WEBVTT
1
00:00:03.120 --> 00:00:14.130
Michelle :: Webinar Producer: Well, good morning good afternoon Hello everyone, welcome to today's webinar titled basics every machine learning practitioner should know in.
2
00:00:15.809 --> 00:00:17.430
Michelle :: Webinar Producer: My name is Michelle i'm here to.
3
00:00:17.430 --> 00:00:19.140
Michelle :: Webinar Producer: produce the session.
4
00:00:19.680 --> 00:00:26.880
Michelle :: Webinar Producer: And before we jump into our introductions I want to give a quick tour of how to navigate the virtual platform.
5
00:00:27.300 --> 00:00:39.540
Michelle :: Webinar Producer: So, as you joined, you probably notice that your microphones are disabled, so if you have any questions comments they are more than welcome, they are invited in the chat icon.
6
00:00:40.050 --> 00:00:53.190
Michelle :: Webinar Producer: If you take your mouse and hover anywhere inside this platform you'll see a toolbar at the bottom of your screen one of those icons is labelled chat click on that to find the chat window and post your questions throughout the session.
7
00:00:54.540 --> 00:01:00.390
Michelle :: Webinar Producer: We are recording this session and we'll email a copy out to everyone who registered by the end of the week.
8
00:01:03.420 --> 00:01:16.440
Michelle :: Webinar Producer: Our presenter today is Chris screen UK Chris is the global machine learning technical lead for the Amazon partner network commonly referred to as the ap n.
9
00:01:16.920 --> 00:01:34.320
Michelle :: Webinar Producer: He co founded the machine learning group for the API and in 2017 Chris started his career as a quantitative strategists at Goldman Sachs developed a predictive maintenance Apps as an Ai engineer at ATT.
10
00:01:34.830 --> 00:01:50.310
Michelle :: Webinar Producer: And prior to Amazon was computer vision architect at gopro very cool portfolio He founded and sold to Silicon Valley startups in finance and networking monitoring clear station and site rock.
11
00:01:50.970 --> 00:01:59.040
Michelle :: Webinar Producer: As a technical lead for horizontal support of Ai ml he's the first line of support for aws partners.
12
00:01:59.250 --> 00:02:17.670
Michelle :: Webinar Producer: Looking to develop or expand their use of technologies across all markets, including finance healthcare and life sciences, public services, telecommunications and E commerce all right, where you are in for an excellent webinar today and Chris, the floor is all yours.
13
00:02:18.150 --> 00:02:29.130
Kris Skrinak: Well, thank you very much for that introduction i'm very happy to be here today um you know I, I would like to start off by just saying we're at a.
14
00:02:29.940 --> 00:02:40.440
Kris Skrinak: At a very special moment in machine learning where an enormous amount of highly you know, important and break through code has matured.
15
00:02:41.070 --> 00:02:53.910
Kris Skrinak: And because of that, there are certain things now that really every practitioner needs to know it doesn't matter sort of what your role is whether you're a developer, a data scientist, whether you do devops.
16
00:02:54.690 --> 00:03:03.930
Kris Skrinak: And if you're just getting started in the field, then this is what I consider sort of my best starting point for folks who really want to get into ml.
17
00:03:04.680 --> 00:03:16.140
Kris Skrinak: First of all, there's good reason to panic when you just take a look at the services that are available on aws well the way we describe it as with this three tier stack.
18
00:03:16.710 --> 00:03:26.340
Kris Skrinak: And on the top level, we have API services, these are services like every other service on aws like our storage s3 or our instances etc to.
19
00:03:26.820 --> 00:03:37.470
Kris Skrinak: What I mean by that is they're accessible through the graphic user interface in the console they're available through your command line interface just using bash or a windows.
20
00:03:39.330 --> 00:03:46.860
Kris Skrinak: panel and they're also available in a higher and lower level SDK so when it comes to that top level of the stack.
21
00:03:47.100 --> 00:03:59.430
Kris Skrinak: i'm really going to be focusing quite a bit on that today, because at this point, you know, no one would think you were a storage expert if you knew how to get a file out of deep storage or glaciers we call it on s3.
22
00:03:59.850 --> 00:04:04.260
Kris Skrinak: it's you do it through the user interface you'd make a simple call from the coi.
23
00:04:04.500 --> 00:04:14.010
Kris Skrinak: no big deal right and the analogy i'm trying to make here is that that's exactly where we are with machine learning today when it comes to most of this three layer stack.
24
00:04:14.370 --> 00:04:25.740
Kris Skrinak: The exception to that rule rule is the Center Amazon sage maker, is the seminal end to end machine learning bespoke model development ID.
25
00:04:26.010 --> 00:04:33.510
Kris Skrinak: And it is a breakthrough product, it has so many sub components that any data scientist who really wants to get into.
26
00:04:33.810 --> 00:04:39.960
Kris Skrinak: Creating high performance models models that not just run in the cloud but might run on a PC.
27
00:04:40.320 --> 00:04:51.600
Kris Skrinak: or MAC on android or ios I mean it is the place where you make these models i'm not going to cover much of that today because frankly there's a lot of material out there.
28
00:04:51.810 --> 00:05:04.080
Kris Skrinak: And it's pretty advanced and sometimes you need a couple of weeks or months of training to really begin to understand that, but I think we're going to be able to create a bridge for that today to on the bottom layer of that.
29
00:05:04.860 --> 00:05:16.650
Kris Skrinak: stack three layer stack we have our infrastructure, now we include in that infrastructure of some of the deep learning frameworks like tensorflow pi torch and apac Apache mx net.
30
00:05:16.980 --> 00:05:25.950
Kris Skrinak: We make it very easy to use those frameworks for data scientists, by putting them in images that you can load up on an EC to or, of course, load up.
31
00:05:26.340 --> 00:05:34.920
Kris Skrinak: with various configurations in sage maker itself, we also have the most robust set of gpu instances out there.
32
00:05:35.130 --> 00:05:44.100
Kris Skrinak: We actually make our own chips now that are specifically designed for machine learning training and influence influences when that model goes into production.
33
00:05:44.310 --> 00:05:56.070
Kris Skrinak: We also have something called elastic inference So if you just have an easy to instance, you know, one of the you know $5 an hour ones, are they are really five cent I should say an hour one's really cheap.
34
00:05:56.370 --> 00:06:10.080
Kris Skrinak: instance, you can attach a GP you to it just by clicking a checkbox in the gooey recall that elastic inference and we also have field programmable gate arrays that's a lot but wait there's more.
35
00:06:10.710 --> 00:06:19.950
Kris Skrinak: machine learning is really just software development, and it is found its way into almost every aspect of the services that aws importantly.
36
00:06:20.310 --> 00:06:26.280
Kris Skrinak: Our relational database systems like aurora Athena redshift and Neptune and some of our.
37
00:06:26.640 --> 00:06:36.330
Kris Skrinak: relational database partners such as snowflake can easily now automatically create models with sequel then there's glue data Bruce, this is a breakthrough.
38
00:06:37.200 --> 00:06:42.930
Kris Skrinak: Data prep tool that has a parallel in Stage maker called data wrangler.
39
00:06:43.770 --> 00:06:59.610
Kris Skrinak: You know our business intelligence suite quick site, we have something called ml insights which not only can understand language, and you can do ad hoc natural language queries to create reports or charts, but we also.
40
00:06:59.970 --> 00:07:10.080
Kris Skrinak: can generate language based off of the data that you're using so it's a data generation tool as well and I mentioned, training and infrared GR custom chips.
41
00:07:10.830 --> 00:07:19.200
Kris Skrinak: But brief if that got you a little bit excited it's going to be okay, in fact, they are evergreen patterns and principles that i'm going to walk through today.
42
00:07:19.500 --> 00:07:27.900
Kris Skrinak: And once you notice them you'll see them throughout the stack so it really makes everything that I just mentioned, a lot more accessible So what are we going to learn.
43
00:07:28.620 --> 00:07:37.440
Kris Skrinak: First of all, i'm going to tell you machine learning dark secret i'm going to give you some information on how to prioritize requests for machine learning.
44
00:07:37.680 --> 00:07:53.100
Kris Skrinak: Development, what are the roles in the life cycle, what is the basic terminology, when I found with executives decision makers, sometimes All they need to know is to understand the words that are being used in a field and that's fit their 50% there.
