Deep Neural Networks as Gaussian Processes

In lab meeting this week, we read Deep Neural Networks as Gaussian Processes by Lee, Bahri, Novak, Schoenholz, Pennington and Sohl-Dickstein, and which appeared at ICLR 2018. The paper extends a result derived by Neal (1994); and the authors show that there is a correspondence between deep neural networks and Gaussian processes. After coming up with an efficient method to evaluate the associated kernel, the authors compared the performance of their Gaussian process model with finite width neural networks (trained with SGD) on an image classification task (MNIST, CIFAR-10). They found that the performance of the finite width networks approached that of the Gaussian process performance as the width increased, and that the uncertainty captured by the Gaussian process correlated with mean squared prediction error. Overall, this paper hints at new connections between Gaussian processes and neural networks; and it remains to be seen whether future work can harness this connection in order to extend Gaussian process inference to larger datasets, or to endow neural networks with the ability to capture uncertainty. We look forward to following progress in this field.

Single Layer Neural Networks as Gaussian Processes – Neal 1994

Let us consider a neural network with a single hidden layer. We can write the ith output of the network, z_{i}^{1}, as

z_{i}^{1}(x) = b_{i}^{1} + \sum_{j}^{N_{1}}W_{ij}^{1}x_{j}^{1}(x)

where x_{j}^{1}(x) = \phi(b_{j}^{0} + \sum_{k}^{d_{in}}W_{jk}^{0}x_{k}) is the post-activation of the jth neuron in the hidden layer; \phi(x) is some nonlinearity, and x_{k} is the kth input to the network.

If we now assume that the weights for each layer in the network are sampled i.i.d. from a Gaussian distribution: W_{ij}^{1} \sim \mathcal{N}(0, \dfrac{\sigma_{w}^{2}}{N_{1}}), W_{ij}^{0} \sim \mathcal{N}(0, \dfrac{\sigma_{w}^{2}}{d_{in}}); and that the biases are similarly sampled: b_{i}^{1} \sim \mathcal{N}(0, \sigma_{b}^{2}) and b_{i}^{0} \sim \mathcal{N}(0, \sigma_{b}^{2}); then it is possible to show that, in the limit of N_{1} \rightarrow \infty, z_{i} \sim \mathcal{GP} (0, K_{\phi}), for a kernel K_{\phi} which depends on the nonlinearity. In particular, this follows from application of the Central Limit Theorem: for a fixed input to the network \vec{x}, z_{i}^{1} (\vec{x}) \rightarrow \mathcal{N}(0, \sigma_{b}^{2}+\sigma_{w}^{2}V_{\phi}(x^{1}(\vec{x}))) as N_{1} \rightarrow \infty where V_{\phi}(x^{1}(\vec{x})) \equiv \mathbb{E}[(x^{1}_{i}(\vec{x}))^{2}] (which is the same for all i).

We can now apply a similar argument to the above in order to examine the distribution of ith output of the network for a collection of inputs: that is we can examine the joint distribution of \{z_{i}^{1}(\vec{x}^{\alpha = 1}), z_{i}^{1}(\vec{x}^{\alpha = 2}), ..., z_{i}^{1}(\vec{x}^{\alpha = k})\}. Application of the Multidimensional Central Limit Theorem tells us that, in the limit of N_{1} \rightarrow \infty,

\{z_{i}^{1}(\vec{x}^{\alpha = 1}), z_{i}^{1}(\vec{x}^{\alpha = 2}), ...,z_{i}^{1}(\vec{x}^{\alpha = k})\} \sim \mathcal{N}(0, K_{\phi}),

where K_{\phi} \in \mathbb{R}^{k \times k} and K_{\phi}(\vec{x}, \vec{x}') \equiv \sigma_{b}^{2} + \sigma_{w}^{2}C_{\phi}(\vec{x}, \vec{x'}) and C_{\phi}(\vec{x}, \vec{x'}) \equiv \mathbb{E}[x_{i}^{1}(\vec{x})x_{i}^{1}(\vec{x}')].

Since we get a joint distribution of this form for any finite collection of inputs to the network, we can write that z_{i}^{1} \sim \mathcal{GP}(0, K_{\phi}), as this is the very definition of a Gaussian process.

This result was shown in Neal (1994); and the precise form of the kernel K_{\phi} was derived for the error function (a form of sigmoidal activation function) and Gaussian nonlinearities in Williams (1997).

