:30:30.759 --> 00:30:35.279
<v Speaker 1>one crucial aspect, the human element. Building a model is

635
00:30:35.319 --> 00:30:38.160
<v Speaker 1>only part of the story. It is what about gaining

636
00:30:38.240 --> 00:30:40.880
<v Speaker 1>the trust and buy in of the people who will

637
00:30:40.920 --> 00:30:42.359
<v Speaker 1>actually be using these models.

638
00:30:42.519 --> 00:30:43.960
<v Speaker 2>That's a great point. You know, we can build the

639
00:30:44.000 --> 00:30:46.880
<v Speaker 2>most sophisticated AI in the world, but if people don't

640
00:30:46.920 --> 00:30:48.920
<v Speaker 2>trust it or understand how.

641
00:30:48.759 --> 00:30:51.279
<v Speaker 1>It works, it's not going to be very useful.

642
00:30:51.039 --> 00:30:54.000
<v Speaker 2>Exactly, and that's why those explainability features that we keep

643
00:30:54.039 --> 00:30:55.920
<v Speaker 2>talking about are so important.

644
00:30:56.119 --> 00:30:59.799
<v Speaker 1>Right. Transparency is key. It is, especially when you're dealing

645
00:30:59.839 --> 00:31:02.839
<v Speaker 1>with decisions that could have a real impact on people's lives.

646
00:31:03.200 --> 00:31:07.799
<v Speaker 2>Absolutely. The book emphasizes the need to be able to

647
00:31:07.960 --> 00:31:12.079
<v Speaker 2>clearly articulate how a model is moving decisions, especially in

648
00:31:12.119 --> 00:31:15.480
<v Speaker 2>those industries with strict regulations or where the stakes are.

649
00:31:15.440 --> 00:31:17.279
<v Speaker 1>High, right, like healthcare or finance.

650
00:31:17.480 --> 00:31:18.039
<v Speaker 2>Exactly.

651
00:31:18.279 --> 00:31:21.480
<v Speaker 1>You need to be able to explain why the model

652
00:31:21.599 --> 00:31:25.960
<v Speaker 1>is recommending a certain treatment or making a certain investment decision.

653
00:31:26.200 --> 00:31:29.079
<v Speaker 1>You do. It's not enough to just say the computer

654
00:31:29.200 --> 00:31:30.039
<v Speaker 1>says so right.

655
00:31:30.160 --> 00:31:31.119
<v Speaker 2>We need more than that.

656
00:31:31.319 --> 00:31:32.799
<v Speaker 1>You need to be able to back it up with

657
00:31:32.880 --> 00:31:34.640
<v Speaker 1>evidence and insights.

658
00:31:34.279 --> 00:31:37.160
<v Speaker 2>And AUTOMML gives us the tools to do just that.

659
00:31:37.160 --> 00:31:38.440
<v Speaker 1>That's good. You know.

660
00:31:38.480 --> 00:31:40.960
<v Speaker 2>You can use those feature important scores to see which

661
00:31:41.000 --> 00:31:44.240
<v Speaker 2>factors are driving the model's predictions. Okay, you can even

662
00:31:44.319 --> 00:31:47.359
<v Speaker 2>drill down into those individual predictions to understand why the

663
00:31:47.359 --> 00:31:49.000
<v Speaker 2>model made a specific decision.

664
00:31:49.240 --> 00:31:52.240
<v Speaker 1>So it's like having this transparent AI exactly where you

665
00:31:52.240 --> 00:31:53.960
<v Speaker 1>can peek under the hood and see what's going on.

666
00:31:54.079 --> 00:31:54.440
<v Speaker 2>That's it.

667
00:31:54.759 --> 00:31:56.880
<v Speaker 1>Now, before we wrap up, I want to highlight something

668
00:31:56.920 --> 00:31:59.079
<v Speaker 1>that really stood out to me while reading the book. Okay,

669
00:31:59.319 --> 00:32:02.160
<v Speaker 1>it mentions that AUTOMML can actually be used in other

670
00:32:02.279 --> 00:32:07.000
<v Speaker 1>Microsoft products besides just Azure Machine Learning Studio. Yes, so

671
00:32:07.039 --> 00:32:09.759
<v Speaker 1>it's not just limited to this one platform, not at all.