45
00:07:53.880 --> 00:08:04.020
Kris Skrinak: what's possible without prior machine learning training and that's what we're going to focus on quite a bit what what kind of models, can you create with no little or no code.
46
00:08:04.380 --> 00:08:14.550
Kris Skrinak: How does devops work with machine learning what is auto ml what our solutions that are ready, out of the box and when do you actually need a specialist so let's break this down.
47
00:08:14.760 --> 00:08:27.660
Kris Skrinak: You may already know much more than you think here it comes machine learnings dark secret is it's not about the algorithms machine learning is really simple when you look at it from this perspective.
48
00:08:28.110 --> 00:08:35.070
Kris Skrinak: Data writes code now I say that and it's very simplistic but it's there's a lot of truth here.
49
00:08:35.460 --> 00:08:44.280
Kris Skrinak: When you're actually creating a machine learning model what you're doing is you're taking data and you're matching it to an algorithm now let's talk about those algorithms for a moment.
50
00:08:44.640 --> 00:08:53.430
Kris Skrinak: there's a bunch of kids out there Python mostly but also are even in Java now where all these algorithms are already there like.
51
00:08:53.880 --> 00:09:05.820
Kris Skrinak: You know graph and sort say and in unix you wouldn't go through the effort of rewriting those utilities you shouldn't do the same in machine learning, it is so robust, now it is so mature and so rich.
52
00:09:06.150 --> 00:09:13.410
Kris Skrinak: That really a smaller and smaller fraction of machine learning tasks actually require writing an algorithm from scratch.
53
00:09:13.890 --> 00:09:20.730
Kris Skrinak: But without this principle, this notion that data is going to write the code for you it's really easy to get lost.
54
00:09:20.970 --> 00:09:31.560
Kris Skrinak: So with you start with this basic principle, then the very next step is to say, well, what really makes machine learning code different from say Java c++ or things like that.
55
00:09:31.890 --> 00:09:38.340
Kris Skrinak: And that is that every output of a machine learning model is a prediction, and because that's true.
56
00:09:38.850 --> 00:09:47.370
Kris Skrinak: you're always going to come up with some accuracy measurement for its use, it could be high, it could be in the high 90s 95% 97%.
57
00:09:48.090 --> 00:09:55.230
Kris Skrinak: I have seen some models go into production with as low as 70 and 80% it really depends on the use case right if you're.
58
00:09:55.410 --> 00:10:05.760
Kris Skrinak: Trying to detect whether tumor is cancerous or not, you might want to have really high 90s if you're trying to select the next product that someone might want on your website, you might go for a lower figure.
59
00:10:06.060 --> 00:10:13.170
Kris Skrinak: But basically all code written with machine learning is going to be predictive in nature, in fact, I often say that.
60
00:10:13.470 --> 00:10:21.030
Kris Skrinak: Ai and machine learning itself or misnomers this whole field should really just be called prediction but it's a lot sexier.
61
00:10:21.330 --> 00:10:29.790
Kris Skrinak: I also want to discriminate very carefully between analytics and machine learning now analytics predictive analytics or the phrase that we use here.
62
00:10:30.360 --> 00:10:33.780
Kris Skrinak: augmented analytics has been around for quite some time.
63
00:10:34.410 --> 00:10:45.630
Kris Skrinak: And when you break this down a lot of this is just based on a lot of simple statistical methods, with the possible exception of deep learning, which has you know another the neural network aspect to it.
64
00:10:45.990 --> 00:10:58.680
Kris Skrinak: But we're going to talk about machine learning and the way I described this difference once again half joking is that analytics creates reports machine learning creates endpoints so we're going to write code with machine learning.
65
00:10:59.010 --> 00:11:09.450
Kris Skrinak: And if you get lost as I go through this just remember data writes code data is absolutely paramount and, as you go through the life cycle of a machine learning model how its.
66
00:11:09.750 --> 00:11:18.300
Kris Skrinak: Developed you'll see frequently people say 80% of their time is spent wrangling, the data well, that is not wasted time because.
67
00:11:18.540 --> 00:11:26.670
Kris Skrinak: you're a data scientist perhaps or a programmer that's using some of these tools for the very first time, you may not be the domain expert.
68
00:11:26.880 --> 00:11:39.900
Kris Skrinak: The domain expert knows a lot about that data, you might not so that 50 to 70% of your time that's spent wrangling the code and getting it ready to fit to an algorithm is not wasted time it's an absolutely.
69
00:11:40.140 --> 00:11:46.620
Kris Skrinak: essential part of the process, so let's talk about that process we start by collecting training data.
70
00:11:46.920 --> 00:11:54.810
Kris Skrinak: One of the things I often say about machine learning models is every machine learning model is obsolete on production So how could that be true.
71
00:11:55.050 --> 00:12:10.440
Kris Skrinak: it's true because that model was written with historical data once you put it into production, you have fresh data, you need to monitor it i'll talk about that in a minute, but that's where we get into the devops and ml OPS aspect of this but data is absolutely key.
72
00:12:10.980 --> 00:12:17.160
Kris Skrinak: Then you choose an algorithm here's another dirty little secret there's basically only three things you can do with machine learning.
73
00:12:17.520 --> 00:12:26.460
Kris Skrinak: That is predicted number predict a class or discover a class sounds really simple that is regression or classification and clustering respectively.
74
00:12:26.730 --> 00:12:34.350
Kris Skrinak: Really three things doom all done almost 90% of machine learning and solve most business problems with machine learning.
75
00:12:34.560 --> 00:12:47.010
Kris Skrinak: So when you go to choose your algorithm that's one of the first questions you ask, am I predicting a number, am I trying to predict a class like cat dog, for example, pedestrian you know things like this and computer vision.
76
00:12:47.850 --> 00:12:52.710
Kris Skrinak: You know customer that that will come back customer that one come back these are all classes.
77
00:12:53.700 --> 00:13:03.060
Kris Skrinak: Then you have to set up and manage that environment train and tune your model deploy that model in production and then you need to scale and manage it in a production environment.
78
00:13:03.330 --> 00:13:12.930
Kris Skrinak: So when you take a look at that big picture, the actual Code, the actual code that creates the algorithm is a really small part of the process.
79
00:13:13.380 --> 00:13:18.930
Kris Skrinak: It is an important part of the process it's kind of like saying the hardest just one organ in the body.
80
00:13:19.230 --> 00:13:28.410
Kris Skrinak: But the reality is this is really the big picture of how machine learning code goes into production and who puts it in production well there's a couple of teams.
81
00:13:28.650 --> 00:13:38.970
Kris Skrinak: You definitely have data, scientists and data engineers and frequently in large companies, these are two independent groups and a startup it might be one person doing everything you see here.
82
00:13:39.270 --> 00:13:47.040
Kris Skrinak: But more frequently data engineers are folks who've been working with relational or graph databases for decades that's their area of expertise.
83
00:13:47.250 --> 00:13:55.440
Kris Skrinak: And they need to communicate with the data scientists, you know where is this data How should it be for formatted, how do you need it formatted for your model.
84
00:13:55.740 --> 00:14:05.490
Kris Skrinak: If it's tabular data, there may be peculiarities so, for example, all machine learning models need numbers to work with, so if you have a lot of words.
85
00:14:05.760 --> 00:14:14.310
Kris Skrinak: You have to vector eyes those words so that's a data prep concept now you put that model into production and you're going to be working with QA.
86
00:14:15.000 --> 00:14:21.090
Kris Skrinak: will be working with machine learning engineers and other developers and the whole set of devops personnel.
87
00:14:21.420 --> 00:14:26.220
Kris Skrinak: what's critical here is that each of these roles understands the goal.
88
00:14:26.520 --> 00:14:32.610
Kris Skrinak: of each other, so they could they can better communicate I won't go through this in detail, and of course we're being recorded here.
89
00:14:32.820 --> 00:14:42.600
Kris Skrinak: But you basically your data scientists data engineer and devops have very narrow responsibilities in this process once again if you're a startup you have to do it all.
90
00:14:43.230 --> 00:14:48.960
Kris Skrinak: So let's talk about the terminology like I said this is 50% of it, beginning right now.
