Originally posted on my Github repo [paper-notes](https://github.com/karpathy/paper-notes/blob/master/vin.md).

# Value Iteration Networks

By Berkeley group: Aviv Tamar, Yi Wu, Garrett Thomas, Sergey Levine, and Pieter Abbeel

This paper introduces a poliy network architecture for RL tasks that has an embedded differentiable *planning module*, trained end-to-end. It hence falls into a category of fun papers that take explicit algorithms, make them differentiable, embed them in a larger neural net, and train everything end-to-end.

**Observation**: in most RL approaches the policy is a "reactive" controller that internalizes into its weights actions that historically led to high rewards.

**Insight**: To improve the inductive bias of the model, embed a specifically-structured neural net planner into the policy. In particular, the planner runs the value Iteration algorithm, which can be implemented with a ConvNet. So this is kind of like a model-based approach trained with model-free RL, or something. Lol.

NOTE: This is very different from the more standard/obvious approach of learning a separate neural network environment dynamics model (e.g. with regression), fixing it, and then using a planning algorithm over this intermediate representation. This would not be end-to-end because we're not backpropagating the end objective through the full model but rely on auxiliary objectives (e.g. log prob of a state given previous state and action when training a dynamics model), and in practice also does not work well.

NOTE2: A recurrent agent (e.g. with an LSTM policy), or a feedforward agent with a sufficiently deep network trained in a model-free setting has some capacity to learn planning-like computation in its hidden states. However, this is nowhere near as explicit as in this paper, since here we're directly "baking" the planning compute into the architecture. It's exciting.


## Value Iteration

Value Iteration is an algorithm for computing the optimal value function/policy $V^*, \pi^*$ and involves turning the Bellman equation into a recurrence:

![Screen Shot 2016-08-13 at 3.26.04 PM](https://raw.githubusercontent.com/karpathy/paper-notes/master/img/vin/Screen%20Shot%202016-08-13%20at%203.26.04%20PM.png)
This iteration converges to $V^*$ as $n \rightarrow \infty$, which we can use to behave optimally (i.e. the optimal policy takes actions that lead to the most rewarding states, according to $V^*$). 



## Grid-world domain

The paper ends up running the model on several domains, but for the sake of an effective example consider the grid-world task where the agent is at some particular position in a 2D grid and has to reach a specific goal state while also avoiding obstacles. Here is an example of the toy task:

 ![Screen Shot 2016-08-13 at 4.43.04 PM](https://raw.githubusercontent.com/karpathy/paper-notes/master/img/vin/Screen%20Shot%202016-08-13%20at%204.43.04%20PM.png)

The agent gets a reward +1 in the goal state, -1 in obstacles (black), and -0.01 for each step (so that the shortest path to the goal is an optimal solution).



## VIN model

The agent is implemented in a very straight-forward manner as a single neural network trained with TRPO (Policy Gradients with a KL constraint on predictive action distributions over a batch of trajectories). So the only loss function used is to maximize expected reward, as is standard in model-free RL. However, the policy network of the agent has a very specific structure since it (internally) runs value iteration.

First, there's the core Value Iteration **(VI) Module** which runs the recurrence formula (reproducing again):

![Screen Shot 2016-08-13 at 3.26.04 PM](https://raw.githubusercontent.com/karpathy/paper-notes/master/img/vin/Screen%20Shot%202016-08-13%20at%203.26.04%20PM.png)
The input to this recurrence are the two arrays R (the reward array, reward for each state) and P (the dynamics array, the probabilities of transitioning to nearby states with each action), which are of course unknown to the agent, but can be predicted with neural networks as a function of the current state. This is a little funny because the networks take a _particular_ state **s** and are internally (during the forward pass) predicting the rewards and dynamics for all states and actions in the entire environment. Notice, extremely importantly and once again, that at no point are the reward and dynamics functions explicitly regressed to the observed transitions in the environment. They are just arrays of numbers that plug into value iteration recurrence module.

But anyway, once we have **R,P** arrays, in the Grid-world above due to the local connectivity, value iteration can be implemented with a repeated application of convolving **P** over **R**, as these filters effectively *diffuse* the estimated reward function (**R**) through the dynamics model (**P**), followed by max pooling across the actions. If **P** is a not a function of the state, it would simply be the filters in the Conv layer. Notice that posing this as convolution also assumes that the env dynamics are position-invariant. See the diagram below on the right:![Screen Shot 2016-08-13 at 4.58.42 PM](https://raw.githubusercontent.com/karpathy/paper-notes/master/img/vin/Screen%20Shot%202016-08-13%20at%204.58.42%20PM.png)

Once the array of numbers that we interpret as holding the estimated $V^*$ is computed after running **K** steps of the recurrence (K is fixed beforehand. For example for a 16x16 map it is 20, since that's a bit more than the amount of steps needed to diffuse rewards across the entire map), we "pluck out" the state-action values $Q(s,.)$ at the state the agent happens to currently be in (by an "attention" operator $\psi$), and (optionally) append these Q values to the feedforward representation of the current state $\phi(s)$, and finally predicting the action distribution.



