Publications and Preprints

Cautious Deep Learning [Paper]
Yotam Hechtlinger, Barnabás Póczos, Larry Wasserman

Main Idea: Standard classification methods are designed to predict an output for all input, and are usually incapable of returning "I don't know". For example, a classifier trained on ImageNet dataset predicts Jackson Pollock painting on the right as a "Coil" with a probability of 0.91. This overconfident prediction originates from the normalization term in P(Y|X). In this paper we suggest a cautious method to do classification based on conformal prediction which utilizes P(X|Y) instead. This method only associate an observation with a class if there are similar observations that support this classification.

A generalization of Convolutional Neural Networks to Graph-Structured Data [Paper] [Code]
Yotam Hechtlinger, Purvasha Chakravarti, Jining Qin

Main Idea: The convolution used in Convolutional Neural Networks (CNN) implicitly assumes that spatially close variables are the most correlated, requiring the data to have a grid structure. It is possible to generalize the convolution to data sets lacking the grid structure by learning the data graph structure (e.g. using the Correlation matrix) and convolve each variable with its closest neighbors, selected by a random walk on the data graph. This provides a generalization of CNN to new data structures.

Interpretation of Prediction Models Using the Input Gradient [Paper]
Yotam Hechtlinger

Main Idea: We suggest a simple method to interpret the behavior of any predictive model, both for regression and classification. Given a particular model, the information required to interpret it can be obtained by studying the partial derivatives of the model with respect to the input. We exemplify this insight by interpreting convolutional and multi-layer neural networks in the field of natural language processing.

Advances in Neural Information Processing Systems (NIPS), Interpretable Machine Learning Workshop, 2016
Most effective expression
Positive Sentiment Negative Sentiment
fantastic film total unfortunately they suffer
moving highly entertaining dull after five
unconvincing i can't painfully dull after

Discussion: An estimate of the science-wise false discovery rate and applications to top medical journals by Jager and Leek [Paper]
Yoav Benjamini, Yotam Hechtlinger

Main Idea: In a paper by Jager and Leek, the authors estimate the Science-Wise FDR of medical research at 14% by analyzing the P-values distribution in top medical journals. In a discussion we have been asked to write on the paper, we address possible drawbacks with the analysis, suggest ways to improve the estimation of the Science-Wise FDR and consider the multiplicity problem of selective inference scientists must face.

Biostatistics, 2013


I was born and raised in Israel and currently reside in Pittsburgh. I enjoy Rock Climbing, Theatre and random stuff that makes you wonder where the time went. My office is in FMS 132.

Github, Linkedin and Email (yhechtli['at']stat['dot']cmu[.]edu).