91
00:14:49.680 --> 00:14:59.910
Kris Skrinak: I mentioned regression classification and clustering as the principal goals of almost every machine learning model, and I said it was about 90% so what's in that last 10%.
92
00:15:00.390 --> 00:15:13.530
Kris Skrinak: That is going to be a number of things, but, most importantly, reinforce reinforcement learning reinforcement learning is when you actually learn through trial and error regression and classification are frequently called.
93
00:15:14.460 --> 00:15:25.620
Kris Skrinak: supervised learning clustering is frequently called unsupervised learning, because when you're looking at data you don't know how many classes are in there you have to discover that and you do it through trial and error.
94
00:15:26.550 --> 00:15:34.350
Kris Skrinak: Deep learning simply means that you're employing a neural network which is simply a matrix of nodes of numbers that are.
95
00:15:35.340 --> 00:15:43.200
Kris Skrinak: automatically generated for you, through iteration when you're making a deep learning model will often hear structured and unstructured.
96
00:15:43.710 --> 00:15:55.200
Kris Skrinak: Data structured data just think of it as a spreadsheet rows returned by a relational database and unstructured data is everything else, but this is also principally key.
97
00:15:55.500 --> 00:16:01.920
Kris Skrinak: Thanks to the advancements in deep learning almost everything digital Now you can consider data.
98
00:16:02.190 --> 00:16:12.420
Kris Skrinak: that's usually photographs and videos and audio but anything any sensor data coming out of an iot device what temperature is It is this machine vibrating etc.
99
00:16:12.780 --> 00:16:18.000
Kris Skrinak: This is basically going to be unstructured or sometimes time series data in those last two instances.
100
00:16:18.630 --> 00:16:29.370
Kris Skrinak: People use the word model, a lot at aws we're very careful about discriminating between a model that's in your code that's an algorithm that's.
101
00:16:29.610 --> 00:16:37.440
Kris Skrinak: fifth year data to produce a model and model artifacts model artifacts you can think of is compiled code it's the model.
102
00:16:37.710 --> 00:16:48.000
Kris Skrinak: That was actually generated by training your data and it's the model artifacts that go into production, so we did make a very careful to distinguish how those words are used.
103
00:16:48.810 --> 00:16:59.100
Kris Skrinak: Training is when you, as I mentioned fit that data to an algorithm that is the principal job of a data scientist or a developer who's using some of these higher end tools.
104
00:16:59.520 --> 00:17:10.560
Kris Skrinak: inference is the prediction that's returned when you send fresh data to that model in production that's actually how you're using the model in production and, by the way.
105
00:17:11.010 --> 00:17:22.620
Kris Skrinak: Up to about two or three years ago, most of the cost was associated with training because gpu instances were very expensive now, thanks to aws and cloud models.
106
00:17:23.040 --> 00:17:30.270
Kris Skrinak: such as we have really all the biggest expenses now in in France and that's good that means you created a model.
107
00:17:30.510 --> 00:17:40.530
Kris Skrinak: You put it in production, hopefully you're making money with that model and so it's paying for itself, but now, a lot more attention and cost reduction is on that model in production.
108
00:17:40.920 --> 00:17:41.610
Kris Skrinak: Okay there.
109
00:17:42.060 --> 00:17:42.900
Kris Skrinak: Oh go ahead i'm sorry.
110
00:17:43.080 --> 00:17:48.180
Myles Brown :: Moderator: Just before we move on it, I think there's a question that might actually fit those definitions really well.
111
00:17:48.690 --> 00:18:02.760
Myles Brown :: Moderator: I don't know if you can see, the one it's asking how much labeled data, do you need to create a class identifier, what a couple hundred do it, and then they asked about rephrase seems to only need a few I think it's the difference between training a model and applying model.
112
00:18:04.200 --> 00:18:11.550
Kris Skrinak: Well, so the short answer to that is it depends on the use case and frequently you don't have a lot of training data to begin with.
113
00:18:12.300 --> 00:18:24.660
Kris Skrinak: On aws we have a lot of tools and sage maker, one of them is augmented Ai what augmented Ai does is you take you build a model with whatever data, you have it could be 100 rose.
114
00:18:24.930 --> 00:18:33.780
Kris Skrinak: You put it in production and then, when the results come in at a number lower than his production, where the say 69%.
115
00:18:34.170 --> 00:18:44.670
Kris Skrinak: That data gets set out to a separate file for human labeling so that's really what I recommend for most folks who don't have a lot of data to start with.
116
00:18:44.940 --> 00:18:53.010
Kris Skrinak: And we're moving into this phase now and machine learning where it kind of all the low hanging fruits been taken right and or it's covered by the services that i'm going to go through.
117
00:18:53.280 --> 00:19:02.220
Kris Skrinak: So you know when you don't have a lot of data augmented Ai human in the loop labeling and iterating through model after model after model, you could quickly get up to model.
118
00:19:04.200 --> 00:19:17.520
Kris Skrinak: Using this strategy that I just described, but you can you get a very rich highly accurate training set very, very quickly, but it requires engagement right from all those stakeholders that I mentioned earlier.
119
00:19:18.240 --> 00:19:27.150
Kris Skrinak: All right, it's important to know the names of the tools of the trade, these are the basic tools, the basic terms, you have to know in this field number one anaconda.
120
00:19:27.660 --> 00:19:38.610
Kris Skrinak: Python is the lingua franca of ml and most people are using anaconda anacondas Open Source it can run in the cloud on a machine a current on a little raspberry pi if you want.
121
00:19:38.940 --> 00:19:47.880
Kris Skrinak: But it makes it real easy to change environments to you know do ad hoc where, and if you hear anaconda they're really referring to the Python instance of anaconda.
122
00:19:48.270 --> 00:19:59.220
Kris Skrinak: Jupiter lab is the principal ID for developing machine learning models or orchestrating the services that will build that model.
123
00:19:59.520 --> 00:20:10.020
Kris Skrinak: So Jupiter lab is an essential part of sage maker aws is an important contributor to that an open source project now I put a little crown by psychic learn.
124
00:20:10.410 --> 00:20:17.430
Kris Skrinak: Because psychic learn is so basic a lot of people talk about super sexy things, and you know tensorflow or.
125
00:20:17.880 --> 00:20:31.050
Kris Skrinak: pi torch etc psychic learn does all the basics and it's also a great educational tool, I always recommend folks that I work with no psychic learn cold before they get into any kind of deep learning.
126
00:20:31.410 --> 00:20:46.020
Kris Skrinak: pandas is just basically how we do tabular data manipulation and open CV like psychic learn now does a lot of things that were formerly only possible with deep learning so if you're doing computer vision open CV is an important tool.
127
00:20:46.470 --> 00:20:52.260
Kris Skrinak: Deep learning frameworks, I mentioned them, you know before tensorflow pi torching mx net, these are the big three.
128
00:20:53.280 --> 00:20:57.420
Kris Skrinak: i'm a bit of a pie towards bigger i'll just put my bias out there, I use pi torch.
129
00:20:58.470 --> 00:21:02.640
Kris Skrinak: Pretty extensively and I mentioned the deep learning containers that we have.
130
00:21:02.850 --> 00:21:14.430
Kris Skrinak: Our army, which is Amazon machine image which actually are managed products at aws we could coordinate all of those tools, you see above into something you could just stand up on an easy to instance.
131
00:21:14.730 --> 00:21:23.430
Kris Skrinak: and get right to work, if you use the deep learning army and it's free the instances aren't free and here are the categories for the instances that aws.
132
00:21:23.640 --> 00:21:33.930
Kris Skrinak: The PE class at the end they're the ones, without really high powered nvidia gpus last but not least is the neuron SDK near an SDK is does a number of things, including.
133
00:21:34.380 --> 00:21:42.930
Kris Skrinak: dramatically increases efficiency of your production models before you put them into production but they're required if you're going to use our custom.
134
00:21:43.830 --> 00:21:57.150
Kris Skrinak: cpus and gpus alright so popping to the next set here it's really easy for any program or even people who just know bash the bash Shell to put machine learnings into production i'm not kidding.
135
00:21:57.450 --> 00:22:07.920
Kris Skrinak: It whether it's vision chat bots etc and it's because of the growth and the maturity of these high level Ai services and i'm going to walk through some of the key ones right now.
