Sentence embedding skip thought vectors

Machine learning metrics

Methods: Cross-validation, usually only used with smaller datasets

Evaluation Metrics Tricks: Incremental Dev Set Testing - At every ~100th iteration, make predictions on dev set read more...

Lecture 9 analysis of large graphs link analysis, pagerank

Lecture 9- Analysis of Large Graphs: Link Analysis, PageRank

• Informational Retrieval investigates: Find relevant docs in a small and trusted set.
• Link Analysis read more...

Lecture 8 latent factor recommender systems

Netflix Prize

• The winning entry factorized the problem into global (baseline), regional (SVD), and local levels (collab filtering).

18 April 2020

Lecture 7 recommender systems 1

Hybrid Recommender Systems

• Implement two or more different recommenders and combine predictions (ie linearly)
• Add content based methods to collaborative fitlering to solve read more...

Lecture 6 dimensionality reduction

Lecture 5- Dimensionality Reduction

THe first latent dimension is the direction in which the points have the highest variance. THe second is the dimension read more...

Lecture 5 clustering

Lecture 5 - Clustering

The curse of dimensionality - in large dimensional spaces, almost all points are far away from each other, and almost read more...

Lecture 4 theory of locality sensitive hashing

Lecture 4 - Theory of Locality Sensitive Hashing

LSH requires some dimensionality reduction first. You do need to do some preprocessing beforehand to get read more...

Lecture 3 locality sensitive hashing

Lecture 3 - Locality Sensitive Hashing

Locality Sensitive hashing is a brilliant use of hashing (Jure: The most brilliant), very important, part of the read more...

Lecture 2 frequent itemset mining and association rules

Lecture 2 - Frequent Itemset Mining and Association Rules

For motivation of problem see below (2). For formal definition of problem see below (3)

Lecture 1a spark and spark vs mapreduce

Lecture 1a - Spark and Spark vs MapReduce

Generally spark is faster with Hadoop but with caveats - Spark can process data in memory, read more...

Lecture 18 computational advertising 2

Explore-Exploit

• Applications
  • Web advertising (Estimating CTR)
  • Cold start for Recommendation Engines
  • Clinical Trials
  • Adaptive Routing (Package Routing)
  • Asset Pricing
  • Product Pricing
  • Promo read more...

Lecture 17 computational advertising

Def: Competitive Ratio:

• In the worst assignment, how much money you would get with a particular algorithm compared to how much money you would read more...

Lecture 16 data streams 2

Bloom Filters (Hash Filters)- Filtering a Data Stream

Applications:

• We know 1 billion ‘whitelisted email addresses’ for each user. If an email read more...

Lecture 14 svm and distributed deep learning

Use SVMs if you have very high-dimensional data (ie text). Tree based agos don’t work well with feature count > 1000.

SVM:

Assume a read more...

Lecture 13 planet  decision trees, large scale machine learning

Large Scale Machine Learning

Size of data affects model performance much more than a fancy model (even Naive Bayes performs very well). - “The read more...

Lecture 12 network embeddings (graph representations)

Network node embeddings (Graph Representation Learning)

• Goal: Efficient task-independent graph embeddings, for applications including clustering/community detection, node classification, and link prediction (Example (1)).
read more...

Lecture 11 finding network clusters commuinty detection

Intro

Applications

• Finding Micro-markets (e.g. in sponsored search), customer segmentation (10)
• Finding clusters in read more...

Lecture 10 link analysis trustrank and webspam

Lecture 10 - Link Analysis: TrustRank and WebSpam

Base PageRank. Teleport set $S$ is $S = [0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, read more...

Ee261 lecture 8, 9, 10 scales, shifts and convolutions

1. Fourier Transforms of Transformed Functions

Delays (shifts in time) change the phase of the Fourier Transform: $Ff(s-b) = e^{-2 \pi i s b}Ff(s)$.

Ee261 lecture 5,6,7 fourier transforms

0. Intro To Fourier Transforms Fourier Series - Periodic Phenomena Fourier Transforms - Nonperiodic Phenomena

View view non-periodic function as the limiting case of a read more...

Ee261 lecture 11 convergence of integrals

Fourier Analysis: This is the field of convergence of Fourier Integrals and is a subfield of Functional Analysis. Some example issues include:

1. Recall, $F(\pi(t)) read more...

Ee261 lecture 1, 2, 3, 4 fourier series

0. The Fourier Transforms and its Applications - **Intro**

The Fourier Series - the studies of periodic phenomenon.

The Fourier Transform - a limiting case read more...

Cs224n lecture 12 subword models

Lingusistics Aside

In Appendix (1). Not really important but simply because it is cool.** **

** **

Purely Character Level NMT Models Aside

In Appendix read more...

Cs234 lecture 9 policy gradient 2

Desired Properties of a Policy Gradient RL Algorithm

• Obtain large monotonic improvements with each timestep. => Make it more efficient and less noisy.
• Note read more...

Cs234 lecture 6 deep q learning and nn

Nonlinear function approximations for RL:

• Kernel Based Approaches (Convergence Properties but don’t scale well)
• Neural Networks

History of NN approaches to RL:

read more...

Cs234 lecture 5 linear value function approximation

Today we estimate the Value and Q function as a function with parameters $w$: $(V(s),Q(s,a)) \rightarrow (\hat V(s;w), \hat Q(s,a;w))$. This helps us reduce memory, read more...

Cs234 lecture 4 model free control

Backgammon, Go, Robot Locomution, Helicoper flight, Robocup soccer, Autonomous driving, Customer Ad selection, Patient Management Def: On-policy learning: Get direct experience from the world and read more...

Cs234 lecture 3 model free policy evaluation

Def: Policy Evaluation is computing the value of all states given a particular policy $\pi$. Policy evaluation can be wither model-based (requires the transition and read more...

