ML Session-2
- 1. Introduction to ML Part 2 Explore ML
- 2. What is Machine Learning?
- 3. Machine Learning systems take inputs (data) to make useful predictions and decisions about previously unseen pieces of data. ML Extended PROPRIETARY + CONFIDENTIAL
- 4. Machine learning is a specific field of AI where a system learns to find patterns in examples in order to make predictions. ML Extended PROPRIETARY + CONFIDENTIAL
- 5. Computers learning how to do a task without being explicitly programmed to do so. ML Extended PROPRIETARY + CONFIDENTIAL
- 6. Machine Learning systems might: ● Label or classify data ● Predict numerical values ● Cluster similar pieces of data together ● Infer association patterns in data ● Create complex outputs
- 7. Machine Learning could be used for early dementia diagnosis Automating drone-based wildlife surveys saves time and money "Machine Learning: Why or Why not?" Read a couple of news articles involving applications of ML. 1. Would a traditional programming solution be more efficient? 1. Could a human perform the same task in less time? 1. What are the benefits of a Machine Learning model in these instances?
- 8. Model learns patterns from unlabelled data. Machine Learning Supervised Unsupervised Model is trained on labeled data stop_sign_4 stop_sign_1 stop_sign_2 stop_sign_3
- 9. See it in action! Image label verification Supervised learning Visit https://crowdsource.app to try these tasks Semantic Similarity Unsupervised learning
- 10. Loss
- 11. Predicting the Price of a House Features ● Location ● Number of bedrooms ● Size of property ● Number of light switches? ● Color of house?
- 12. Recommending which video a user should watch next Features ● Topic ● Popularity of a video/Number of views ● Creator of video ● Length of video? ● Age of video?
- 13. A baker would like to optimize pricing of cakes in their bakery depending on previous pricing, cost to make, and time of year. Checkpoint on Learning 1. What about the scenario make it suitable for ML? 2. What is the benefit to the business? 3. Would a human perform the job better? 4. What are the inputs to the system? 5. How could the ML model go wrong?
- 14. An ocean conservationist would like to track fish populations over time. Checkpoint on Learning 1. What about the scenario make it suitable for ML? 2. What is the benefit for the organization? 3. Would a human perform the job better? 4. What are the inputs to the system? 5. How could the ML model go wrong?
- 15. An online brand influencer wants a model that can predict the number of ‘likes’ that a particular post may get. Checkpoint on Learning 1. What about the scenario make it suitable for ML? 2. What is the benefit for the business or individual? 3. Would a human perform the job better? 4. What are the inputs to the system? 5. How could the ML model go wrong?
- 16. Create your own scenario Checkpoint on Learning 1. What about the scenario make it suitable for ML? 2. What is the benefit for the business or individual? 3. Would a human perform the job better? 4. What are the inputs to the system? 5. How could the ML model go wrong?
Editor's Notes
- #2: This session will go into a little more detail about machine learning. It is helpful if you have already completed some introduction to machine learning but if you have any questions please ask.
- #3: Let's review. How would you explain machine learning? Answers could include (shown on the following slides): Machine learning is a specific field of AI where a system learns to find patterns in examples in order to make predictions. Computers learning how to do a task without being explicitly programmed to do so.
- #4: Here's one definition.
- #5: Here's another definition.
- #6: And another definition.
- #7: Here is a list representing some of the things a machine learning system can be trained to do. Question: What are some examples of features and products that you assume use machine learning? How could you tell it was ML?
- #8: Use your computer and go to news.google.com or some other search engine to find and read a couple of online news articles involving applications of ML. Note: Sometimes news stories refer to ML as artificial intelligence (AI). While you are reading, think about these questions. Give learners 20 minutes to complete this task.
- #9: Machine learning looks for patterns in data. The majority of ML applications today are supervised learning where labeled data is used to teach the model. This includes tasks like classification and regression. Supervised machine learning is analogous to a student learning a tricky maths concept by studying a set of questions and their corresponding answers. After mastering the mapping between questions and answers, the student can then provide answers to new (never-before-seen) questions on the same topic. The signs on the left are labeled data so a supervised machine learning system could learn from these labels what stop signs around the world look like. The signs on the right are unlabeled data. Even the sign with the word stop because labeled data requires a human's rating to indicate what it is. Images: https://wikipedia.org/wiki/Stop_sign https://wikipedia.org/wiki/Comparison_of_European_road_signs
- #11: In the previous slide, I said supervised machine learning is analogous to a student taking a test. Let's say I created the 4 machine learning regression models above. Which one is the best? It depends on your goal and what variable you are trying to optimize for. To grade how well a model is doing on its test, machine learning practitioners measure the distance between the model's prediction (indicated in these graphs by the blue line) and the example data. This is known as loss. [Click to animate slide] With the loss displayed. Which model is best at achieving the goal? [Click to animate slide] Model #3 has the lowest loss which indicates this model is best at achieving the goal.
- #12: Features are the variables which distinguish one example from another. They tell the machine learning model what parts of the data to look for patterns for achieving the goal. The first three variables would probably help the model determine a home price the other two probably would not. So lots of data is crucial to a machine learning system but it needs to be helpful and relevant data. Though you never know until you experiment to see what variables truly make an impact.
- #13: Ask learners to come up with a list of features which might be helpful in recommending the next video to watch. The last two might turn out to be very important variables, there's no way to know unless you experiment.
- #14: What about the scenario make it suitable for ML? multiple variables, potentially lots of data examples, a goal to optimize What is the benefit to the business? the business will benefit by finding the ideal price for their cakes accounting for multiple variables. Would a human perform the job better? Data analysis could find the optimal value but would need to be continuously adjusted as new data comes in. The more variables that play a role, the more effort would be needed. What are the inputs to the system? Previous prices Cost to make Time of year How could the ML model go wrong? If there was not enough data It was biased in some way (e.g. only trained for certain times of year) It was mislabeled Stretch question: Why might time of year be a factor in the price of cake? A: Wedding season! Because demand is high during the summer a baker can charge more for cakes. (Other seasonal answers, such as a holiday time, also work)
- #15: What about the scenario make it suitable for ML? Using prior examples to train a system to recognize new examples. Lots of examples over time What is the benefit for the individual? Reduce time labeling photos of fish Could identify issues more quickly invasive species reduction in population Would a human perform the job better? It would be a very slow process and even experts make mistakes and have to refresh their knowledge What are the inputs to the system? Type of fish Region found Time of year Number of fish How could the ML model go wrong? If the data is not diverse enough, a fish could be misclassified Fish are not homogeneously distributed, the count could be off depending on the sampling. Stretch question for 2: What would happen if the conservationist found a fish the model had never seen before? A: ML models will always provide an answer- even if it is totally wrong!! This is why (at least for the time being) ML models still need humans to check up on their work.
- #16: What about the scenario make it suitable for ML? Lots of examples of data Can run experiments to see which variables have an effect What is the benefit for the business or individual? The model can get the most visibility for their customer Would a human perform the job better? With so many variables, a human could get a sense for optimal times for a specific situation but it would become complex if the product, influencer, audience, etc were changed What are the inputs to the system? Potential variables could include: Time of day, person posting, brand, product, social network How could the ML model go wrong? If trained only for a specific type of post, product, etc it might not generalize to another scenario If the training data is for the wrong demographic (gender, age, marital status, etc) the patterns might not be relevant.
- #17: Give learners 2 minutes to create their own scenario. Have them turn to the person next to them and present the scenario. The other person has 2 minutes to answer the questions on the slide. Learners swap roles and repeat steps 2 and 3.