WEBVTT 00:01.280 --> 00:04.280 Hello everyone and welcome. 00:04.680 --> 00:07.560 In today's session we'll be learning about models. 00:07.720 --> 00:12.200 We'll go over the different kinds of models that are available as of today. 00:12.760 --> 00:17.440 We'll start with foundation models, then move on to language models. 00:17.840 --> 00:20.120 We'll cover large language models. 00:20.520 --> 00:24.360 And finally we'll wrap up with multimodal models. 00:25.320 --> 00:31.880 It's very important to understand models in general are a vast topic. 00:32.320 --> 00:36.200 It's hard to cover all the intrinsic details in a single course. 00:36.440 --> 00:44.080 However, in this section I've introduced all the essential jargon and terminology you'll need throughout 00:44.080 --> 00:44.920 the course. 00:46.160 --> 00:52.640 This will not only help you understand the content, it will also help you grasp how different models 00:52.640 --> 00:56.360 work and how they perform across categories. 00:56.720 --> 01:00.490 So having said that, let's begin. 01:01.330 --> 01:02.770 Foundation models. 01:03.490 --> 01:06.130 The first model we'll cover is the Foundation model. 01:06.450 --> 01:13.810 Foundation models, also known as FMS, are large deep learning neural networks trained on massive data 01:13.850 --> 01:14.530 sets. 01:15.130 --> 01:22.810 The term foundation model was coined by researchers to describe ML models that are trained on broad, 01:22.970 --> 01:31.610 generalized and unlabeled data and capable of performing a wide variety of general tasks like understanding 01:31.610 --> 01:36.890 language, generating text, and even generating images. 01:37.890 --> 01:44.370 The size and purpose of an FM is what sets it apart from traditional machine learning models. 01:44.490 --> 01:52.370 You can use a foundation model as a base model for developing more specialized downstream applications. 01:53.450 --> 01:55.820 How do foundation models work? 01:56.460 --> 01:59.660 Foundation models are a form of generative AI. 02:00.260 --> 02:07.540 They generate output based on one or more inputs, usually in the form of human language instructions. 02:07.980 --> 02:15.940 These models learn patterns and relationships and use that learning to predict the next item in a sequence. 02:16.460 --> 02:22.020 In the case of text, they predict the next word in a sentence based on context. 02:22.900 --> 02:29.180 With image generation, they analyze image features and create clearer, sharper versions. 02:29.580 --> 02:36.140 It's crucial to remember FM's use learned patterns to predict the next logical item in a sequence. 02:36.620 --> 02:38.980 Examples of foundation models. 02:39.900 --> 02:45.100 Some examples include Bert bidirectional encoder representations. 02:45.100 --> 02:48.260 Stars Transformers released in 2018. 02:48.740 --> 02:54.510 It analyzes the full context of a sentence bidirectionally and then makes predictions. 02:55.070 --> 02:55.470 G. 02:56.830 --> 03:02.790 Generative pre-trained transformer also released in 2018 by OpenAI. 03:03.550 --> 03:10.350 It uses a 12 layer transformer decoder with a self-attention mechanism, and was trained on a dataset 03:10.390 --> 03:13.070 of over 11,000 free novels. 03:13.670 --> 03:19.590 Amazon Titan a powerful pre-trained foundation model designed for general purpose use. 03:19.950 --> 03:26.830 There are many others too, from organizations like cohere, Anthropic, and Meta's Llama family. 03:27.790 --> 03:29.030 Language models. 03:29.150 --> 03:32.590 Now let's understand language models. 03:33.390 --> 03:39.870 A language model is a type of machine learning model designed specifically to represent language. 03:40.430 --> 03:47.680 It forms the basis for many language tasks like Q and a summarization or semantic search. 03:48.800 --> 03:50.640 How are models trained? 03:51.120 --> 03:57.760 The training process involves neural networks like Bert processing millions of data points. 03:58.120 --> 04:00.800 This process is what we call training. 04:02.200 --> 04:05.480 Once you train a model, the next phase is fine tuning. 04:05.960 --> 04:12.000 There are many tasks that benefit from language model like sentiment analysis, question answering, 04:12.000 --> 04:12.920 and others. 04:13.440 --> 04:18.080 Adapting a general purpose model to such task is known as fine tuning. 04:18.480 --> 04:22.560 Now let's understand diagrammatically how this looks like. 04:23.320 --> 04:24.480 General purpose. 04:25.320 --> 04:32.200 Purpose models like Bert or its bigger sister, like Roberta, require a huge amount of data to learn 04:32.200 --> 04:34.080 a language's regularities. 04:34.560 --> 04:42.320 NLP practitioners often use Wikipedia and other freely available collections of textual data to train 04:42.320 --> 04:42.760 them. 04:43.050 --> 04:45.730 Here are the general purpose model available. 04:45.730 --> 04:48.570 As of now we have Bert code. 04:48.610 --> 04:53.850 Roberta, Roberta and others that are listed here from the Bert family. 04:54.290 --> 05:01.410 So once you have a general purpose model, you would want to fine tune a language model to make it like 05:01.410 --> 05:03.570 Bert to a specific domain. 05:03.810 --> 05:06.610 You'd fine tune it using domain specific data. 05:07.530 --> 05:16.290 For example, Bert, plus scientific papers equals sy Bert Bert plus financial texts equals fin. 05:16.290 --> 05:19.210 Bert plus legal documents equals legal. 05:19.250 --> 05:19.810 Bert. 05:20.370 --> 05:23.370 That's how fine tuning helps specialize a model. 05:24.810 --> 05:32.690 Large language models llms with advancements and language models, researchers built even larger ones, 05:33.010 --> 05:37.210 requiring massive computing power and huge training data sets. 05:37.610 --> 05:43.100 That's how large language models or less were born. 05:44.860 --> 05:47.300 Model size comparison. 05:47.300 --> 05:58.500 Let's compare Bert Base Bert Large, bigger GPT three with 175 billion parameters, GPT four even larger. 05:58.500 --> 06:00.860 That's why they're called large language models. 06:01.740 --> 06:03.860 Training Llms the phases. 06:04.620 --> 06:06.180 There are three key phases. 06:06.620 --> 06:11.580 Pre-training the model learns by predicting the next word in a sequence. 06:12.140 --> 06:14.940 Supervised fine tuning. 06:15.220 --> 06:21.420 Researchers teach the model how to respond to prompts by using high quality Q and A datasets. 06:22.340 --> 06:26.700 Reinforcement learning from human feedback, also known as RL. 06:27.900 --> 06:31.940 This involves human rankings to help guide the model toward better answers. 06:32.300 --> 06:34.820 You can imagine the model evolving like this. 06:35.260 --> 06:37.950 At first it's like a wild dragon. 06:38.310 --> 06:45.750 Supervised fine tuning tames it r f allows it to collaborate with humans effectively. 06:47.310 --> 06:48.870 Multimodal models. 06:48.910 --> 06:52.030 Now let's talk about multimodal models. 06:52.310 --> 06:58.310 These models can understand and generate content across multiple modalities, not just text. 06:58.630 --> 07:06.270 So instead of only processing and generating text, they can now process text, images, audio and video 07:06.630 --> 07:10.550 and output text, images, audio and video. 07:11.390 --> 07:17.110 This enables much richer, more context aware and versatile AI experiences. 07:17.550 --> 07:23.630 I hope you now have a solid understanding of the different kinds of models and how they evolved over 07:23.630 --> 07:24.270 time. 07:25.270 --> 07:28.190 I'll see you in the next video.