#### Goal:

Predict interventions on ICU patients using LSTM and CNN. 

#### Dataset:

MIMIC-III v.1.4 https://mimic.physionet.org/

+ Patients over 15 years of age with intensive care stay between 12h and 240h. (Only the first stay is considered for each patient) - 34148 unique records.
+ 5 static variables.
+ 29 vital signs and test results.
+ Clinical notes of patients (presented as time series).

#### Feature Engineering:

+ Topic Modeling of clinical notes: Vector of topics using Latent Dirichlet Allocation (LDA)
+ Physiological Words: Vital / Laboratory results converted to z-scores - [integer values ​​between -4 and 4] and score is one-hot encoded (each vital / lab is replaced by 9 columns). It is good idea to avoid the imputation of missing values as the physiological word in this case is the all-zero vector.

Feature vector:
+ is the concatenation of the static variables, physiological words for each vital/lab and the  topic vector.
+ 1 feature vector / patient / hour.
+ 6-hour slice used to predict a 4-hour window after a 6-hour gap. All the features values are normalized between 0 and 1. (static variables are replicated).

#### Target Classes:

For some of the procedures to be predicted there are 4 classes:
+ Onset: Y goes from 0 to 1 during the prediction window.
+ Wean: Y goes from 1 to 0 during the prediction window.
+ Stay On: Y stays at 1 throughout prediction window.
+ Stay Off: Y stays at 0 for the entire prediction window.

#### Setup of the Experiments:

+ Dataset Split: 70% training, 10% validation, 20% test.

Long Short Term Memory (LSTM) Networks:
+ Dropout P(keep) = 0.8, L2 regularization.
+ 2 hidden layers: 512 nodes in each.

Convolutional Neural Networks
+ 3 different temporal granularities (3, 4, 5 hours). 64 filters in each.
+ Features are treated as channels. 1D temporal convolution.
+ Dropout between fully connected layers. P (keep) = 0.5.

TensorFlow 1.0.1 - Adam optimizer. Minibatches of size 128. 
Validation set used for early stopping (metric: AUC).

#### Results:

+ Baseline for comparison: L2-regularized Logistic Regression
+ Metrics: 
 + AUC per class.
 + AUC macro = Arithmetic mean of AUC per class.

+ Proposed architectures outperforms baseline.
+ Physiological words improve performance (especially on high class imbalance scenario).

#### Model Interpretability:
+ LSTM: feature occlusion like analysis. The feature is replaced by uniformly distributed noise between 0 and 1 and variation in AUC is computed.
+ CNN: analysis of the maximally activating trajectories.

#### Positive Aspects:
+ Relevant work: In the healthcare domain is very important to anticipate events.
+ Built on top of rich and heterogeneous data: It leverages large amounts of ICU data. 
+ The proposed model is not a complete black-box. Interpretability is crucial if the system is to be adopted in the future. 

#### Caveats:
+ Some of the methodology is not clearly explained:
 + How the split of the dataset was performed? Was it on a patient-level?
 + When testing the logistic regression baseline it is not clear how the feature vector was built. Was it built by simply flattening the 6-hour chunk?
 + For the raw data test, it was not mentioned the way the missing values were treated. 

The Escherichia coli chemotaxis network is a model system for biological signal processing. In E. coli, transmembrane receptors responsible for signal transduction assemble into large clusters containing several thousand proteins. These sensory clusters have been observed at cell poles and future division sites. Despite extensive study, it remains unclear how chemotaxis clusters form, what controls cluster size and density, and how the cellular location of clusters is robustly maintained in growing and dividing cells. Here, we use photoactivated localization microscopy (PALM) to map the cellular locations of three proteins central to bacterial chemotaxis (the Tar receptor, CheY, and CheW) with a precision of 15 nm. We find that cluster sizes are approximately exponentially distributed, with no characteristic cluster size. One-third of Tar receptors are part of smaller lateral clusters and not of the large polar clusters. Analysis of the relative cellular locations of 1.1 million individual proteins (from 326 cells) suggests that clusters form via stochastic self-assembly. The super-resolution PALM maps of E. coli receptors support the notion that stochastic self-assembly can create and maintain approximately periodic structures in biological membranes, without direct cytoskeletal involvement or active transport.