672
00:32:09.920 --> 00:32:13.480
<v Speaker 2>Tell me more So, the book talks about integrating AutoML

673
00:32:13.559 --> 00:32:18.079
<v Speaker 2>with tools like powerbi Powerbio, which is that powerful data

674
00:32:18.160 --> 00:32:20.880
<v Speaker 2>visualization and business intelligence platform.

675
00:32:21.039 --> 00:32:24.480
<v Speaker 1>So you could create these interactive dashboards that not only

676
00:32:24.559 --> 00:32:27.599
<v Speaker 1>display the data, yeah, but also use auto mL to

677
00:32:27.640 --> 00:32:29.240
<v Speaker 1>generate predictions and insights.

678
00:32:29.920 --> 00:32:30.480
<v Speaker 2>Exactly.

679
00:32:30.640 --> 00:32:32.799
<v Speaker 1>That's amazing. It's like taking data storytelling to a whole

680
00:32:32.799 --> 00:32:33.240
<v Speaker 1>new level.

681
00:32:33.279 --> 00:32:35.960
<v Speaker 2>It is. And the book also touches on using auto

682
00:32:36.079 --> 00:32:38.720
<v Speaker 2>mL with Azure synaps Analytics.

683
00:32:38.400 --> 00:32:43.400
<v Speaker 1>Okay, which is that cloud based data warehousing and analytics service. Right,

684
00:32:43.480 --> 00:32:46.319
<v Speaker 1>So this opens up even more possibilities for working with

685
00:32:46.559 --> 00:32:50.720
<v Speaker 1>massive data sets, right and building these enterprise scale AI

686
00:32:50.799 --> 00:32:54.279
<v Speaker 1>solutions Exactly. It sounds like the possibilities are pretty much endless.

687
00:32:54.680 --> 00:32:55.039
<v Speaker 1>They are.

688
00:32:55.200 --> 00:32:56.680
<v Speaker 2>The possibilities are endless.

689
00:32:57.079 --> 00:32:59.720
<v Speaker 1>We've covered so much ground, but it feels like we've

690
00:32:59.759 --> 00:33:03.559
<v Speaker 1>only just scratch the surface of what AutoML on Azure

691
00:33:03.599 --> 00:33:03.960
<v Speaker 1>can do.

692
00:33:04.160 --> 00:33:06.279
<v Speaker 2>We've just scratched the surface.

693
00:33:06.359 --> 00:33:08.160
<v Speaker 1>So as we move into the final part of our

694
00:33:08.200 --> 00:33:12.440
<v Speaker 1>deep dive, Okay, I'm curious where should someone start if

695
00:33:12.440 --> 00:33:15.720
<v Speaker 1>they want to explore this world of AutoML on Azure.

696
00:33:15.880 --> 00:33:18.480
<v Speaker 2>Well, this book we've been discussing. Is an excellent place

697
00:33:18.519 --> 00:33:19.039
<v Speaker 2>to begin.

698
00:33:19.039 --> 00:33:20.359
<v Speaker 1>Right, It's a great starting point.

699
00:33:20.400 --> 00:33:24.279
<v Speaker 2>Dennis Sawyers did a phenomenal job creating this really practical

700
00:33:24.359 --> 00:33:25.920
<v Speaker 2>and easy to follow guide.

701
00:33:26.000 --> 00:33:26.559
<v Speaker 1>I agree.

702
00:33:26.640 --> 00:33:30.359
<v Speaker 2>It's packed with real world examples, code snippets, and tons

703
00:33:30.359 --> 00:33:31.319
<v Speaker 2>of helpful advice.

704
00:33:31.400 --> 00:33:33.839
<v Speaker 1>And of course there are tons of online resources out there,

705
00:33:33.880 --> 00:33:37.680
<v Speaker 1>including Microsoft's owned documentation and tutorials. There are plus the

706
00:33:37.680 --> 00:33:40.559
<v Speaker 1>Azure community is incredibly active and supportive. It is.

707
00:33:40.599 --> 00:33:42.039
<v Speaker 2>It's a great community, So.

708
00:33:41.960 --> 00:33:44.279
<v Speaker 1>If you have questions, you'll definitely find someone who can

709
00:33:44.319 --> 00:33:47.519
<v Speaker 1>help you will. I think the key takeaway here is

710
00:33:47.559 --> 00:33:51.599
<v Speaker 1>that auto mL is not some futuristic concept. No it's not.

