Day #182 - Hasgeek Interesting 2018 Talks

Talk #1 - Product Size Recommendation for Fashion E-commerce - lavanya TS
Key Lessons

• Product Size recommendations at Ecommerce
• Brands and Sizes issues are different across each brands, This problem is the statement taken

Data Challenges

• Sparsity
• Unavailability of data
• Incorrect return data issues

Approach

• Leverage purchase and return data
• Compute True Size, Size of product
• Fit / Large / Small - Fit the data in the intervals
• A bunch of assumptions to interpret this data / eliminate noise from data

First Approach
Loss function Approach

• Point estimate function
• Sparsity is still a problem

Bayesian Approach

• Plot Distribution
• Reflective of data
• Probablistic formulation
• Smooth transition modelled using logit function
• Generative process - Mathematical model from Gaussian (mean catalog size), Inference - From Data define the unknown variable
• Approximate Inference computation

Talk #2 - Going beyond what and asking why: Explainability in ML/DL - Vineeth N Balasubramanian

He is my Prof for my masters. Very detailed and Impressive talk.

Key Lessons
Explainability in ML

• Accuracy measure by improved revenues / improved conversions
• Complex real world systems (Medical, cockpit decision support)
• Cost of bad decision is very high in critical applications
• Interpretability
• Explainability

Visual Interpretation of CNN

• Linear Proxy Models
• Saliency models
• Automatic Rule Extraction 

LIME

• Lime gives decision of what feature led to decision
• Regress on instances of output
• Popularly used in industry to explain decisions

Visual Interpretation

• Visualize the weights 
• Interpretable only in first layer
• Backpropagation methods to interpret (with respect to input instead of weights)
• Backpropagage with respect to particular class
• Deconvolution / Guided - Backprop

Grad-CAM

• Class Activation Maps
• Take feature Maps
• Average and represent by one particular value
• Gradient based CAM
• Retraining not required

Grad-CAM++

Code




Next Talks
Learning Real-time Object Detection In The Absence of Large-scale Datasets
Sarcasm Detection: Achilles Heel of sentiment analysis - Anuj Gupta
Looking beyond LSTMs: Alternatives to Time Series Modelling using Neural Nets - Aditya Patel

Happy Mastering DL!!!