## Experiments

**Baseline 1**: A vanilla ConvNet policy trained with TRPO. [(50 3x3 filters)\*2, 2x2 max pool, (100 3x3 filters)\*3, 2x2 max pool, FC(100), FC(4), Softmax].

**Baseline 2**: A fully convolutional network (FCN), 3 layers (with a filter that spans the whole image), of 150, 100, 10 filters. i.e. slightly different and perhaps a bit more domain-appropriate ConvNet architecture.

**Curriculum** is used during training where easier environments are trained on first. This is claimed to work better but not quantified in tables. Models are trained with TRPO, RMSProp, implemented in Theano.

Results when training on **5000** random grid-world instances (hey isn't that quite a bit low?):![Screen Shot 2016-08-13 at 5.47.23 PM](https://raw.githubusercontent.com/karpathy/paper-notes/master/img/vin/Screen%20Shot%202016-08-13%20at%205.47.23%20PM.png)

TLDR VIN generalizes better.

The authors also run the model on the **Mars Rover Navigation** dataset (wait what?), a **Continuous Control** 2D path planning dataset, and the **WebNav Challenge**, a language-based search task on a graph (of a subset of Wikipedia). Skipping these because they don't add _too_ much to the core cool idea of the paper.

## Misc

**The good**: I really like this paper because the core idea is cute (the planner is *embedded* in the policy and trained end-to-end), novel (I don't think I saw this idea executed on so far elsewhere), the paper is well-written and clear, and the supplementary materials are thorough.

**On the approach**: Significant challenges remain to make this approach more practicaly viable, but it also seems that much more exciting followup work can be done in this framework. I wish the authors discussed this more in the conclusion. In particular, it seems that one has to explicitly encode the environment connectivity structure in the internal model $\bar{M}$. How much of a problem is this and what could be done about it? Or how could we do the planning in more higher-level abstract spaces instead of the actual low-level state space of the problem? Also, it seems that a potentially nice feature of this approach is that the agent could dynamically "decide" on a reward function at runtime, and the VI module can diffuse it through the dynamics and hence do the planning.  A potentially interesting outcome is that the agent could utilize this kind of computation so that  an LSTM controller could learn to "emit" reward function subgoals and the VI planner computes how to meet them. A nice/clean division of labor one could hope for in principle.

**The experiments**. Unfortunately, I'm not sure why the authors preferred breadth of experiments and sacrificed depth of experiments. I would have much preferred a more in-depth analysis of the gridworld environment. For instance:

- Only 5,000 training examples are used for training, which seems very little. Presumable, the baselines get stronger as you increase the number of training examples?
- Lack of visualizations: Does the model actually learn the "correct" rewards **R** and dynamics **P**? The authors could inspect these manually and correlate them to the actual model. This would have been reaaaallllyy cool. I also wouldn't expect the model to exactly learn these, but who knows.
- How does the model compare to the baselines in the number of parameters? or FLOPS? It seems that doing VI for 30 steps at each single iteration of the algorithm should be quite expensive.
- The authors should study the performance as a function of the number of recurrences **K**. A particularly funny experiment would be K = 1, where the model would be effectively predicting **V*** directly, without planning. What happens?
- If the output of VI $\psi(s)$ is concatenated to the state parameters, are these Q values actually used? What if all the weights to these numbers are zero in the trained models?
- Why do the authors only evaluate success rate when the training criterion is expected reward?



Overall a very cute idea, well executed as a first step and well explained, with a bit of unsatisfying lack of depth in the experiments in favor of breadth that doesn't add all that much.

This paper’s high-level goal is to evaluate how well GAN-type structures for generating text are performing, compared to more traditional maximum likelihood methods. In the process, it zooms into the ways that the current set of metrics for comparing text generation fail to give a well-rounded picture of how models are performing. 

In the old paradigm, of maximum likelihood estimation, models were both trained and evaluated on a maximizing the likelihood of each word, given the prior words in a sequence. That is, models were good when they assigned high probability to true tokens, conditioned on past tokens. However, GANs work in a fundamentally new framework, in that they aren’t trained to increase the likelihood of the next (ground truth) word in a sequence, but to generate a word that will make a discriminator more likely to see the sentence as realistic. Since GANs don’t directly model the probability of token t, given prior tokens, you can’t evaluate them using this maximum likelihood framework. 