711
00:33:51.759 --> 00:33:53.559
<v Speaker 1>It's here. It's here now, it's now.

712
00:33:53.680 --> 00:33:54.680
<v Speaker 2>It's a powerful tool.

713
00:33:54.799 --> 00:33:57.279
<v Speaker 1>It's a powerful tool that's available right now.

714
00:33:57.079 --> 00:33:59.440
<v Speaker 2>And it's making AI more accessible than ever before.

715
00:33:59.559 --> 00:34:03.720
<v Speaker 1>Exactly, and with the power of AUTOMML on Azure, anyone

716
00:34:03.720 --> 00:34:05.039
<v Speaker 1>could become an AI innovator.

717
00:34:05.079 --> 00:34:05.640
<v Speaker 2>Anyone can.

718
00:34:05.960 --> 00:34:11.559
<v Speaker 1>So to our listener, we challenge you, what problem will

719
00:34:11.599 --> 00:34:12.840
<v Speaker 1>you solve with auto mail?

720
00:34:13.079 --> 00:34:13.800
<v Speaker 2>That's question?

721
00:34:13.840 --> 00:34:14.920
<v Speaker 1>What will you create?

722
00:34:15.000 --> 00:34:15.800
<v Speaker 2>What will you create?

723
00:34:16.480 --> 00:34:19.000
<v Speaker 1>We'll be back in a moment with the final part

724
00:34:19.119 --> 00:34:27.199
<v Speaker 1>of our deep dive into auto mL on Asure welcome

725
00:34:27.199 --> 00:34:29.079
<v Speaker 1>back to the final part of our deep dive into

726
00:34:29.159 --> 00:34:32.719
<v Speaker 1>auto mL on Asure. You know, this book has really

727
00:34:32.760 --> 00:34:35.480
<v Speaker 1>opened my eyes to the potential of auto mL.

728
00:34:35.559 --> 00:34:36.920
<v Speaker 2>It really is a game changer, and.

729
00:34:36.880 --> 00:34:39.719
<v Speaker 1>Azure seems like the perfect platform. Oh, absolutely to really

730
00:34:39.760 --> 00:34:41.519
<v Speaker 1>explore it. It is now for those who want to

731
00:34:41.559 --> 00:34:44.360
<v Speaker 1>kind of delve a little deeper into those technical aspects. Yeah,

732
00:34:44.480 --> 00:34:46.760
<v Speaker 1>the book actually doesn't shy away from the code, No

733
00:34:46.840 --> 00:34:49.719
<v Speaker 1>it doesn't. It guides you through using the Azure Machine

734
00:34:49.760 --> 00:34:52.760
<v Speaker 1>Learning SDK for Python. It does so for those who

735
00:34:52.840 --> 00:34:55.639
<v Speaker 1>are comfortable with coding, there's a way to get even

736
00:34:55.719 --> 00:34:59.000
<v Speaker 1>more granular control over the auto mL process.

737
00:34:59.119 --> 00:35:01.280
<v Speaker 2>Absolutely, the book that shows you how to use Jupiter

738
00:35:01.360 --> 00:35:04.760
<v Speaker 2>notebooks within Azure machine Learning, which is a really popular

739
00:35:04.800 --> 00:35:08.079
<v Speaker 2>way to write and execute Python code for data science.

740
00:35:08.159 --> 00:35:10.679
<v Speaker 1>And you can even use those powerful compute clusters that

741
00:35:10.679 --> 00:35:13.159
<v Speaker 1>we talked about yes to run your code in the

742
00:35:13.199 --> 00:35:16.920
<v Speaker 1>cloud exactly, tapping into massive processing power.

743
00:35:16.719 --> 00:35:18.119
<v Speaker 2>And all that processing power.

744
00:35:18.239 --> 00:35:21.960
<v Speaker 1>It's incredible how cloud computing has really made these complex

745
00:35:22.039 --> 00:35:23.639
<v Speaker 1>tasks so much more accessible.

746
00:35:23.760 --> 00:35:24.360
<v Speaker 2>It really has.

747
00:35:24.719 --> 00:35:27.119
<v Speaker 1>I'm curious about the practical side of things. Okay, we've

748
00:35:27.159 --> 00:35:29.440
<v Speaker 1>trained our models, but how do we actually put them

749
00:35:29.440 --> 00:35:32.719
<v Speaker 1>to work right. The book mentions batch scoring and real

750
00:35:32.760 --> 00:35:36.320
<v Speaker 1>time scoring, but it also talks about something called mL pipelines.