Reward functions are a funny part of modern reinforcement learning: enormously salient from the inside, if you’re coding or working with RL systems, yet not as clearly visible from the outside perspective, where we just see agents playing games in what seem to be human-like ways. Just seeing things from this angle, it can be easy to imagine that the mechanisms being used to learn are human-like as well. And, it’s true that some of the Atari games being examined are cases where there is in fact a clear, explicit reward in the form of points, that human players would also be trying to optimize. But in most cases, the world isn’t really in the habit of producing clear reward signals, and it definitely doesn’t typically do so on time scales that account for most of the learning humans do. 

So, it’s generally hypothesized that in addition to updating on (sparse) environmental rewards, humans also operate according to certain pre-coded, possibly evolutionarily-engineered heuristics, of which one is curiosity. The intuition is: it sure seems like, especially early in life, humans learn by interacting with objects purely driven by curiosity, and we’d love to somehow harness that same drive to allow our learning systems to function in environments lacking dense, informative reward signals. One such environment is the video game Montezuma’s Revenge, which in addition to being amusingly difficult to search for, is a game with sparse, long-range rewards, on which typical reward-based agents have historically performed poorly, and on which this current paper focuses. 

A strong existing tradition of curiosity objectives focuses on incentivizing agents to be able to better predict the next observation, given the current observation and their action within it. Intuitively, by training such a network on historical observations, and giving agents a bonus according to that prediction’s error on a given observation. The theory behind this is that if an agent isn’t able to predict the observation-transition dynamics at a given state, that probably means it hasn’t visited many nearby states, and so we want to incentivize it doing so to gain information. If this sounds familiar to the classic “explore vs exploit” trade-off, it’s very much a similar idea: in cases of clear reward, we should take the reward, but in cases of low or uncertain reward, there’s value to exploration. 

One difficulty of systems like the one described above is that they reward the agent for being in environments where the next observation is difficult to predict from the current one. And while that could describe novel states about which the agent needs to gain information, it can also describe states that are inherently stochastic; the canonical example being random static on a TV screen. The agent has a lot of trouble predicting the next observation because it’s fundamentally non-deterministic to a greater degree than even the random-but-causal dynamics of most games. The proposed alternative of this paper is a little strange, but makes more sense in the context of responding to this stochasticity problem. The authors propose to create a random mapping, in the form of an initialized but untrained neural network, taking in observations and spitting out embedding vectors. Then, they incentivize their agent to go to places that have high prediction error on a network designed to predict these random embeddings. Since the output is just a function mapping, it’s deterministic with respect to observations. The idea here is that if you’ve seen observations similar to your current observation, you’ll be more able to predict the corresponding embedding, even if there’s no meaningful relationship that you’re learning. 

https://i.imgur.com/Ds5gHDE.png

The authors found that this performed well on Montezuma’s Revenge and Private Eye, but only middlingly-well on other environments. I’m a bit torn on this paper overall. On one hand, it seems like a clever idea, and I’m in general interested in seeing more work on curiosity. It does clearly seem to be capturing something that corresponds to novelty-seeking, and the agent trained using it explores a higher number of rooms than alternative options. On the other, I’m a little skeptical of the fact that it only has consistent performance in two environments, and wish there had been more comparisons to simpler forms of observation similarity, since this really does just seem like a metric of “how similar of observation vectors to this have you seen before”. I find myself wondering if some sort of density modeling could even be effective here, especially if (as may be the case, I’m unsure) the input observations are metadata rather than pixels. 