This paper surveys a range of prior work that has evaluated GANs and MLE models on two broad categories of metrics, occasionally showing GANs to perform better on one or the other, but not really giving a way to trade off between the two. 
- The first type of metric, shorthanded as “quality”, measures how aligned the generated text is with some reference corpus of text: to what extent your generated text seems to “come from the same distribution” as the original. BLEU, a heuristic frequently used in translation, and also leveraged here, measures how frequently certain sets of n-grams occur in the reference text, relative to the generated text. N typically goes up to 4, and so in addition to comparing the distributions of single tokens in the reference and generated, BLEU also compares shared bigrams, trigrams, and quadgrams (?) to measure more precise similarity of text. 
- The second metric, shorthanded as “diversity” measures how different generated sentences are from one another. If you want to design a model to generate text, you presumably want it to be able to generate a diverse range of text - in probability terms, you want to fully sample from the distribution, rather than just taking the expected or mean value. Linguistically, this would be show up as a generator that just generates the same sentence over and over again. This sentence can be highly representative of the original text, but lacks diversity. One metric used for this is the same kind of BLEU score, but for each generated sentence against a corpus of prior generated sentences, and, here, the goal is for the overlap to be as low as possible 

The trouble with these two metrics is that, in their raw state, they’re pretty incommensurable, and hard to trade off against one another. The authors of this paper try to address this by observing that all models trade off diversity and quality to some extent, just by modifying the entropy of the conditional token distribution they learn. If a distribution is high entropy, that is, if it spreads probability out onto more tokens, it’s likelier to bounce off into a random place, which increases diversity, but also can make the sentence more incoherent. By contrast, if a distribution is too low entropy, only ever putting probability on one or two words, then it will be only ever capable of carving out a small number of distinct paths through word space. 
The below table shows a good example of what language generation can look like at high and low levels of entropy 
https://i.imgur.com/YWGXDaJ.png

The entropy of a softmax distribution be modified, without changing the underlying model, by changing the *temperature* of the softmax calculation. So, the authors do this, and, as a result, they can chart out that model’s curve on the quality/diversity axis. Conceptually, this is asking “at a range of different quality thresholds, how good is this model’s diversity,” and vice versa. I mentally analogize this to a ROC curve, where it’s not really possible to compare, say, precision of models that use different thresholds, and so you instead need to compare the curve over a range of different thresholds, and compare models on that.  

https://i.imgur.com/C3zdEjm.png

When they do this for GANs and MLEs, they find that, while GANs might dominate on a single metric at a time, when you modulate the temperature of MLE models, they’re able to achieve superior quality when you tune them to commensurate levels of diversity. 

This paper starts by introducing a trick to reduce the variance of stochastic gradient variational Bayes (SGVB) estimators. In neural networks, SGVB consists in learning a variational (e.g. diagonal Gaussian) posterior over the weights and biases of neural networks, through a procedure that (for the most part) alternates between adding (Gaussian) noise to the model's parameters and then performing a model update with backprop. 

The authors present a local reparameterization trick, which exploits the fact that the Gaussian noise added into the weights could instead be added directly into the pre-activation (i.e. before the activation fonction) vectors during forward propagation. This is due to the fact that computing the pre-activation is a linear operation, thus noise at that level is also Gaussian. The advantage of doing so is that, in the context of minibatch training, one can efficiently then add independent noise to the pre-activation vectors for each example of the minibatch. The nature of the local reparameterization trick implies that this is equivalent to using one corrupted version of the weights for each example in the minibatch, something that wouldn't be practical computationally otherwise. This is in fact why, in normal SGVB, previous work would normally use a single corrupted version of the weights for all the minibatch. 

The authors demonstrate that using the local reparameterization trick yields stochastic gradients with lower variance, which should improve the speed of convergence.

Then, the authors demonstrate that the Gaussian version of dropout (one that uses multiplicative Gaussian noise, instead of 0-1 masking noise) can be seen as the local reparameterization trick version of a SGVB objective, with some specific prior and variational posterior. In this SGVB view of Gaussian dropout, the dropout rate is an hyper-parameter of this prior, which can now be tuned by optimizing the variational lower bound of SGVB. In other words, we now have a method to also train the dropout rate! Moreover, it becomes possible to tune an individual dropout rate parameter for each layer, or even each parameter of the model.

Experiments on MNIST confirm that tuning that parameter works and allows to reach good performance of various network sizes, compared to using a default dropout rate. 


