MLconf SF 2015
2015-11-13
San Francisco, United States
[MLconf SF 2014, MLconf SF 2015, MLconf SF 2016, MLconf SF 2017, MLconf SF 2019]
who it's for
research scientists, machine learning engineers, data scientists, deep learning engineers, engineering managers, platform engineers, product managers
topics
parameter server for large models; MXNet on AWS GPUs; visual question answering; distributed functional Python runtime; cross-compiled Python for cloud; cause-effect pairs causality; network deconvolution for networks; TensorFlow sequence-to-sequence models; large-scale training on thousands of machines; elastic net brain biomarkers; Numenta anomaly detection benchmark; tensor decomposition for latent-variable models
| title | author | length |
|---|---|---|
| Alex Smola, Professor in the Machine Learning Department, Carnegie Mellon University | Alex Smola | 42m |
Distributed key‑value Parameter Server coordinator plus careful single‑machine caching and partitioning let large models run fast: 100 billion variables and… | ||
| Braxton McKee, CEO & Founder, Ufora | Braxton McKee | 21m |
An automatically distributed, implicitly parallel Python platform that cross‑compiles a restricted, purely functional subset of Python and runs on AWS spot… | ||
| Isabelle Guyon, President, ChaLearn | Isabelle Guyon | 25m |
ChaLearn organizes causality challenges that convert network reconstruction into concrete tests: a 2013 cause-effect pairs challenge supplies 20% real and 80%… | ||
| Quoc Le, Software Engineer, Google | Quoc Le | 26m |
Code and models are released at tensorflow.org, with full source so researchers can reproduce results. A sequence-to-sequence recurrent model handles… | ||
| Irina Rish, Research Staff, IBM T.J. Watson Research Center | Irina Rish | 32m |
Uses elastic net and group-sparse graphical models to get interpretable brain biomarkers, with elastic net reaching about 0.8 correlation to self-reported pain… | ||
| Allison Gilmore, Data Scientist, Ayasdi | Allison Gilmore | 20m |
Uses a lens function (for example y-axis projection or a density estimator), creates overlapping value bins, clusters each bin, and outputs a graph whose nodes… | ||
| Subutai Ahmad, VP of Research, Numenta | Subutai Ahmad | 19m |
Numenta publishes an open-source Numenta Anomaly Detection Benchmark (NAB) that streams ~350,000 points across 47 real-world datasets and 11 artificial… | ||
| Xavier Amatriain, VP of Engineering, Quora | Xavier Amatriain | 42m |
Presents ten concrete lessons from building Quora’s ML stack: prefer implicit signals beat explicit (more data, but combine with explicit to avoid clickbait)… | ||
| Ben Hamner, Co-founder and CTO, Kaggle | Ben Hamner | 21m |
Kaggle reports a 410,000-strong community making about 100,000 submissions per month and shows concrete wins: a diabetic retinopathy model at ~85% agreement… | ||
| Justin Basilico, Research/ Engineering Manager, Netflix | Justin Basilico | 22m |
Netflix runs almost 70 million members in ~50 countries on 1,000+ device types streaming 3 billion hours/month and credits 80% of viewing to recommendations… | ||
| Brad Klingenberg, Director of Styling Algorithms, Stitch Fix | Brad Klingenberg | 20m |
Stitch Fix runs a two-stage system where models predict matches from client attributes, item attributes and feedback and human stylists pick the final five… | ||
| Anima Anandkumar, Electrical Engineering and CS Dept, UC Irvine | Anima Anandkumar | 22m |
Presents a concrete recipe: form higher‑order moments, run SVD on pairwise moments with random projections to whiten, then perform tensor decomposition via… | ||
| Alessandro Magnani, Data Scientist, WalmartLabs | Alessandro Magnani | 22m |
Walmart catalog labels get stored with a recorded sampling probability for each item so accuracy estimates use a self‑normalized importance sampling estimator… | ||
| Narayanan Sundaram, Research Scientist, Intel Labs | Narayanan Sundaram | 25m |
GraphMat implements a vertex-centric front end but compiles those primitives to generalized sparse matrix-vector and matrix-matrix operations, yielding a… | ||
| Melanie Warrick, Deep Learning Engineer, Skymind.io | Melanie Warrick | 20m |
Explains how attention lets models focus on parts of input rather than process everything, contrasting soft attention/backprop which uses the full input and… | ||
| Neural Networks Demystified: Part 1 | 04m | |
Shows how to build a working neural regressor in Python and NumPy to predict test score from hours studied and slept, scaling inputs to [0,1], using exactly… | ||
| Neural Networks Demystified: Part 2 | 06m | |
Implementation of a small feedforward net in Python: 2 inputs, 3 hidden units, 1 output; weights initialized randomly and learned parameters separate from… | ||
| Neural Networks Demystified: Part 3 | 07m | |
Builds on a simple Python network and shows exactly how cost is measured as the sum of squared errors with a 1/2 factor then minimized by changing weights… | ||
| Neural Networks Demystified: Part 4 | 08m | |
Derives exact gradients DJ/DW1 and DJ/DW2 for a 2-layer network, showing DJ/DW matrices match W1 and W2 shapes and the output is a 3x1 backprop error delta3… | ||
| Neural Networks Demystified: Part 5 | 04m | |
Explains a practical method to verify backprop gradients by computing a numerical gradient check: perturb each weight by a small epsilon perturbation, evaluate… | ||
| Neural Networks Demystified: Part 6 | Grant Sanderson | 05m |
Builds and trains a neural network using the BFGS optimizer via scipy.optimize.minimize, wrapping the network so the objective returns cost and gradients and… | ||
| Neural Networks Demystified: Part 7 | Grant Sanderson | 06m |
A neural net maps hours slept and hours studied to predicted test score and initially fits the training data so tightly it produces implausible dips where more… | ||
| nothing matches | ||