136
00:22:08.430 --> 00:22:18.930
Kris Skrinak: So, first of all language translation is essential it's common outside of the US for people to speak multiple languages, of course, we're English centric in North America.
137
00:22:19.290 --> 00:22:31.380
Kris Skrinak: And just about everyone speaks English around the world, however, when you want to highly personalize your experience for your customers it's important to translate between not only languages, but dialects and that's what translate does.
138
00:22:31.620 --> 00:22:40.230
Kris Skrinak: transcribe will take speech from audio and video files and output it in a json machine readable format.
139
00:22:40.890 --> 00:22:47.310
Kris Skrinak: Next is extract text track to use his computer vision to look at the page a page of.
140
00:22:47.700 --> 00:22:57.750
Kris Skrinak: text, the same way we do so when we look at a piece of paper you know we generally will take a look at it and say okay there's a header there's a footer oops I guess there we go.
141
00:22:58.410 --> 00:23:06.660
Kris Skrinak: header footer there's something that looks like a table there's a photograph in here text track goes through, and does highly accurate.
142
00:23:07.500 --> 00:23:17.160
Kris Skrinak: semantic segmentation is what it's called in computer vision and then applies appropriate algorithms to each of those components to create machine readable code.
143
00:23:18.660 --> 00:23:24.360
Kris Skrinak: We can also do branded text to speech, and we have two great examples.
144
00:23:25.200 --> 00:23:35.520
Kris Skrinak: And i'll give you just a when I do the DEMO i'll do a quick DEMO of what that sounds like we have the Colonel for Kentucky fried chicken and an Australian bank nav.
145
00:23:36.060 --> 00:23:49.140
Kris Skrinak: So next up is personalized obviously the Amazon recommends system is very famous and we've taken all the exact same tools that we use it Amazon with the billions of SK use that we have.
146
00:23:49.560 --> 00:23:59.460
Kris Skrinak: And made all of those tools accessible to you through this high level API called personalized time series forecasting is an essential part of prediction.
147
00:23:59.880 --> 00:24:04.680
Kris Skrinak: We have methods within forecast where all you have to do is provide data.
148
00:24:05.070 --> 00:24:17.760
Kris Skrinak: Actually, no programming whatsoever and create highly accurate models models that used to take months to put together, we have a number of algorithms that we apply to that such as profit Rima.
149
00:24:18.450 --> 00:24:24.960
Kris Skrinak: And even deep learning auto aggressors so it's you basically it's a hands off service you just give a data.
150
00:24:25.350 --> 00:24:35.370
Kris Skrinak: Next is fraud detector it has the exact same model it identifies potentially fraudulent ad online activities such as payment fraud fake accounts, etc.
151
00:24:35.730 --> 00:24:49.290
Kris Skrinak: Now, the reason I group these three together is that they all have that exact same model, all you need is data you create a data set up, you can do it all in the gooey you push a button and outcomes and actually high.
152
00:24:49.680 --> 00:24:56.730
Kris Skrinak: High accuracy machine learning model it's the same for all three of these what's a little bit different are these two right here.
153
00:24:57.090 --> 00:25:11.070
Kris Skrinak: kendra uses intelligence search with natural language to create you know highly effective search mechanisms and what it does, is it actually enables you to just take your salesforce your sharepoint pdfs.
154
00:25:11.490 --> 00:25:27.690
Kris Skrinak: faqs just put them in a data set, we will search it for you, we do everything, sometimes we call this the fourth layer of aws because it uses all of those elements that I previously described to make a very effective search experience for your website.
155
00:25:28.050 --> 00:25:37.920
Kris Skrinak: Then, last but not least, here is context Center and our introductions now by integrating all of those media tools to create excellent customer support.
156
00:25:38.970 --> 00:25:45.480
Kris Skrinak: Just a little bit in Stage maker, if you do want to dive in there, we have a product called autopilot where you simply provide data.
157
00:25:46.170 --> 00:26:00.000
Kris Skrinak: And then I have always advocated, creating a web APP to DEMO my models, we have something called amplify on aws that makes it really easy to create react models even react native that can run on.
158
00:26:00.840 --> 00:26:04.710
Kris Skrinak: On phones, you know the executives the customers that you're creating code.
159
00:26:05.100 --> 00:26:15.270
Kris Skrinak: For really don't care about data science statistics like route means square your area under curve, or something like that they want to see the model in production amplify makes it easy to call any.
160
00:26:15.600 --> 00:26:24.300
Kris Skrinak: Of the High Level service that I mentioned, and I have a github repo there that I call NUTS because it's NUTS that people don't use it it uses react to just.
161
00:26:24.750 --> 00:26:34.050
Kris Skrinak: Do this very simply, I don't have a lot of time for a DEMO here, but I do want to point out just very quickly how easy, this is.
162
00:26:34.680 --> 00:26:44.400
Kris Skrinak: I have a video that came out of reinvent two years ago it's this gentleman talking about Ek s with cube flow and I really liked it, so I used.
163
00:26:44.910 --> 00:27:03.150
Kris Skrinak: You know, a tool, a plugin for chrome to download that and I went to transcribe and I created a transcription job, all I had to do there was given my s3 location and push play and it basically transcribed that entire thing.
164
00:27:03.900 --> 00:27:16.020
Kris Skrinak: I would also like to just very quickly do one or two things I have i'm just going to cat out in a terminal here, which I made kind of big.
165
00:27:16.560 --> 00:27:36.030
Kris Skrinak: The lyrics to a song by a band like King crimson, and here it is right here, and just using the just using the coi here, using polly I can actually output that and play it now, I obviously copying and pasting here, but you can see i'm using the aws see alive.
166
00:27:37.980 --> 00:27:47.610
Kris Skrinak: On the instruments adapt the sunlight right leg leaves went everywhere torn apart with nightmares and with dreams will know one lady Laurel review in silence drowns the screen there we go so.
167
00:27:49.170 --> 00:27:51.720
Kris Skrinak: Stop this player, now we can all sit there you go.
168
00:27:52.050 --> 00:28:01.470
Kris Skrinak: Alright cool, so the point here is, you know this is how you call any of those Ai Ai services from the command line I mentioned, you know there's four ways to do this.
169
00:28:01.740 --> 00:28:07.710
Kris Skrinak: This is nine times out of 10 and early stages, I will pick one of these Ai services in this case polly.
170
00:28:08.040 --> 00:28:16.170
Kris Skrinak: polly has some arguments here to synthesize speech and if you're curious about what all the options are there you just basically do that and say help.
171
00:28:16.500 --> 00:28:35.190
Kris Skrinak: So you can imagine here now going through every one of those high level services and getting whatever you want, whether that is a photograph, what are the objects in the photograph the activity, who are the personalities, you can detect up to 20 million faces with.
172
00:28:36.450 --> 00:28:48.120
Kris Skrinak: Recognition and recognition and comprehend both have something called custom labels were once again zero programming you just provide custom labels for the data that you're giving it and we create a highly accurate model.
173
00:28:48.630 --> 00:28:58.350
Kris Skrinak: I want to very briefly i'm a solutions architect, so I love seeing architectures show you an architecture for this kind of thing and production can be this simple.
174
00:28:58.800 --> 00:29:12.330
Kris Skrinak: And I have you know something here, using contact Center where we're combining now in a pipeline, all of these different tools, you can see here were using connect transcribe glue lex and quick site.
175
00:29:13.050 --> 00:29:19.020
Kris Skrinak: here's another model where we're just doing basic text track a synchronous asynchronous API workflow.
176
00:29:19.410 --> 00:29:24.570
Kris Skrinak: And then in sage maker, because I don't want to ignore sage maker, even though I said i'm not going to emphasize it today.
177
00:29:24.990 --> 00:29:38.070
Kris Skrinak: here's a full sample architecture from the data change a new data element being detected in s3 it kicks off a retraining job that goes and kicks off a job in code pipeline.
178
00:29:38.340 --> 00:29:48.510
Kris Skrinak: That could go then to step functions and it will go all the way to your hosted in point, by the way, that could also be kicked off by the model monitor here as well.
179
00:29:48.990 --> 00:29:55.650
Kris Skrinak: Now we do have one thing that's relatively new we call it the Ai rather the ml OPS framework.