Deep Neural Networks as Gaussian Processes

Lee et al. use similar arguments to those presented in Neal (1994) to show that the ith output of the lth layer of a network with a Gaussian prior over all of the weights and biases is a sample from a Gaussian process in the limit of N_{l} \rightarrow \infty. They use an inductive argument of the form: suppose that z_{j}^{l-1} \sim \mathcal{GP}(0, K_{\phi}^{l-1}) (the jth output of the (l-1)th layer of the network is sampled from a Gaussian process). Then:

z_{i}^{l} \equiv b_{i}^{l} + \sum_{j=1}^{N_{l}}W_{ij}^{l}x_{j}^{l}(\vec{x})

is Gaussian distributed as N_{l} \rightarrow \infty and any finite collection of \{z_{i}^{l}(\vec{x}^{\alpha=1}), ..., z_{i}^{l}(\vec{x}^{\alpha=k})\} will have a joint multivariate Gaussian distribution, i.e., z_{i}^{l} \sim \mathcal{GP}(0, K_{\phi}^{l}) where

K_{\phi}^{l}(\vec{x}, \vec{x'}) \equiv \mathbb{E}[z_{i}^{l}(\vec{x})z_{i}^{l}(\vec{x'})] = \sigma_{b}^{2} + \sigma_{w}^{2} \mathbb{E}_{z_{i}^{l-1}\sim \mathcal{GP}(0, K_{\phi}^{l-1})}[\phi(z_{i}^{l-1}(\vec{x})) \phi(z_{i}^{l-1}(\vec{x'}))].

If we assume a base kernel of the form K^{0}(\vec{x}, \vec{x'}) \equiv \sigma_{b}^{2} + \sigma_{w}^{2}(\dfrac{\vec{x}\cdot \vec{x'}}{d_{in}}), these recurrence relations can be solved in analytic form for the ReLU nonlinearity (as was demonstrated in Cho and Saul (2009)), and they can be solved numerically for other nonlinearities (and Lee et al., give a method for finding the numerical solution efficiently).

Comparison: Gaussian Processes and Finite Width Neural Networks

Lee et al. went on to compare predictions made via Gaussian process regression with the kernels obtained by solving the above recurrence relations (for nonlinearities ReLU and tanh), with the predictions obtained from finite width neural networks trained with SGD. The task was classification (reformulated as a regression problem) of MNIST digits and CIFAR-10 images. Overall, they found that their “NNGP” often outperformed finite width neural networks with the same number of layers for this task; and they also found that the performance of the finite width networks often approached that of the NNGP as the width of these networks was increased:

Figure 1 of Lee et al. The authors compare the performance of their NNGP to finite width neural networks of the same depth and find that, for many tasks, the NNGP outperforms the finite width networks and that the performance of the finite width networks approaches that of the NNGP as the width is increased.
Advertisements

oi-VAE: output interpretable VAEs

We recently read oi-VAE: Output Interpretable VAEs for Nonlinear Group Factor Analysis, which was published at ICML in 2018 by Samuel Ainsworth, Nicholas Foti, Adrian Lee, and Emily Fox. This paper proposes a nonlinear group factor analysis model that is an adaptation of VAEs to data with groups of observations. The goal is to identify nonlinear relationships between groups and to learn latent dimensions that control nonlinear interactions between groups. This encourages disentangled latent representations among sets of functional groups. A prominent example in the paper is motion capture data, where we desire a generative model of human walking and train on groups of observed recorded joint angles. 

Let us consider observations \mathbf{x} that we group into G groups \mathbf{x} = [\mathbf{x}^{(1)}, \cdots, \mathbf{x}^{(G)}].  We note that the paper does not discuss how to choose the groups and assumes that a grouping has already been specified. The generative model maps a set of latent variables to group-specific generator networks via group-specific mapping matrices \mathbf{W}^{(g)}  such that

\mathbf{z} \sim \mathcal{N}(0, \mathbf{I}) \\ \mathbf{x}^{(g)} \sim \mathcal{N}( f_{\theta_g}^{(g)} (\mathbf{W}^{(g)} \mathbf{z} ))

for each g.

oi-vae

oi-VAE generative model (Ainsworth et al., 2018)

For learning interpretable sets of latents that control separate groups, the key feature of this approach is to place a sparsity-inducing prior on the columns of each \mathbf{W}^{(g)}. The authors use a hierarchical Bayesian sparse prior that when analytically marginalized corresponds to optimizing a group lasso penalty on the columns of \mathbf{W}^{(g)}.