This paper feels a bit like watching a 90’s show, and everyone’s in denim and miniskirts, except it’s a 2017 ML paper, and everything uses attention. (I’ll say it again, ML years are like dog years, but more so). That said, that’s not a critique of the paper: finding clever ways to cobble together techniques for your application can be an important and valuable contribution. This paper addresses the problem of text to image generation: how to take a description of an image and generate an image that matches it, and it makes two main contributions: 1) a GAN structure that seems to merge insights from Attention and Progressive GANs in order to select areas of the sentence to inform details in specific image regions, and 2) a novel discriminator structure to evaluate whether a sentence matches an image. 

https://i.imgur.com/JLuuhJF.png

Focusing on the first of these first: their generation system works by an iterative process, that gradually builds up image resolution, and also pulls specific information from the sentence to inform details in each region. The first layer of the network generates a first “hidden state” based on a compressed representation of the sentence as a whole (the final hidden state of a LSTM text encoder, I believe), as well as random noise (typical input to a GAN). Subsequent “hidden states” are calculated by calculating attention weightings between each region of the image, and each word in the sentence, and pulling together a per-region context vector based on that attention map. (As far as I understand it, “region” here refers to the fact that when you’re at lower spatial scales of what is essentially a progressive generation process, 64x64 rather than 256x256, for example, each “pixel” actually represents a larger region of the image). I’m using quotes around “hidden state” in the above paragraph because I think it’s actually pretty confusing terminology, since it suggests a recurrent structure, but this model isn’t actually recurrent: there’s a specific set of weights for resolution block 0, and 1, and 2. This whole approach, of calculating a specific attention-weighted context vector over input words based on where you are in the generation process is very conceptually similar to the original domain of attention, where the attention query would be driven by the hidden state of the LSTM generating the translated version of some input sentence, except, here, instead of translating between languages, you’re translating across mediums. 

The loss for this model is a combination of per-layer loss, and a final, special, full-resolution loss. At each level of resolution, there exists a separate discriminator, which seems to be able to take in both 1) only an image, and judge whether it thinks that image looks realistic on it’s own, and 2) an image and a global sentence vector, and judge whether the image matches the sentence. It’s not fully clear from the paper, but it seems like this is based on just feeding in the sentence vector as additional input? 
https://i.imgur.com/B6qPFax.png
For each non-final layer’s discriminator, the loss is a combination of both of these unconditional and conditional losses. 

The final contribution of this paper is something they call the DAMSM loss: the Deep Attention Multimodal Similarity Model. This is a fairly complex model structure, whose ultimate goal is to assess how closely a final generated image matches a sentence. The whole structure of this loss is based on projecting region-level image features (from an intermediate, 17x17 layer of a pretrained Inception Net) and word features into the same space, and then calculating dot product similarities between them, which are then used to build “visual context vectors” for each word (for each word, created a weighted sum of visual vectors, based on how similar each is to the word). Then, we take each word’s context vector, and see how close it is to the original word vector. If we, again, imagine image and word vectors as being in a conceptually shared space, then this is basically saying “if I take a weighted average of all the things that are the most similar to me, how ultimately similar is that weighted average to me”. This allows there to be a “concept representation” match found when, for example, a particular word’s concept, like “beak”, is only present in one region, but present there very strongly: the context vector will be strongly weighted towards that region, and will end up being very close, in cosine similarity terms, to the word itself. By contrast, if none of the regions are a particularly good match for the word’s concept, this value will be low. DAMSM then aggregates up to an overall “relevance” score between a sentence and image, that’s simply a sum over a word’s “concept representation”, for each word in a sentence. It then calculates conditional probabilities in two directions: what’s the probability of the sentence, given the image (relevance score of (Sent, Imag), divided by that image’s summed relevance with all possible sentences in the batch), and, also, what’s the probability of the image, given the sentence (relevance score of the pair, divided by the sentence’s summed relevance with all possible images in the batch). 

In addition to this word-level concept modeling, DAMSM also has full sentence-level versions, where it simply calculates the relevance of each (sentence, image) pair by taking the cosine similarity between the global sentence and global image features (the final hidden state of an encoder RNN, and the final aggregated InceptionNet features, respectively). All these losses are aggregated together, to get one that uses both global information, and information as to whether specific words in a sentence are represented well in an image. 

The paper proposes integrating a pre-trained language model into a sequence labeling model. The baseline model for sequence labeling is a two-layer LSTM/GRU. They concatenate the hidden states from pre-trained language models onto the output of the first LSTM layer. This provides an improvement on NER and chunking tasks.

https://i.imgur.com/Hso3mL9.png