Tensorly: A Flexible Python Framework for Machine Learning
summary
Tensorly is a Python front end that wires to NumPy SciPy or PyTorch TensorFlow and MXNet backends and provides Jupyter notebooks that define tensorized neural networks and tensor decompositions such as Tucker and tensor train. The claim is that replacing fully connected layers with low‑rank tensor contractions yields up to 65% parameter reduction while retaining accuracy, and custom primitives like strided batched matmul approach or exceed BLAS performance. Experiments run on SageMaker/AWS show faster topic modeling and improved long‑term forecasting.
AI-written summary · about
chapters (12)
- 0:00Keynote introduction
- 2:12Tensor vs matrix algebra
- 5:26Tensorly for tensor computations
- 10:18Tensors in ML relevance
- 16:05Tensorized neural networks
- 19:44Tensorly architecture
- 22:17Optimized tensor primitives
- 26:57Tensors in time series
- 31:15Multimodal feature fusion
- 33:47Topic modeling with tensors
- 41:55Tensors in robotics
- 43:39Future of tensor research
related talks
28mNext-generation frameworks for Large-scale Machine Learning
44mTensorflow, Tensorflow, What is it that you see?
07mWTT: What the Tensor?
15mEnabling Tensor Language Model to Assist in Generating High-Performance Tensor...
38mCutting Edge TensorFlow: New Techniques
30mTensor Logic: The Language of AI
20mTensorFlow.js Bringing Machine Learning to the Web and Beyond
10mDale Smith, Data Scientist, Nexidia