1 |
One model for the learning of language.
|
|
|
|
In: Proceedings of the National Academy of Sciences of the United States of America, vol 119, iss 5 (2022)
|
|
BASE
|
|
Show details
|
|
2 |
International School Elementary Educators' Transition to a Revised Mathematics Curriculum: A Case Study
|
|
|
|
In: Doctoral Dissertations and Projects (2021)
|
|
BASE
|
|
Show details
|
|
3 |
Investigating Global Englishes-oriented teacher development through the lens of transformative learning theory: A study of South Korean elementary EFL teachers’ experiences ...
|
|
Choi, Ko Un. - : Apollo - University of Cambridge Repository, 2021
|
|
BASE
|
|
Show details
|
|
5 |
A theoretical journey from social constructivism to digital storytelling
|
|
|
|
BASE
|
|
Show details
|
|
7 |
Understanding Teacher Experience with Instructional Coaching to Inform Program Improvement Reflecting the Principles of Adult Learning Theory: A Mixed Method Case Study
|
|
|
|
In: Graduate Theses & Dissertations (2020)
|
|
BASE
|
|
Show details
|
|
8 |
Towards provably efficient algorithms for learning neural networks ...
|
|
|
|
BASE
|
|
Show details
|
|
9 |
Using a blend of the Knowledge Framework and the Project Framework to develop materials for teaching public speaking
|
|
|
|
In: Creative Components (2020)
|
|
BASE
|
|
Show details
|
|
10 |
Towards provably efficient algorithms for learning neural networks
|
|
|
|
Abstract:
Neural networks (NNs) have seen a surge in popularity due to their unprecedented practical success in fields such as computer vision, robotics, and natural language. Developing provably efficient algorithms for learning commonly used neural network architectures continues to be a core challenge in understanding deep learning. In particular, even the problem of learning very basic architectures remains open. The underlying difficulty arises from the highly non-convex nature of the optimization problems posed by neural networks. Despite their practical success, the standard neural network training algorithms based on gradient descent (GD) and its variants have almost no provable guarantees. This necessitates a paradigm shift towards developing new principled algorithms with provable guarantees. In this thesis, we give the first set of efficient algorithms for learning commonly studied neural network architectures under minimal assumptions. In the first part of the thesis, we focus on characterizing the computational complexity of learning a single non-linear unit. We combine techniques from kernel methods and polynomial approximation to give the first dimension-efficient algorithm for learning a single ReLU (rectified linear unit), the most popular activation function, in the agnostic learning model (arbitrary noise) for any distribution on the unit sphere. We further show that if the input distribution is assumed to be Gaussian, the problem is hard. Our results unconditionally imply that GD cannot agnostically learn a single ReLU. Lastly, we show that if we relax our learning guarantee, then there is a fully polynomial time algorithm that achieves a constant factor approximation for all isotropic log-concave distributions. We further extend our results to shallow NNs. We give the first dimension efficient algorithm for learning norm-bounded one layer fully connected NNs. We subsequently show that if the marginal distribution on the input exhibits sufficient eigenvalue decay (low dimensional structure), then one-hidden-layer NNs can be learned in polynomial time in all parameters. For one hidden layer convolutional NNs, we propose a simple iterative algorithm that efficiently recovers the underlying parameters for commonly used convolutional schemes from computer vision. We further give the first polynomial time algorithm for networks with more that one hidden layer in a weaker noise model. The techniques from this work also give improved results for problems related to boolean concept learning. Lastly, we shift focus to the unsupervised learning setting through the lens of graphical models. We study Restricted Boltzmann Machines (RBMs) which are simple generative neural networks that model a probability distribution. We give the first algorithm for learning RBMs with non-negative interactions under arbitrary biases on binary as well as non-binary input. ; Computer Sciences
|
|
Keyword:
Efficient algorithms; Learning theory; Neural networks
|
|
URL: https://doi.org/10.26153/tsw/13349 https://hdl.handle.net/2152/86398
|
|
BASE
|
|
Hide details
|
|
11 |
Understanding (De)alienation in Social Movements: Resident Activism and Anti-poverty Organizing in the Toronto-area
|
|
|
|
BASE
|
|
Show details
|
|
12 |
Exploring the constructivist mentoring program in developing EFL teacher professionalism: A qualitative approach
|
|
|
|
In: Indonesian Journal of Applied Linguistics, Vol 10, Iss 1, Pp 132-142 (2020) (2020)
|
|
BASE
|
|
Show details
|
|
13 |
Evaluating the impact of online discussion boards on student engagement with group work
|
|
|
|
In: Delaney, D, Kummer, T-F, Singh, KH, (2019). Evaluating the impact of online discussion boards on student engagement with group work. British Journal of Educational Technology, Vol. 50, No. 2, p. 902-920 http://dx.doi.org/10.1111/bjet.12614 (2019)
|
|
BASE
|
|
Show details
|
|
15 |
Language coaching for speakers of Spanish as a heritage language : novice-expert interactions via videoconference ...
|
|
|
|
BASE
|
|
Show details
|
|
17 |
Foreign language classroom communities of practiceS: An extended social learning theory for classroom foreign language education
|
|
|
|
BASE
|
|
Show details
|
|
18 |
How recorded audio-visual feedback can improve academic language support
|
|
|
|
In: Journal of University Teaching & Learning Practice (2019)
|
|
BASE
|
|
Show details
|
|
19 |
The Power of Transformation: A Grounded Theory Study of Cultivating Teacher Growth Mindset towards Student Intelligence
|
|
|
|
In: Doctoral Dissertations and Projects (2018)
|
|
BASE
|
|
Show details
|
|
20 |
Language coaching for speakers of Spanish as a heritage language : novice-expert interactions via videoconference
|
|
|
|
BASE
|
|
Show details
|
|
|
|