Timezone: »

 
Oral
Learning One Convolutional Layer with Overlapping Patches
Surbhi Goel · Adam Klivans · Raghu Meka

Thu Jul 12 04:30 AM -- 04:50 AM (PDT) @ K1
in Deep Learning (Theory) 3 »
We give the first provably efficient algorithm for learning a one hidden layer convolutional network with respect to a general class of (potentially overlapping) patches under mild conditions on the underlying distribution. We prove that our framework captures commonly used schemes from computer vision, including one-dimensional and two-dimensional ``patch and stride'' convolutions. Our algorithm-- {\em Convotron}-- is inspired by recent work applying isotonic regression to learning neural networks. Convotron uses a simple, iterative update rule that is stochastic in nature and tolerant to noise (requires only that the conditional mean function is a one layer convolutional network, as opposed to the realizable setting). In contrast to gradient descent, Convotron requires no special initialization or learning-rate tuning to converge to the global optimum. We also point out that learning one hidden convolutional layer with respect to a Gaussian distribution and just {\em one} disjoint patch $P$ (the other patches may be arbitrary) is {\em easy} in the following sense: Convotron can efficiently recover the hidden weight vector by updating {\em only} in the direction of $P$.
[ PDF]  [ Video] 

Author Information

Surbhi Goel (University of Texas at Austin)
Adam Klivans (University of Texas at Austin)
Raghu Meka (UCLA)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors