WEBVTT 00:00.840 --> 00:02.040 Hello everyone! 00:02.520 --> 00:06.640 In today's video we will learn about a hallucination in AI. 00:07.200 --> 00:09.200 What is AI hallucination? 00:09.760 --> 00:15.360 AI hallucinations are incorrect or misleading results that AI model generates. 00:15.800 --> 00:22.600 These errors can be caused by a variety of factors, including insufficient training data, incorrect 00:22.600 --> 00:29.000 assumptions made by the model, or biases in the data used to train the model. 00:29.920 --> 00:36.960 AI hallucinations can be a problem for AI systems that can be used to make important decisions, such 00:36.960 --> 00:40.280 as medical diagnosis or financial training. 00:40.600 --> 00:43.640 Now let's understand how to hallucination occur. 00:44.240 --> 00:51.480 AI models are trained on data, and they learn to make predictions by finding patterns in the data. 00:51.880 --> 00:58.480 However, the accuracy of these predictions often depend on the quality and completeness of the training 00:58.480 --> 00:59.120 data. 00:59.520 --> 01:06.920 If the training data is incomplete, biased, or otherwise flawed, the AI model may learn incorrect 01:06.920 --> 01:11.400 patterns, leading to incorrect predictions or hallucination. 01:12.280 --> 01:19.710 For example, an AI model that is trained on a data set of medical images may learn to identify cancer 01:19.710 --> 01:20.350 cells. 01:20.710 --> 01:27.710 However, if the data set does not include any images of healthy tissue, the AI model may incorrectly 01:27.710 --> 01:31.470 predict that the healthy tissue is cancerous. 01:32.150 --> 01:34.430 Let's move on to learn this a little more. 01:34.910 --> 01:38.750 Another factor that may contribute is a lack of proper grounding. 01:39.150 --> 01:46.710 An AI model may struggle to accurately understand real world knowledge, physical properties, or factual 01:46.710 --> 01:47.790 information. 01:48.750 --> 01:55.550 This lack of grounding can cause the model to generate outputs that are seemingly possible, but are 01:55.550 --> 01:58.070 factually incorrect or irrelevant. 01:58.470 --> 02:06.350 For example, Google's Bard stated that the James Webb Space Telescope took the very first picture of 02:06.350 --> 02:15.270 a planet outside of our solar system when the first such photo was taken, 16 years before Google Webb 02:15.270 --> 02:17.630 Space Telescope was even launched. 02:18.030 --> 02:21.150 This happened in a public demo of Google Bard. 02:21.510 --> 02:24.270 So what are examples of hallucination? 02:25.230 --> 02:28.630 They're categorized into three different categories. 02:29.070 --> 02:31.270 First is incorrect predictions. 02:31.750 --> 02:36.710 An AI model may predict that an event will occur when it's unlikely to happen. 02:36.990 --> 02:44.630 For example, an AI model that is used to predict the weather may predict it will rain tomorrow when 02:44.630 --> 02:46.910 there is no rain in the forecast. 02:47.190 --> 02:52.150 There are false positives when working with an AI model. 02:52.310 --> 02:56.830 It may identify something as being a threat when it is not a threat. 02:57.230 --> 03:05.470 For example, an AI model that is used to detect fraud may flag a transaction as fraudulent when it 03:05.470 --> 03:08.750 is actually not a fraudulent transaction. 03:09.230 --> 03:16.030 The false negatives AI model may fail to identify something as being a threat, but it is actually a 03:16.030 --> 03:16.590 threat. 03:16.990 --> 03:24.710 For example, an AI model that is used to detect cancer may fail to identify a cancerous tumor. 03:25.550 --> 03:28.670 So these are some high level examples of hallucination. 03:29.070 --> 03:32.950 Let's understand how you can prevent AI hallucination. 03:33.350 --> 03:38.100 Limit possible outcomes to prevent AI model from overfitting. 03:38.300 --> 03:42.420 It's crucial to limit its prediction range using regularization. 03:42.980 --> 03:49.220 This technique penalizes extreme predictions, improving the model's accuracy on new data. 03:50.140 --> 03:58.140 The training data for effective I training use data relevant to the task, such as medical images for 03:58.140 --> 03:58.660 cancer. 03:58.660 --> 04:04.100 Identifying model irrelevant data can lead to incorrect predictions. 04:04.460 --> 04:06.980 Template of your I to follow. 04:07.460 --> 04:13.340 If you are training an AI model to write text, you could create a template that includes the following 04:13.380 --> 04:19.500 elements a title, an introduction, a body, and a conclusion. 04:20.380 --> 04:28.060 No matter how much we try, it's hard to prevent AI hallucination, but it's very important that once 04:28.100 --> 04:33.180 AI hallucination happened, we also should be able to detect it. 04:33.700 --> 04:39.300 In our next video we will learn about techniques to detect AI hallucination. 04:39.820 --> 04:40.500 Thank you. 04:40.980 --> 04:42.620 See you in the next video.