← Back to Home

Papers i'm reading #1

I’ve recently set myself the goal of reading one academic paper a week relating to the ML/AI things i’m working on i’m my current role.

To try help keep me honest and diligent in this regard, I’ve decided to get into the habit of jotting down some quick notes on each paper and every now and then as i get through a batch of them, stick them into blog post (because i like to try squeeze everything and anything into a blog post if i can get away with it, even better if is minimal extra effort on my part :) ).


Anomaly Detection in Streaming Non-stationary Temporal Data

Link

My Summary: Really interesting paper and application, considers a lot of different design aspects in it. Nice example of a different approach leveraging feature extraction and statistical techniques to get the job done.

Notes:


A comparative evaluation of unsupervised anomaly detection algorithms for multivariate data

Link

My Summary: Very good reference paper for more traditional methods as opposed to deep learning based approaches. Good discussion on complexity and online setting too. Primarily concerned with traditional tabular data as opposed to time series but still some good ideas to pursue.

Notes:


Deep Learning for Anomaly Detection: A Survey

Link

My Summary: A looot of references and got some good ideas out of it. Not much else to it. 

Notes:


Time2Vec: Learning a Vector Representation of Time

Link

My Summary: Nice little paper and idea of learning the frequency functions and representations seems really interesting. 

Notes:


catch22: CAnonical Time-series CHaracteristics

Link

My Rating: 8/10

My Summary: Well done paper, limited application potentially to an online setting but great food for thought on the range of ts feature transformations literature already out there.

Notes:


Time Series Anomaly Detection; Detection of anomalous drops with limited features and sparse examples in noisy highly periodic data

Link

My rating: 6/10

My summary: Good example of simple regression based approach, not very generalisable, data and results not really powerful. 

Notes:


Time-series anomaly detection service at Microsoft

Link

My rating: 8/10

My summary: Good walkthrough of end to end service, interesting computer vision application, some good leads to follow up. 

Notes:

← Back to Home