180
00:29:55.920 --> 00:30:07.890
Kris Skrinak: It is a template really for putting ml OPS into production, so I encourage you to if you're a devops person to start there, because it really covers absolutely everything in depth.
181
00:30:08.220 --> 00:30:16.440
Kris Skrinak: Now ml and other services, I only have a few more minutes here, I just want to cover like I said we've put machine learning in almost every service at aws.
182
00:30:16.800 --> 00:30:26.130
Kris Skrinak: All of the language that you see in this report daily revenue year to date revenue was generated by an algorithm that's included within.
183
00:30:26.490 --> 00:30:36.960
Kris Skrinak: Quick site, which is our product that API is available, you could put this in your products as well, and the natural language query is available as well with a product we call cube.
184
00:30:38.190 --> 00:30:51.630
Kris Skrinak: Getting models built with relational databases from redshift is a simple as a sequel query so you make that query you can number one just call any service from aws like getting a sentiment score.
185
00:30:51.870 --> 00:31:04.200
Kris Skrinak: On some language, and you can do that with sequel, but you can also create complex models that go into production so that works with aurora as well, and even with our graph database, which is called Neptune.
186
00:31:04.860 --> 00:31:20.010
Kris Skrinak: All right, last but not least, here, look out for metrics is a new family of tools that we have that once again makes it easy for non data scientists to detect anomalies in virtually any time series data set.
187
00:31:20.520 --> 00:31:32.550
Kris Skrinak: All right, I covered a lot of material here and i'm going to pass it off to my colleague now miles Brown, but I do want to just say one or two closing comments, and that is, you know.
188
00:31:33.300 --> 00:31:41.700
Kris Skrinak: This famous quote by George fox all models are wrong, but some are useful, so it kind of brings it back to this concept that data writes code.
189
00:31:41.940 --> 00:31:47.970
Kris Skrinak: you're going to be making predictions and the predictions are the essential elements of what goes into production.
190
00:31:48.510 --> 00:31:56.280
Kris Skrinak: Some of the trends coming up this year include everything that I talked about here ml OPS hybrid cloud auto ml so.
191
00:31:57.000 --> 00:32:13.710
Kris Skrinak: Just ask yourself, can I write sequel yeah can I use the aws console yeah how do I know how to use a cla a command window and do you know enough javascript Python to make a restful call, then you already know enough to do high level highly accurate machine learning.
192
00:32:15.300 --> 00:32:17.580
Kris Skrinak: So miles hand it off to you.
193
00:32:17.940 --> 00:32:21.720
Myles Brown :: Moderator: Sure, Chris just before we move on on me, maybe people aren't.
194
00:32:23.610 --> 00:32:28.290
Myles Brown :: Moderator: Clear on a term hops and what that means, maybe you can expand on that a bit.
195
00:32:28.890 --> 00:32:38.610
Kris Skrinak: Sure, so um if you just Google, you know agile software development lifecycle you'll see that there's a lot of literature and really formalized process.
196
00:32:38.910 --> 00:32:48.960
Kris Skrinak: Around how we get traditional programming languages like C c++ Java etc from development into production and generally we call this devops.
197
00:32:49.260 --> 00:32:57.750
Kris Skrinak: So devops is a very different job from the programmers you know you, you have to know a little bit about development, but a lot about infrastructure and how you put that in production.
198
00:32:58.320 --> 00:33:11.610
Kris Skrinak: And there was a term added to devops called SEC devops where security became a real principal concern of devops in general now we're at ml OPS, but it's really ml SEC devops.
199
00:33:12.000 --> 00:33:20.730
Kris Skrinak: And what it really boils down to is in addition to tracking the code which we've done very well right we have github repos we QA knows how to look at a spec.
200
00:33:21.090 --> 00:33:30.300
Kris Skrinak: But when you introduce data as one of the principal authors of the code, it introduces nuances and completely new architectures.
201
00:33:30.540 --> 00:33:45.030
Kris Skrinak: For how you put that in production generally we just call that ml OPS, but you have to remember when you go into production you can't ignore security you can't ignore disaster recovery and backups and things like that so it's really devops with machine learning.
202
00:33:45.990 --> 00:33:48.450
Myles Brown :: Moderator: yeah Thank you I and that's one of the things I noticed.
203
00:33:48.450 --> 00:33:48.660
Kris Skrinak: Was.
204
00:33:48.750 --> 00:34:01.380
Myles Brown :: Moderator: You know, in the early days when you had maybe one data scientists within a large organization, they were sort of on their own doing their own thing you know and devops might start to make its way into your processes, but then.
205
00:34:01.800 --> 00:34:10.200
Myles Brown :: Moderator: All the analytics and data science That was all its own little thing and now that's that's probably the biggest trend i'm seeing now is.
206
00:34:10.380 --> 00:34:11.280
Kris Skrinak: is huge.
207
00:34:11.340 --> 00:34:12.120
Kris Skrinak: And it's fine.
208
00:34:12.240 --> 00:34:26.790
Myles Brown :: Moderator: maturity yeah they're trying to operationalize the whole thing and like you said, if this isn't the purview of just a few really you know really highly specialized data scientists it's it's sort of on every developer's radar these days.
209
00:34:27.120 --> 00:34:37.200
Kris Skrinak: Absolutely yeah it also you know it also says that this train has not left the station right you don't want to try to chase a train after it leaves leaves the station.
210
00:34:37.560 --> 00:34:46.590
Kris Skrinak: And i'll OPS is maturing it's expanding we need good talented engineers in every single aspect of that software development lifecycle.
211
00:34:47.010 --> 00:34:54.690
Myles Brown :: Moderator: Yes, and there's that you know there's a million surveys out there, saying that this is an area where you know, we need a lot of people.
212
00:34:56.070 --> 00:35:00.510
Myles Brown :: Moderator: And so it's it's definitely in demand, let me just share my slides here.
213
00:35:02.340 --> 00:35:02.730
sure.
214
00:35:04.230 --> 00:35:08.280
Myles Brown :: Moderator: I got just a couple slides about that well first off, I want to thank Chris.
215
00:35:08.940 --> 00:35:18.390
Myles Brown :: Moderator: for doing this for us it, you know it's rare to have somebody who's you know so deeply technical who's also a great presenter and can break these complex things down and make it.
216
00:35:18.870 --> 00:35:38.790
Myles Brown :: Moderator: You know, clear, I think that this this is within everybody's REACH, you know to certain degree, we had exit certified are an authorized training partner have been since 2014 of aws we're also partner with some other vendors typically we mostly do vendor authorized training.
217
00:35:40.320 --> 00:35:44.670
Myles Brown :: Moderator: But but 30 different vendors, but aws is our biggest partner.
218
00:35:45.330 --> 00:35:54.930
Myles Brown :: Moderator: And we've been doing virtual training since 2012 you know this isn't something new that we've we've just added in our regular classes, if you do want virtual training.
219
00:35:55.230 --> 00:36:04.980
Myles Brown :: Moderator: And they are based on zoom like this, but it's not like a webinar style we really encourage two way audio and video people asking questions is as much like a regular class as possible.
220
00:36:05.520 --> 00:36:16.170
Myles Brown :: Moderator: And you might see the logo here exit certified is sort of the go to market training division of tech data so tech data is you know, a very large distributor that you may have heard of.
221
00:36:17.100 --> 00:36:34.500
Myles Brown :: Moderator: We have a lot of training on the aws side, these are these are authorized aws classes built by the vendor delivered by aws authorized instructors, we have a number on staff like quite a large number, and we have these classes running you know publicly.
222
00:36:35.700 --> 00:36:51.750
Myles Brown :: Moderator: You know some of the classes are a little more rare than others, they might only run every three or four weeks there are classes, you know your sort of entry level aws classes that run every week, but if you have a group of people, we can put on a special event just for you and.
223
00:36:53.190 --> 00:36:56.760
Myles Brown :: Moderator: Okay, great oh somebody's asking I don't see a Q amp a.
224
00:36:57.780 --> 00:37:00.540
Myles Brown :: Moderator: Oh yeah that's it's just the questions are going to be in the chat.
225
00:37:00.900 --> 00:37:09.450
Myles Brown :: Moderator: So we encourage anybody has questions throw them in the chat we're going to come back and ask Chris any any leftover Q amp a questions so throw them in the chat if you have any.
