The Michigan Student Artificial Intelligence Lab (MSAIL) is a student organization for discussion of artificial intelligence and machine learning. Andrew Ng said:
“ ...if you read research papers consistently, if you seriously study half a dozen papers a week and you do that for two years, after those two years you will have learned a lot... But that sort of investment, if you spend a whole Saturday studying rather than watching TV, there's no one there to pat you on the back or tell you you did a good job. ” — Andrew NgMSAIL is a community in which motivated students can read and discuss modern machine learning literature together. We welcome students of all backgrounds and ability. To join MSAIL and stay up to date, simply join our Slack team! Also be sure to check out our sister organization: the Michigan Data Science Team! We are both graciously sponsored by the Michigan Institute for Data Science.
Image classification models have become quite good at predicting what is in a given image. In fact, many models even outperform reported human classification performance on the ImageNet dataset. But can models generalize to unseen labels? If we only train on images of dogs and dogs with bones, can we correctly classify an image of just a bone? Label-Set Operations attempt to do this by learning a feature space and set operations functions that can perform the intersection, union, and difference of label sets. This is useful for automatically generating data for unseen labels that will help a model generalize better.
Most of the learning mechanisms we have discussed in MSAIL fall into two categories: supervised or unsupervised. In a sense, we either learn from feeding in data to models and telling them what the correct output should be or we attempt to recognize patterns inherent to the data presented to us to make a prediction. Reinforcement learning is a third category on its own and Q-learning is an implementation that falls into that category. The general premise is that we put a model in an open environment with a constrained set of actions it can take and attempt to learn a function that will predict the long-term value of each action taken. It does this by experimenting in the environment over and over again, until it can discern meaningful patterns. Deep Q-learning is an attempt to l earn this value function through a neural architecture. This week we discusssed this topic of Deep Q-learning, and furthermore how it can be used to solve difficult problems, like training a model to play Mariro.
A robot carrying out a natural language instruction has been a fantasy since the Jetsons cartoon show. However, recent advances in vision and language methods have made large strides towards this being reality. This week we discussed this topic of vision-and-language navigation. Vision-and-language navigation is the act of a robot interpreting a natural language command based on what it sees.
This week we discussed the release of a TensorFlow version of PoseNet, a machine learning model which allows for real-time human pose estimation in the browser. Pose estimation is a Computer vision technique that detects human figures (elbows, arms, legs, etc.) in images and video. Its use cases range from augmented reality, to animation, to fitness.
This week we also began forming project teams. The projects are student-led and range in topics from natural language processing, to computer vision, to medical AI. If you wish to join one of the project teams message one of us on administration in the Slack channel, and we will put you on a team!