IntroductionStarting an AI projectWorkflow of projectsSelecting AI projectsOrganizing data and team for the projectsWorkflow of a machine learning projectHow do you build, say a speech recognition engineKey Steps:Collect Data: people saying "Alexa", and other wordsTrain model: learns A to B mapping... audio clip to "word"many iterationsDeploy the model: implement in to a smart speakerWill collect new data (get data back), to maintain /update the modelHow do you build, say a self driving carKey steps:Collect Data: images - > positions of other cars, draw rectangles around carsTrain model: need to iterate and precisely identify carsDeploy model: may learn that golf carts are identified and positions well. keep iterating.Workflow of a data science projectoutput: actionable insightsOptimize a sales funnelKey steps:Collect Data: where are people coming from, time of day, machines type, etsc...Analyze the data: Iterate many time to get good insights insights from the data collected.Suggest hypotheses/actions: Deploy changes, re-analyze new data periodically.Optimizing a manufacturing lineKey steps:Collect Data: clay supplier, mixing time, ingredients, lead times, relative humidity, temperature, kiln duration, etc...Analyze the data: Iterate many time to get good insights insights from the data collected. Suggest hypotheses/actions: Deploy changes, re-analyze new data periodically.Every job function needs to learn how to use dataUse data to optimize workflows through data science based analysis, and to take on tasks with machine learning (remember less than a second), Inputs (A) to Output (B).From Sales, recruiting, marketing, to agriculture, and beyond DS and ML are having huge impactsHow to choose an AI projectBring together a cross-functional team knowledgeable in AI, plus domain experts.Brainstorming framework:Think about automating "tasks," vs automating "jobs."what are the main drivers of business values?What are the main pain point in your business?Note: you can make progress without big dataHaving more data almost never hurtsData makes some business [Google, Facebook, Netflix, Amazon] defensible.With small datasets, you can still make progress. The amount of data you need is problem dependant.Due diligence on projectWhat AI can do + Valuable for your businessTechnical diligenceCan AI system meet desired performance (e.g. accuracy, speed, etc)How much data is need to meet performance goalsEngineering timlineBusiness diligenceCurrent business: Lower costsCurrent business: Increase revenue ( getting more people to check out)New business: New product or business*Ethical diligence*money vs impact on societyBuild vs. buyML projects can be in-house or outsourcedDS projects are more commonly in-houseSome things will be industry standard, avoid building those."Don't sprint in front of a train."some times it makes sense to adopt another's platform or approach than to build your own. resource constraints, capability constraints...Working with an AI teamSpecify your acceptance criteriaGoal: defects with 95% accuracy...How do you measure accuracyTest Set (n1000): labelled training dataset to measure performance. Training Set: Pictures with labelsLearn mapping from A to BTest Set: Another data set to test the mappings. Often more than 1 test set will be requested.Pitfall of expecting 100% accuracy. Discuss with AI engineers what's reasonable.Limitations of MLInsufficient dataMislabeled dataAmbiguous labelsTechnical Tools for AI teamsCPU vs. GPU [Great for deep Learning/Neural Networks] NvidiaCloud vs. On-prem, ....Edge [Processor, where data is collected.]
0
0