226
00:37:09.870 --> 00:37:14.610
Myles Brown :: Moderator: In the meantime, I just want to run down some of the major classes that sort of apply here.
227
00:37:15.300 --> 00:37:23.280
Myles Brown :: Moderator: You know, one of the things Chris said is that data is paramount to machine learning and so, if you want to learn about that whole data analytic stack.
228
00:37:24.090 --> 00:37:30.630
Myles Brown :: Moderator: In aws we have a three day class called the big data on aws class and it talks all about how we, you know.
229
00:37:31.110 --> 00:37:45.390
Myles Brown :: Moderator: Starting from collecting the data with things like denise's all the way back to you know where we're going to start long term like redshift or maybe a data lake there's also a one day data lake class I didn't include it in this list, but that's a class we run fairly often as well.
230
00:37:46.470 --> 00:37:54.390
Myles Brown :: Moderator: more specific to some of this ml stack we have a one day client class called practical data science with Amazon sage maker.
231
00:37:55.560 --> 00:38:02.820
Myles Brown :: Moderator: As Chris mentioned, you know, not everybody needs to be that deep you know data scientists specialist.
232
00:38:03.120 --> 00:38:13.560
Myles Brown :: Moderator: But if you are starting down the road and you want to build your own models with sage maker, if you already have sort of that knowledge of data, science and you just want to know about.
233
00:38:13.860 --> 00:38:25.140
Myles Brown :: Moderator: This specific aws tool we got that Nice one day class, but if you want a broader you know view of hey what is the job of the data scientists, how do I build and train those models.
234
00:38:25.650 --> 00:38:34.110
Myles Brown :: Moderator: We do have this nice four days machine learning pipeline on aws class and that that does serve pretty well to help you get ready for the.
235
00:38:34.980 --> 00:38:40.080
Myles Brown :: Moderator: machine learning specialty certification if that's something you're going down that road.
236
00:38:40.620 --> 00:38:47.940
Myles Brown :: Moderator: Now, typically if you're going to be working in these things you're probably already well versed in a language like Python.
237
00:38:48.390 --> 00:38:53.730
Myles Brown :: Moderator: I think, Chris mentioned, I think you you literally said it's the lingua franca of machine learning.
238
00:38:54.210 --> 00:38:59.700
Myles Brown :: Moderator: And so, if you're starting with nope you know you might be a Java developer, you might be, you know C sharp or something.
239
00:39:00.690 --> 00:39:07.860
Myles Brown :: Moderator: If you just want a nice gentle introduction to Python we've got those kinds of classes, but we also have some classes, where it says well.
240
00:39:08.190 --> 00:39:15.810
Myles Brown :: Moderator: You know if you know a little bit about Python let's take you the rest of the way because you're going to want to learn about pandas and then you know, so we have some.
241
00:39:16.140 --> 00:39:24.480
Myles Brown :: Moderator: You know, comprehensive data science with Python So these are not not necessarily aws classes, these are sort of language classes and like I mentioned, you know.
242
00:39:24.900 --> 00:39:30.660
Myles Brown :: Moderator: We were partnered with a bunch of vendors and then you know for open source things we have you know various.
243
00:39:31.320 --> 00:39:45.210
Myles Brown :: Moderator: places that we gather and we have you know quite a roster of instructors that can help us with those things, and if you want to know more, you know you can go to exit certified.com slash aws actually let me throw that in the.
244
00:39:48.000 --> 00:39:55.470
Myles Brown :: Moderator: In the chat you should be able to click through that I think in the chat it should be a hyperlink, but you can poke around our website and see what's there.
245
00:39:56.400 --> 00:40:09.420
Myles Brown :: Moderator: If you have a group of people, then you know, then you don't have to be so set in the like nine to five kind of classes, that we have opportunities to do a half day over maybe more days.
246
00:40:10.200 --> 00:40:20.490
Myles Brown :: Moderator: We have you know total flexibility when you do a private class, but if you only have one or two people that you want to send on training, well then we've got quite a robust public schedule for you.
247
00:40:21.240 --> 00:40:35.520
Myles Brown :: Moderator: And right now we have a promo where you can get some cheaper training it's our summer promo you have to register by August 27 and then you have to take the class, I think, by the end of September.
248
00:40:36.120 --> 00:40:45.630
Myles Brown :: Moderator: And the way it works is basically if it's a one day class you get $100 off if it's due date classic 200 all the way up to $500 off a five day class.
249
00:40:46.350 --> 00:40:54.990
Myles Brown :: Moderator: If you're doing learning subscriptions or on demand or self paced courses or 10% off, and so this is this our current promo we run these in the summer.
250
00:40:55.530 --> 00:41:03.960
Myles Brown :: Moderator: You know the training business is a little bit cyclical so the summer is usually a little lighter a lot of people on vacation, but it is an opportunity for you, if you want.
251
00:41:04.410 --> 00:41:11.520
Myles Brown :: Moderator: To get a good deal on training and then also $500 off per day if you're doing private training so Those are all.
252
00:41:12.240 --> 00:41:27.750
Myles Brown :: Moderator: available, you can learn more at exit certified.com slash summer 500, but this will all be I think the follow up email that you'll get you know telling you about the recording of the presentation, you know well we'll have information about that promo as well.
253
00:41:28.890 --> 00:41:34.380
Myles Brown :: Moderator: And I think that brings us to the Q amp a let me see.
254
00:41:36.750 --> 00:41:48.330
Myles Brown :: Moderator: If we have any more questions for Chris he's a great resource, this is a good time to get Adam he's also been kind enough to leave his Oh, maybe Chris if you want to throw your email address in the chat.
255
00:41:48.660 --> 00:41:53.790
Kris Skrinak: So absolutely happy to i'll also point out i'm going to put in.
256
00:41:56.670 --> 00:42:08.280
Kris Skrinak: The chat right now, a link to my deck so that deck is accessible to everybody and it includes some of the DEMO code, so you can play with that.
257
00:42:09.420 --> 00:42:13.620
Kris Skrinak: But yeah any questions definitely eager to hear them.
258
00:42:13.860 --> 00:42:15.360
Myles Brown :: Moderator: yeah we really appreciate that because.
259
00:42:15.390 --> 00:42:17.520
Myles Brown :: Moderator: Some of those architectural diagrams.
260
00:42:18.240 --> 00:42:22.200
Myles Brown :: Moderator: But it coming in and you might you might want to take a closer look at this yeah exactly.
261
00:42:25.710 --> 00:42:27.150
Myles Brown :: Moderator: Oh sorry and your email address.
262
00:42:27.330 --> 00:42:30.540
Kris Skrinak: Oh yeah and then, I have a question for you miles, I mean if you.
263
00:42:30.540 --> 00:42:36.990
Kris Skrinak: don't mind um so do do your courses oh wait let's that we have some coming in from our.
264
00:42:37.380 --> 00:42:38.670
Myles Brown :: Moderator: Our phone is is Matthew.
265
00:42:38.760 --> 00:42:43.650
Kris Skrinak: Putting the promo I see I see if I wanted to make vision, Ai that identifies birds.
266
00:42:43.830 --> 00:42:49.230
Kris Skrinak: fly over my house off the top of your head, how much effort, do you think that would take.
267
00:42:50.160 --> 00:42:58.890
Kris Skrinak: yeah that's a great question to ask Well, first of all, the courses that miles just outlined will give you that introduction to Python.
268
00:42:59.130 --> 00:43:10.110
Kris Skrinak: And the ice and the access to where you're going to store those pictures of those birds on on aws and that's the starting point right that's the.
269
00:43:10.440 --> 00:43:16.110
Kris Skrinak: that's the we used to call when you get in a taxi and you have to pay something just to sit in the cab That was the unfair right like.
270
00:43:17.100 --> 00:43:26.190
Kris Skrinak: So that's what you've got to get through before you can start identifying those birds, but then now once you've got your data set or maybe you're assuming that.
271
00:43:26.700 --> 00:43:28.680
Kris Skrinak: You know you're going to have a live camera.
272
00:43:29.370 --> 00:43:39.780
Kris Skrinak: cameras take individual pictures they do it at a rate of 30 or 60 frames per second for machine learning what i've generally seen is like five to six images per second.