##### My two cents

This is another thought provoking connection between Bayesian learning and dropout. Indeed, while Deep GPs have allowed to make a Bayesian connection with regular (binary) dropout learning \cite{journals/corr/GalG15}, this paper sheds light on a neat Bayesian connection for the Gaussian version of dropout. This is great, because it suggests that Gaussian dropout training is another legit way of modeling uncertainty in the parameters of neural networks. It's also nice that that connection also yielded a method for tuning the dropout rate automatically.

I hope future work (by the authors or by others) can evaluate the quality of the corresponding variational posterior in terms of estimating uncertainty in the network and, in particular, in obtaining calibrated output probabilities.

Little detail: I couldn't figure out whether the authors tuned a single dropout rate for the whole network, or used many rates, for instance one per parameter, as they suggest can be done.

This paper performs a comparitive study of recent advances in deep learning with human-like learning from a cognitive science point of view. Since natural intelligence is still the best form of intelligence, the authors list a core set of ingredients required to build machines that reason like humans.

- Cognitive capabilities present from childhood in humans.  
    - Intuitive physics; for example, a sense of plausibility of object trajectories, affordances.
    - Intuitive psychology; for example, goals and beliefs.
- Learning as rapid model-building (and not just pattern recognition).
    - Based on compositionality and learning-to-learn.
    - Humans learn by inferring a general schema to describe goals, object types and interactions. This enables learning from few examples. 
    - Humans also learn richer conceptual models.
        - Indicator: variety of functions supported by these models: classification, prediction, explanation, communication, action, imagination and composition.
        - Models should hence have strong inductive biases and domain knowledge built into them; structural sharing of concepts by compositional reuse of primitives.
- Use of both model-free and model-based learning.
    - Model-free, fast selection of actions in simple associative learning and discriminative tasks.
    - Model-based learning when a causal model has been built to plan future actions or maximize rewards.
- Selective attention, augmented working memory, and experience replay are low-level promising trends in deep learning inspired from cognitive psychology.
    - Need for higher-level aforementioned ingredients.

Disclaimer: I am an author

# Intro

Experience replay (ER) and generative replay (GEN) are two effective continual learning strategies. In the former, samples from a stored memory are replayed to the continual learner to reduce forgetting. In the latter, old data is compressed with a generative model and generated data is replayed to the continual learner. Both of these strategies assume a random sampling of the memories. But learning a new task doesn't cause **equal** interference (forgetting) on the previous tasks!  

In this work, we propose a controlled sampling of the replays. Specifically, we retrieve the samples which are most interfered, i.e. whose prediction will be most negatively impacted by the foreseen parameters update. The method is called Maximally Interfered Retrieval (MIR).

## Cartoon for explanation

https://i.imgur.com/5F3jT36.png

Learning about dogs and horses might cause more interference on lions and zebras than on cars and oranges. Thus, replaying lions and zebras would be a more efficient strategy.

# Method

1) incoming data: $(X_t,Y_t)$

2) foreseen parameter update: $\theta^v= \theta-\alpha\nabla\mathcal{L}(f_\theta(X_t),Y_t)$

### applied to ER (ER-MIR)
3) Search for the top-$k$ values $x$ in the stored memories using the criterion $$s_{MI}(x) = \mathcal{L}(f_{\theta^v}(x),y) -\mathcal{L}(f_{\theta}(x),y)$$

### or applied to GEN (GEN-MIR)
3)   
$$
     \underset{Z}{\max} \, \mathcal{L}\big(f_{\theta^v}(g_\gamma(Z)),Y^*\big) -\mathcal{L}\big(f_{\theta}(g_\gamma(Z)),Y^*\big)
$$
$$
         \text{s.t.}   \quad ||z_i-z_j||_2^2 > \epsilon \forall  z_i,z_j \in Z \,\text{with} \, z_i\neq z_j
$$
i.e. search in the latent space of a generative model $g_\gamma$ for samples that are the most forgotten given the foreseen update.

4) Then add theses memories to incoming data $X_t$ and train $f_\theta$

# Results

### qualitative
https://i.imgur.com/ZRNTWXe.png

Whilst learning 8s and 9s (first row), GEN-MIR mainly retrieves 3s and 4s (bottom two rows) which are similar to 8s and 9s respectively.

### quantitative 

GEN-MIR was tested on MNIST SPLIT and Permuted MNIST, outperforming the baselines in both cases.

ER-MIR was tested on MNIST SPLIT, Permuted MNIST and Split CIFAR-10, outperforming the baselines in all cases.


# Other stuff
### (for avid readers)

We propose a hybrid method (AE-MIR) in which the generative model is replaced with an autoencoder to facilitate the compression of harder dataset like e.g. CIFAR-10.