About this course: What is machine learning, and what kinds of problems can it solve? Google thinks about machine learning slightly differently -- of being about logic, rather than just data. We talk about why such a framing is useful when thinking about building a pipeline of machine learning models. We end with a recognition of the biases that machine learning can amplify and how to recognize this.
A standard introduction to online learning might place Online Gradient Descent at its center and then proceed to develop generalizations and extensions like Online Mirror Descent and second-order methods. Here we explore the alternative approach of putting exponential weights (EW) first. We show that many standard methods and their regret bounds then follow as a special case by plugging in suitable surrogate losses and playing the EW posterior mean. For instance, we easily recover Online Gradient Descent by using EW with a Gaussian prior on linearized losses, and, more generally, all instances of Online Mirror Descent based on regular Bregman divergences also correspond to EW with a prior that depends on the mirror map. Furthermore, appropriate quadratic surrogate losses naturally give rise to Online Gradient Descent for strongly convex losses and to Online Newton Step. We further interpret several recent adaptive methods (iProd, Squint, and a variation of Coin Betting for experts) as a series of closely related reductions to exp-concave surrogate losses that are then handled by Exponential Weights. Finally, a benefit of our EW interpretation is that it opens up the possibility of sampling from the EW posterior distribution instead of playing the mean. As already observed by Bubeck and Eldan, this recovers the best-known rate in Online Bandit Linear Optimization.
Talk to someone with programming skills and discuss any subject about deep learning with them so that you could quickly jump in as a newbie. Though some people figure out various libraries embedding math is used universally, you needn't understand the theory to implement deep learning tasks, I still recommend you learn some math knowledge like partial derivative. Some resources could give you a good starting point like Stanford's online course CS231n, Deep Learning at Oxford 2015and Andrew Ng's Coursera class. Also, some interesting online books like Neural Networks and Deep Learning could also give you an assistance to deep learning. Facilities and toolkits should also be available.
Whenever you are about to be oppressed, you have a right to resist oppression: whenever you conceive yourself to be oppressed, conceive yourself to have a right to make resistance, and act accordingly. In proportion as a law of any kind--any act of power, supreme or subordinate, legislative, administrative, or judicial, is unpleasant to a man, especially if, in consideration of such its unpleasantness, his opinion is, that such act of power ought not to have been exercised, he of course looks upon it as oppression: as often as anything of this sort happens to a man--as often as anything happens to a man to inflame his passions,--this article, for fear his passions should not be sufficiently inflamed of themselves, sets itself to work to blow the flame, and urges him to resistance. Submit not to any decree or other act of power, of the justice of which you are not yourself perfectly convinced. If a constable call upon you to serve in the militia, shoot the constable and not the enemy;--if the commander of a press-gang trouble you, push him into the sea--if a bailiff, throw him out of the window. If a judge sentence you to be imprisoned or put to death, have a dagger ready, and take a stroke first at the judge.
Alex Irpan, a software engineer at Google, wrote an excellent article on the current difficulties of getting deep reinforcement learning to work. For example, even after weeks of optimizing hyperparameters and explotation-exploration rates, these models are still highly sensitive to initial conditions. A 30% failure rate is seen as "working."
We introduce several new black-box reductions that significantly improve the design of adaptive and parameter-free online learning algorithms by simplifying analysis, improving regret guarantees, and sometimes even improving runtime. We reduce parameter-free online learning to online exp-concave optimization, we reduce optimization in a Banach space to one-dimensional optimization, and we reduce optimization over a constrained domain to unconstrained optimization. All of our reductions run as fast as online gradient descent. We use our new techniques to improve upon the previously best regret bounds for parameter-free learning, and do so for arbitrary norms.