273
00:43:40.080 --> 00:43:49.860
Kris Skrinak: And there's a way to run that through a pipeline at aws to call recognition so we'll low code way to do this, but like I said you're going to need a Jupiter notebook.
274
00:43:50.070 --> 00:44:04.500
Kris Skrinak: And Python knowledge is to tie all that that architecture together so that it calls recognition every time one of those new frames hits s3 cause a Lambda that cause recognition and it identifies the bird.
275
00:44:06.390 --> 00:44:11.460
Myles Brown :: Moderator: So you're saying that that recognition already has a library of what birds, or what.
276
00:44:12.090 --> 00:44:14.580
Kris Skrinak: Probably I know you have to put.
277
00:44:14.610 --> 00:44:16.470
Kris Skrinak: You have to get that data and now there's.
278
00:44:16.770 --> 00:44:29.910
Kris Skrinak: ways to scrape that off of Google images or preferably you know you would have photographs that you took yourself with that camera that you were using to identify it's always better to use your own data set right.
279
00:44:30.690 --> 00:44:35.010
Kris Skrinak: Take those pictures use custom labels in recognition, and you can build a model.
280
00:44:35.520 --> 00:44:35.850
Okay.
281
00:44:37.290 --> 00:44:38.490
Kris Skrinak: going to take the next one miles.
282
00:44:38.850 --> 00:44:45.990
Myles Brown :: Moderator: You know, there was another question here at what point does the code become ml because it's possible to hard code and algorithm that act like ml.
283
00:44:49.890 --> 00:44:59.880
Kris Skrinak: um you know if you don't mind i'll just do this um that's really where psychic learn comes in right psychic learn is sometimes called shallow learning but that's not right.
284
00:45:00.480 --> 00:45:07.890
Kris Skrinak: A better word is classic classic machine learning where you know you're basically just not using deep learning.
285
00:45:08.250 --> 00:45:15.150
Kris Skrinak: And you'll if you go through psychic learn you'll find that most of it is based on you know statistics it's a lot of statistics.
286
00:45:15.420 --> 00:45:26.820
Kris Skrinak: So I do not discount psychic learn or classical machine learning right and where the actual line is I mean it's a big Gray space so.
287
00:45:27.000 --> 00:45:40.680
Kris Skrinak: Deep learning is always going to be ml that's just a given right, but when you're in psychic learn you're doing classification with statistical methods yeah I mean you know it can be very Gray area, is it really machine learning or is it just a statistical prediction.
288
00:45:42.810 --> 00:45:56.640
Kris Skrinak: And then I see the other question here follow up on the bird identification, I was also wondering if aws has a pre built recognition system well yeah that is recognition, with a K, not a C.
289
00:45:57.060 --> 00:46:08.730
Kris Skrinak: And it does identify a lot of objects, but it's not going to identify species of birds, where you live, that is, that requires a custom model, or what we often call bespoke model.
290
00:46:11.400 --> 00:46:24.090
Kris Skrinak: Is Mr specific to the different differentiation of sensory input, I meant for differentiating birds from things like planes or flies well if you don't mind I also.
291
00:46:24.690 --> 00:46:37.320
Kris Skrinak: mentioned, you know it's machine learning and object identification is so sophisticated now that if you can label it, you can identify it it's almost that simple at this point.
292
00:46:37.920 --> 00:46:46.560
Kris Skrinak: Individual dogs species, if you have the correct label, for example, that's the classic example that we use different dog species.
293
00:46:47.070 --> 00:47:00.960
Kris Skrinak: load them up into either fast Ai fast Ai has a really good framework for that mx net has something called auto blue on which will create a custom species detection model for those birds.
294
00:47:01.890 --> 00:47:08.760
Kris Skrinak: So yeah there's many roads to roam here it's really just a question of what degree of accuracy, are you going to be satisfied with.
295
00:47:09.480 --> 00:47:20.850
Myles Brown :: Moderator: But it didn't even out of the box recognition, if you if you give it a bunch of images and you say hey which ones are birds vs which ones are airplanes it'll do a pretty good.
296
00:47:20.910 --> 00:47:23.550
Kris Skrinak: that'll that'll do it without of the yes.
297
00:47:23.610 --> 00:47:29.070
Myles Brown :: Moderator: that's that's an A but yeah It might not know it might not know too many species of birds.
298
00:47:29.370 --> 00:47:33.360
Kris Skrinak: yeah it's probably no son probably knows it Robin versus a crow.
299
00:47:33.750 --> 00:47:44.580
Myles Brown :: Moderator: yeah but but yeah if you want to really get into it you're probably going to have to go and custom label and feed it a bunch of images of of a specific species right.
300
00:47:44.970 --> 00:47:50.160
Kris Skrinak: I have an ongoing battle with the crows in my neighborhood so a little bit of a sore point.
301
00:47:51.930 --> 00:47:52.350
Myles Brown :: Moderator: Nice.
302
00:47:54.570 --> 00:47:59.430
Myles Brown :: Moderator: Okay, well, it sounds like sounds like christopher's well on his way to.
303
00:48:00.510 --> 00:48:11.310
Myles Brown :: Moderator: to having a good starting job with the you know that that that's the kind of project that you can really delve into and learn a lot on your own.
304
00:48:11.970 --> 00:48:15.150
Kris Skrinak: Absolutely it's a great starter project yeah.
305
00:48:19.230 --> 00:48:22.530
Myles Brown :: Moderator: We got a couple more minutes if there's more questions you can throw them in there.
306
00:48:25.170 --> 00:48:34.200
Kris Skrinak: And and miles, the question I was going to ask you is what kind of skills do folks usually have coming into your courses and.
307
00:48:34.620 --> 00:48:37.800
Kris Skrinak: Okay, and are their requirements, really, for that matter.
308
00:48:38.010 --> 00:48:48.300
Myles Brown :: Moderator: yeah so so typically for something like that for day machine learning pipeline, there is an assumption that you've used aws before, and so, if you're brand new.
309
00:48:48.600 --> 00:48:54.270
Myles Brown :: Moderator: Then you should probably take the one day tech essentials class where we sort of cover hey what is cloud.
310
00:48:54.570 --> 00:49:05.220
Myles Brown :: Moderator: You know what is aws, what are the, what are the basic services that everybody uses and get some hands on you know, with a few labs and so that's like it's sort of like an eight hour day class.
311
00:49:06.120 --> 00:49:14.160
Myles Brown :: Moderator: If you've used aws at all you've logged into the management console and you know, maybe launched an easy to instance or something, then you can safely skip that class.
312
00:49:14.580 --> 00:49:24.870
Myles Brown :: Moderator: But that that's usually the the grounding for the machine learning class, but something like the big data class, because there is this expectation that.
313
00:49:25.530 --> 00:49:33.450
Myles Brown :: Moderator: You know we're touching on a lot of services that build on the basic services, you probably have sort of what we would call an associate level knowledge.
314
00:49:33.930 --> 00:49:38.100
Myles Brown :: Moderator: So, so the next level up from the tech essentials, would be the sort of.
315
00:49:38.700 --> 00:49:47.010
Myles Brown :: Moderator: Three day classes, like you know we have we have three of them one for SIS OPS sort of administrators, one for architects one for developers.
316
00:49:47.430 --> 00:50:00.090
Myles Brown :: Moderator: And so that'll get you the hey here's here's the 25 services that most people use in aws and get hands on with those so you get sort of three days, under your belt, so if you've taken either actually.
317
00:50:00.450 --> 00:50:12.150
Myles Brown :: Moderator: Whether you take the architect or the SIS OPS, or the developer class then you'll be ready to go into that big data class where we talked about you know some of the things you mentioned redshift and glue and can he says.
318
00:50:12.690 --> 00:50:23.610
Myles Brown :: Moderator: emr you know some of those data wrangling kind of techniques, but the you know it sort of depends, we actually have on our website kind of a.
319
00:50:24.630 --> 00:50:28.710
Myles Brown :: Moderator: path, actually, let me just show it to you so just so people can see it.
320
00:50:30.780 --> 00:50:33.030
Yes, everything's always in the way when I do this.
321
00:50:34.140 --> 00:50:40.320
Myles Brown :: Moderator: Okay, so here's here's our website, and let me just go to the resources section.
