Papers and Reading :

-  Mikolov et al Word2vec
-  1402.3722v1 arxiv
-  stack overflow why does word2vec use 2 representations for each word
-  https://arxiv.org/pdf/1310.4546.pdf
-  Paragraph2Vec Mikolov ICML 2014
-  MRNet-Product2Vec ECML-PKDD 2017.
-  Crowd Sourcing Howe book.
-  Raykar et al JMLR 2010 EM Algo
-  Missing labels Raykar et al JMR 2010
-  Modeling task complexity Welinder et al 2010 and whitehill et al NIPS 2010. 
-  Sequential Crowdsourced labeling as MDP raykar et al
-  Deep structured semantic embedding, Microsoft Research 
-  Visw2v CVPR 2016
-  1602.05568 arxiv
-  1801.03244, ICLR 2018
-  iWGAN Guljarani et al 
-  EcommerceGAN
-  

T-SNE : To get data from high dimension to low dimension for visualizing.

For crowdsourced labeling : CrowdFlower, AmazonMechanicalTurk

Lecture 1 - Representation Learning

• The layer just before classification/regression layer is a good example of a learned representation.
• A representation is basically a feature.
• These learned representations may work well in other kind of tasks too.

Word2Vec

• Earlier words were expressed through a sparse vector , if there were 10000 words and dog was the 100th word. then the 100th value was 1 while all others were zero. This one hot encoding although worked very well, this couldn't work well for a complex task because it didnt capture semantic embeddings.
• Word2Vec : A word is known by the company it keeps.
• How do you learn these embedding? Based on the neighbors. By taking a huge text corupus , we basically see which set of words are co occurrring with each other. Done through NN, Shallow ones albeit.
• Done with continous bag of words or skip gram.
• CBOW Arch : Full connected NN, Each node is inputted a one hot encoding. (CxV,N,V) Dimension. C = Window Size.
• Skip gram Arch : The reverse Arch of CBOW ARCH.
• Skip gram works with small amount of training data, while cbow works several times faster.
• gensim is a popular way to learn word2vec.

• There's an inherent challenge for word2vec, if you have 10000 vocabulary size and your embedding is 300 dimensions then you have 3 million weights. Can use negative sampling to overcome this.

• Limitations :

  1. Sometimes vulnerable and not a robust concept.
  2. Takes a long time to train.
  3. Hard to understand and visualize.

Glove :

• Instead of using co occurence probabilities, we use ratio of co-occurrence probability.

Paragraph2vec:

• Using Distributed Memory Model(similar to cbow) and distributed bag of words model(similar to skip gram).

Lecture 2 - Learning From crowds

Crowdsourced Data Annotation

• Getting objective labels[the gold standard] is expensive, tedious, invasive and potentially dangerous.
• What we actually get is subjective labels[approx gold standard] from experts.
• But even that is expensive, therefore crowdsourcing is the way to go.
• The problem with crowdsourcing is how to consolidate these multi-annotated data.
• One way is to do majority voting, the problems with majority voting is that what if only one of them is a good labler and the rest are novices.
• So you need to have annotator models, with sensitivity and specificity in consideration.
• Use EM algorithm , the initialization is majority voting.
• A few extensions:
  1. Bayesian Approaches
  2. Variation Bayes
  3. Modeling task Complexity
  4. Missing Labels
  5. Categorical, ordinal, and continuous annotations.

Learning from crowds

• Proposed EM algo using classifer while doing m-step in EM. Soft logistic Regression in the M step

Accuracy, cost, time

• Reducing costs includes pay based on performance and removing evil annotators.
• Spammer Metric

Sequential Crowdsourced labeling

• as a Markov decision process , by incorporating cost.

Open Problems :

• Incentive Design.

Lecture 3

Deep Structured Semantic Embedding (DSSM)

• Architecture : There's a term vector and then word hashing which are put through multi-laye non linear projections which give out the semantic features, these semantic measures relevance is measured then by a cosine similarity ultimately going through a softmax layer to give an output.
• You want the query being searched for and the document which is clicked to have close embeddings to each other.
• Word hashing is a way to show a word in 3 character trigrams.

Visw2v

• Word2vec for vision.
• Training data : scenes and associated text

Med2Vec

• Distributed representatiosn for both visits and medical codes

Product 2 Vec

• All ecommerce companies typically process billions of orders.
• A semantic representations and understanding of these orders can help recommendations, order similarity and forecasting.
• A Naive representation would include tf-idf representation.
• Trained a discriminative multi-task neural network.
• Different signals pertaiing to a product were explicitly injected. Static signals like color, material, weight etc and dynamic signals like price, popularity views. The goal was to learn a general representation.
• This is referred to as a Multi-Task Recurrent Neural Network(MRNet), the tasks are to predict, color, material, price etc.
• Alternative Optimization is used it means that the gradient is computed according to the task, this task is selected randomly.
• These embeddings can be used in problems like electrical plugs classfication, Ingestible classfication etc.
• Future Work :-
  1. Incorporating more signals
  2. Product image Simulation

EcommerceGAN

Lecture 4 - AI for Fashion

• 4 Key Aspects - Search, Recommend, Discover, Engage.
• Fashion Taxonomy
• Color Taxonomy
• Research Problems - Taxonomy Curation, Taxonomy expansion, Embedding, New Fashion Themes and concepts, Siamese Style Architectures, Learning pose and background invariant representations, Aspect Based Visual search, Multi-scale attention based models for object detection, Create Outfit, Multi Modal Dialog, Normalizing sizes across brands, Color trend Analysis
• Theme Capsules Eg: Athleisure, Shacket and skort
• Thematic Quering
• Visual Fashion Tags
• Dominant Color
• h&m sells about 4.3billion dollar worth clothes since they dont have a second hand store models.
• Cognitive Couture,
• WGSM fashion
• Color trend analysis - Popular color and trending color
• Cognitive Collection - Jason Grech + IBM Watson
• Cognitive Prints - Can Computers be creative ? Falguni Shane example.
• WWD + IBM fashion article