751
00:35:36.480 --> 00:35:40.039
<v Speaker 2>Yes, mL pipelines are a fantastic way to automate that

752
00:35:40.320 --> 00:35:44.360
<v Speaker 2>entire machine learning workflow. Okay, from data preprocessing and feature

753
00:35:44.400 --> 00:35:47.639
<v Speaker 2>engineering to model training and deployment.

754
00:35:47.840 --> 00:35:50.840
<v Speaker 1>So it's like creating an assembly line for your AI exactly.

755
00:35:50.920 --> 00:35:53.519
<v Speaker 1>Each step is executed in the right order with the

756
00:35:53.599 --> 00:35:54.159
<v Speaker 1>right settings.

757
00:35:54.280 --> 00:35:57.599
<v Speaker 2>Yes, automation, efficiency, consistency.

758
00:35:57.840 --> 00:36:01.239
<v Speaker 1>Yeah, it sounds like a well machine. It is. And

759
00:36:01.280 --> 00:36:03.639
<v Speaker 1>to take even a step further, the book introduces us

760
00:36:03.679 --> 00:36:07.840
<v Speaker 1>to Azure Data Factory. Yes, another great tool, this cloud

761
00:36:07.880 --> 00:36:11.519
<v Speaker 1>based data integration service that can handle even more complex

762
00:36:11.599 --> 00:36:15.159
<v Speaker 1>data flows. It can connect to different data sources, transform

763
00:36:15.199 --> 00:36:18.920
<v Speaker 1>your data, and even trigger those mL pipelines exactly. So

764
00:36:18.960 --> 00:36:21.639
<v Speaker 1>if mL pipelines are the assembly lines, then az your

765
00:36:21.719 --> 00:36:24.559
<v Speaker 1>Data factory is like the logistics manager making sure that

766
00:36:24.599 --> 00:36:25.960
<v Speaker 1>everything's flowing smoothly.

767
00:36:26.119 --> 00:36:27.199
<v Speaker 2>A perfect analogy.

768
00:36:27.440 --> 00:36:30.719
<v Speaker 1>Now, this book has given us a really fantastic overview

769
00:36:30.760 --> 00:36:33.960
<v Speaker 1>of auto mL on Azure. It has, and it's clear

770
00:36:34.360 --> 00:36:37.639
<v Speaker 1>that this technology has the potential to really revolutionize the

771
00:36:37.639 --> 00:36:38.599
<v Speaker 1>way we approach AI.

772
00:36:38.880 --> 00:36:39.679
<v Speaker 2>It really does.

773
00:36:39.840 --> 00:36:42.400
<v Speaker 1>But as we wrap up our deep dive. Yeah, what's

774
00:36:42.440 --> 00:36:45.199
<v Speaker 1>the one key message you hope our listener takes away

775
00:36:45.199 --> 00:36:45.840
<v Speaker 1>from all of this.

776
00:36:46.519 --> 00:36:49.440
<v Speaker 2>I think the biggest takeaway is that auto mL is empowering.

777
00:36:49.880 --> 00:36:52.880
<v Speaker 2>It breaks down those barriers to entry for AI, making

778
00:36:52.880 --> 00:36:55.880
<v Speaker 2>it accessible to a much wider audience. I agree, you

779
00:36:55.920 --> 00:36:58.159
<v Speaker 2>don't need to be a data science guru to build

780
00:36:58.199 --> 00:37:01.960
<v Speaker 2>powerful models on Azure. This book provides the guidance and

781
00:37:02.000 --> 00:37:05.679
<v Speaker 2>the tools you need to get started, and with the power.

782
00:37:05.400 --> 00:37:09.360
<v Speaker 1>Of auto mL on Azure, anyone can become an AI innovator.

783
00:37:09.440 --> 00:37:13.519
<v Speaker 1>Absolutely so to our listeners, we encourage you dive in,

784
00:37:14.079 --> 00:37:16.679
<v Speaker 1>explore the world of auto mL and see what incredible

785
00:37:16.679 --> 00:37:17.599
<v Speaker 1>things you can achieve.

786
00:37:17.800 --> 00:37:19.559
<v Speaker 2>Until next time, happy innovating.