322
00:50:41.940 --> 00:50:58.020
Myles Brown :: Moderator: Okay i'm going to get all these things out of the way yeah going to go to the resources section, and we have what we call learning pads and if I go find the one specifically for aws you can also get there from the aws page, but this is this is usually my fastest way and.
323
00:50:59.910 --> 00:51:01.140
Myles Brown :: Moderator: Let me find the.
324
00:51:03.120 --> 00:51:09.450
Myles Brown :: Moderator: Learning path aws yeah here we go, so this is a two page PDF that we have that we've sort of built out.
325
00:51:09.990 --> 00:51:18.240
Myles Brown :: Moderator: And you'll see that most of the tracks start with the tech essentials, one day, then go to either the architect thing or the developer, or the SIS OPS.
326
00:51:18.570 --> 00:51:28.680
Myles Brown :: Moderator: And those would get you ready for your first level of sort of associate level certifications if you're going the certification route and then there's some professional level certifications.
327
00:51:29.340 --> 00:51:30.210
Myles Brown :: Moderator: And then on this.
328
00:51:30.360 --> 00:51:41.550
Myles Brown :: Moderator: The second page is where you see hey here's more of the data analytic side, where you might take the architect or this is OPS, and then the big data class, you might take the the one day data lake class.
329
00:51:43.110 --> 00:51:55.020
Myles Brown :: Moderator: If you're in the machine learning like I said, the one day tech essentials, the four day machine learning pipeline and that pretty much get you ready, you probably have to do a little self study but get you ready for that machine learning specialty.
330
00:51:56.490 --> 00:52:07.170
Myles Brown :: Moderator: And you know it goes on, we have some classes, specifically on service solutions you know Lambda and a gateway we have security glasses so we've got this sort of two page.
331
00:52:08.310 --> 00:52:13.260
Myles Brown :: Moderator: kind of sort of learning and certification path, and these are all aws built.
332
00:52:14.280 --> 00:52:18.720
Myles Brown :: Moderator: They maintain them that's that's the big difference with getting authorized training from.
333
00:52:19.140 --> 00:52:30.120
Myles Brown :: Moderator: From an aws partner is you know, sometimes you go to one of these Gray market things and somebody the class, but two years later, you know the cloud is such a moving target that.
334
00:52:30.990 --> 00:52:42.960
Myles Brown :: Moderator: You know, you need you need the power of that vendor updating those classes, you know we have these updating classes every two weeks pretty much something's changed and we need to update them so.
335
00:52:43.170 --> 00:52:44.520
Kris Skrinak: that's that's also.
336
00:52:44.790 --> 00:52:55.140
Kris Skrinak: mentioned that 80% of employers that you might want to work with after you take this course, and that our customers their data is on aws.
337
00:52:55.500 --> 00:52:55.920
Myles Brown :: Moderator: Oh yeah.
338
00:52:56.130 --> 00:53:02.370
Kris Skrinak: yeah so you know data writes the code if you're going to get into machine learning you better go where the data is yeah.
339
00:53:03.960 --> 00:53:04.440
Myles Brown :: Moderator: alright.
340
00:53:05.850 --> 00:53:10.470
Myles Brown :: Moderator: See if we have any more questions and some thanks for great information.
341
00:53:13.710 --> 00:53:18.000
Myles Brown :: Moderator: we'll just stick around for another minute or two see if any more questions come in.
342
00:53:19.800 --> 00:53:25.080
Myles Brown :: Moderator: Again, thank you very much, Chris really appreciate your your help with this and.
343
00:53:25.320 --> 00:53:26.040
Kris Skrinak: As you're entirely.
344
00:53:26.250 --> 00:53:27.900
Myles Brown :: Moderator: you've got his nomad you've got.
345
00:53:28.140 --> 00:53:39.690
Kris Skrinak: you've got a web based social media there I do a lot on YouTube a lot I run the New York City deep learning meetup and lot we get into a lot of details a lot of data science.
346
00:53:40.590 --> 00:53:46.650
Kris Skrinak: And it's it's there's some good stuff up there, so if you were going to check out one thing there, it would definitely be the YouTube channel.
347
00:53:47.190 --> 00:53:50.220
Myles Brown :: Moderator: Those meetups are you meeting up virtually these days.
348
00:53:50.670 --> 00:53:58.530
Kris Skrinak: yeah and I don't like it way to get back into the loft I love human interaction yeah.
349
00:53:58.890 --> 00:54:01.500
Myles Brown :: Moderator: Oh, you guys were hosting those that the aws.
350
00:54:01.590 --> 00:54:11.670
Kris Skrinak: At the New York City aws loft in soho yeah broadway it's it's it, you can fit like 250 people in there it's fantastic.
351
00:54:12.360 --> 00:54:18.240
Myles Brown :: Moderator: yeah yeah there's there's a few cities that have those and that that sort of makes sense.
352
00:54:18.840 --> 00:54:20.280
Kris Skrinak: yeah London Tokyo.
353
00:54:20.850 --> 00:54:22.740
Kris Skrinak: And I think there's one more question here.
354
00:54:26.550 --> 00:54:29.400
Kris Skrinak: Someone prompted us but I.
355
00:54:30.240 --> 00:54:31.110
Myles Brown :: Moderator: Think he's typing it.
356
00:54:31.440 --> 00:54:31.920
Okay.
357
00:54:33.570 --> 00:54:33.990
cool.
358
00:54:36.480 --> 00:54:38.400
Myles Brown :: Moderator: yeah went to the San Francisco loft.
359
00:54:38.880 --> 00:54:50.220
Kris Skrinak: If i'm going to build out an ml API that integrates about 10,000 transactions, a month, what would that cost be typically around five kb per transaction.
360
00:54:51.150 --> 00:55:01.980
Kris Skrinak: Well, total cost management is an a priority it's one of the well architected principles at aws and when you say it's five K per transaction, you know.
361
00:55:02.850 --> 00:55:07.530
Kris Skrinak: it's it's really, we need to get into there's more questions than answers here.
362
00:55:08.040 --> 00:55:16.440
Kris Skrinak: The short story is there's an optimization path for almost any type of inference I mentioned that we have custom hardware, for instance.
363
00:55:16.950 --> 00:55:29.100
Kris Skrinak: We also have an iot suite that includes a lot of cost optimization tools to get that data if it's coming from an iot device at five K to to the model.
364
00:55:30.030 --> 00:55:38.520
Kris Skrinak: So we have a number of architectures that are in the machine learning blog on aws that talk about cost optimization.
365
00:55:39.450 --> 00:55:46.920
Kris Skrinak: But yeah I mean, I think that number of questions here around you know what this is actually going to cost can be very difficult to answer.
366
00:55:47.490 --> 00:55:57.210
Kris Skrinak: Just one little nuance, for example, it doesn't cost a penny to put data in s3, but when a user reads it then you know you we charge for the read, not the right.
367
00:55:57.600 --> 00:56:14.820
Kris Skrinak: So you know that's just sort of one nuance here is this data, already in the cloud is it you know coming from an iot devices and a web APP and mobile APP you know there's just way too many questions to front of that, but we also go out of our way to keep costs as low as possible.
368
00:56:17.700 --> 00:56:23.010
Myles Brown :: Moderator: yeah and I guess it kind of depends, which ml API you're using to.
369
00:56:24.060 --> 00:56:34.470
Kris Skrinak: write an API service versus a bespoke model versus calling, you can call models from Lambda you can put containerized models in Lambda call them from Lambda.
370
00:56:35.370 --> 00:56:50.820
Kris Skrinak: We have pre built algorithms in sage maker, and they can go into production as a sage maker endpoint and then that stage maker endpoint can host any number of individual models so that's called multi model hosting so it gets nuanced very quickly.
371
00:56:55.050 --> 00:56:56.430
Kris Skrinak: And that's where you need good training.
372
00:56:58.470 --> 00:56:59.970
Myles Brown :: Moderator: Alright, thanks well thanks for that.
373
00:57:00.810 --> 00:57:16.020
Myles Brown :: Moderator: Alright, it looks like we're pretty much at time, so I think we're going to shut things down but, once again, thank you very much, and everybody look forward to the email coming probably I think we said, Michelle was it end of the week, maybe we're thinking, probably the email with the recording.
374
00:57:16.380 --> 00:57:18.030
Michelle :: Webinar Producer: yeah we'll get that out at the end of the week.