Cs234 lecture 2 markov processes

2.1 Types of Markov Processes

Def: A state has the Markov property if and only if: $p(s_{t+1}|s_t,a_t) = p(s_{t+1}|h_t,a_t)$

Markov Processes

Def: A read more...

Cs234 lecture 12,13 fast mdps, pac (fast rl ii)

Note: Everything in this lecture is in the model-based (tabular MDP) setting. Extending it to the general model-free setting is an important on-going research topic.

Cs234 lecture 10, 11 bandits, bayesian bandits, thompson sampling (fast rl i)

Q learning is computationally efficient (needed for example, simulators, robotics), however, we also want to be sample efficient (when experience is costly or hard to read more...

Cs234 lecture 1 reinforcement learning (intro)

1.1 Intro: Reinforcement Learning

Uses: Robotics, games, healthcare, and also used as optimization technique for difficult problems

Core concepts in RL are: Optimization, Delayed Consequences, read more...

Cs231n lecture 7 training neural networks ii

Issues for Training:

• Poor conditioning (High condition number) is an issue for optimisation (Appendix 1)

• Saddle Points

• Local Minima

  read more...

Cs231n lecture 4,5 backprop, neural networks

Backprop

See (1)

History of NN

Rosenblatt Perceptron

Widrow and Hoff: 1960 Adaline /Madaline - multi-layered perceptrion

1986 Rumelhart - Backprop

Hinton and Salakhutdinov 2006

Cs231n lecture 16 adversarial examples ian goodfellow

Ian Goodfellow spent most of his career looking at adversarial examples.

Adversarial Examples are not due to overfitting: [i.e. overfitting to parameter contours that the read more...

Cs231n lecture 15 efficient methods and hardware for deep learning (songhan, tpus)

Note that appendices for this lecture are in a separate note becase of space limitations.

1. Algorithms for Efficient Inference

1.1. Deep Compression

Song Han: read more...

Cs231n lecture 15 appendices

Appendix 1: **Algorithms for Efficient Inference** **

Appendix 1.1 ** **Deep Compression (aka, Pruning, Quantization, and Huffman Encoding)

read more...

Cs231n lecture 14 reinforcement learning

Reinforcement Learning

• Problems involving an agent interacting with an environment, maximizing a reward.

RL Problem:

• At time $t$, environment is at state read more...

Cs231n lecture 13 generative models (pixelcnn, vae, gan)

Unsupervised learning allows us to use a lot more data, since unlabelled data is very abundant.

** **

Generative Models

Generative Models do densitiy estimation read more...

Cs231n lecture 12 visualizing and understanding cnns

First Layer of CNN - you can just directly visualize the weights as images

Middle Layers - However, the weights don’t match to pixels in read more...

Cs231n lecture 11 detection and segmentation

**

** **

1. Semantic Segmentaiton**

• Rather than assigning the entire image to a category, we read more...

Cs231n lecture 1 2 3 computer vision intro, basics, loss functions and optimization, old fashioned image features

Computer Vision Intro

History of Computer Vision

The onset of vision started the Cambrian Evolution.

In Humans, 50% of our neurons are involved in visual read more...

Cs231n lecture 9 10 cnn architectures, rnns

**CNN History - **

** **

AlexNet

read more...

Cs231n lecture 8 tf and pytorch tutorial

* Note this is as of 2017, may be outdated.

CPU vs GPU (Appendix 3)

Deep Learning Framework

Allows you to easily build big computational read more...

Cs231n lecture 6 training neural networks

One Time Setup:

• Activation Functions, Preprocessing, Weight Initialization, Regularization, Gradient Checking

Training Dynamics:

• Babysitting the learning process, parameter updates, hyperparameter optimization

Cs224n lecture 18 constituency parsing, treernns

TreeRNNs are hard to scale and don’t necessarily work better than LSTMs and Transformers.

Part of the tricks of being a good deep learning researcher read more...

Cs224n lecture 17 multitask learning, guest lecture richard socher

**Richard Socher - **

The Natural Langauge Decathlon- MultiTask Learning for Question Answering.

One Model to Rule All The Different Tasks.

One model, using a read more...

Cs224n lecture 16 coreference resolution

Coreference resolution: **Identify all **mentions that refer to the same real world entity. (1)

Uses of coreference resolution:

• Full text understanding, information extraction, question read more...

Cs224n lecture 10 qa systems

Question Answering

Question Answering systems typically have two parts

1. Finding documents that (might) contain an answer using standard information retrieval methods.

2. Performing Reading read more...

Cs224n tips nlp research ideas final project ideas (lecture 9)

Research ideas:

Learning Temporal Events In text - what came before, what came after

Look at the ‘residual encodings’ in NMT systems - the encoding read more...

Cs224n lecture 5 dependency parsing

Two views of linguistic structure

- Phrase Structure = **Context Free Grammars = **The Lingustics Approach. Not used much in CS.

18 April 2020

Cs224n lecture 3,4 neural nets, backprop

$\underbrace{\frac{\partial s}{\partial W}}{[n \times m]} = \overbrace{\underbrace{\delta ^T}{[n \times 1]}\underbrace{x^T}_{[1 \times m]}}^{\text{This is an outer product.}} $

Cs224n lecture 8 machine translation, seq2seq, attention

Alignment

Alighment Can be Complex! (Appendix 1)

Back in read more...

Cs224n lecture 19 20 future of nlp and deep learning

Lecture 20 - Future of NLP and Deep Learning

Kevin Clarke- Areas recently taking off

** **

Harnessing Unlabeled Data - **Backtranslation and unsupervised machine read more...

Cs224n lecture 15 natural language generation (nlg)

Teacher read more...