The model is trained in the standard VAE approach by optimizing the ELBO with an amortized inference network q_{\phi}(\mathbf{z} | \mathbf{x}), with the addition of the group lasso penalty and a prior on the parameters of the generator networks

\mathcal{L}(\phi, \theta, \mathcal{W}) = \mathbb{E}_{q_\phi(\mathbf{z} | \mathbf{x})} [ \log p(\mathbf{x} | \mathbf{z}, \mathcal{W}, \theta)] - D_{\mathrm{KL}} ( q_\phi(\mathbf{z} | \mathbf{x})  || p(\mathbf{z}) ) + \log p(\theta) - \lambda \sum_{g,j} \| w_j^{(g)} \|_2

where w_j^{(g)} is the j-th column of \mathbf{W}^{(g)} and \lambda is a parameter controlling the sparsity. The prior \log p(\theta) fixes the scaling of the neural network parameters relative to the mapping matrices \mathbf{W}^{(g)}.

The above objective consists of a differentiable term (the ELBO plus log prior on \theta) plus a convex but non-differential term (group LASSO). Therefore the authors use proximal gradient methods to optimize it. First, they update all parameters using the gradients of the ELBO plus log-prior with respect to \phi, \theta, and \mathcal{W}. Then, they apply the proximal operator

w_{j}^{(g)} \leftarrow \frac{ w_{j}^{(g)} } { \| w_{j}^{(g)} \|_2 } ( \| w_{j}^{(g)} \|_2 - \eta \lambda )_+

to each \mathbf{W}^{(g)} to respect the group lasso penalty, where \eta is a step-size. The authors fixed \lambda for all of their experiments fitting oi-VAE, so one question I had reading was how the authors determined \lambda and how varying \lambda affects the results.

The authors validate the approach on a toy example. They generated synthetic bars data, where one row of a square matrix was sampled from a Gaussian with non-zero mean while the rest of the matrix was sampled from zero-mean noise. The authors fit oi-VAE with each group set to a row of observations, and found that the model learned the appropriate sparse mapping where each latent component mapped to one of the rows. This latent space improved on the VAE, which did not have any discernible structure in the latent space. Importantly, oi-VAE still successfully identified the correct number of latent components (8) and sparse mapping even when the model was fit with double the amount of components.

 

row_data

(a) Synthetic bar data example and (b) reconstruction from oi-VAE. The learned oi-VAE latent-group mappings (c) match the true structure, while a VAE (d) does not learn a sparse structure. 

 

After validating the approach, the authors applied it to motion capture data. Here, the output groups were different joint angles. They trained the oi-VAE model on walking motion capture data. The learned latent dimensions nicely corresponded to intuitive groups of joint angles, such as the left leg (left foot, left lower leg, left upper leg). Next, the imposed structure in the model helped it generate more realistic unconditional samples of walking than the VAE, presumably because the inductive bias allowed oi-VAE to better learn invalid joint angles.

walking_samples

Unconditional walking pose samples from the VAE and oi-VAE models.

These results suggest the oi-VAE is a useful model for discovering nonlinear interactions between groups. In particular, I liked the approach of adding structure in the generative model to gain interpretability, and hypothesize that adding other forms of structure to VAE generative models is a good way to encourage disentangled representations (see a recent example of this in Dieng et al., 2019).

Two questions when using the approach are how to choose \lambda and how to choose the grouping of the data, as this work assumes a grouping has been chosen. In some data we may have prior knowledge about a natural grouping structure but that will not always be the case. However, even without multiple groups, the approach could be useful for learning the number of latent dimensions useful for explaining the amount of data. Finally, we point the reader to factVAE, where the authors further develop this idea to simultaneous learn complementary sparse structure in the inference network and generative model.

Uncertainty Autoencoders: Learning Compressed Representations via Variational Information Maximization

This week we continued the deep generative model theme and discussed Uncertainty Autoencoders: Learning Compressed Representations via Variational Information Maximization. This work is in the context of statistical compressive sensing, which attempts to recover a high-dimensional signal x \in \mathbb { R } ^ { n } from m low dimensional measurements y \in \mathbb { R } ^ { m } in a way that leverages a set of training data X. This is in contrast to traditional compressive sensing, which a priori asserts a sparsity based prior over x and learns the recovery of a single x from a single compressed measurement y.

The central compressive sensing problem can be written as

y = W x + \epsilon.

In traditional compressive sensing,  W is a known random Gaussian matrix that takes linear combinations of the sparse signal x to encode y. Here, \epsilon describes additive Gaussian noise. The goal of compressive sensing is to decode x, given m measurements y. This is traditionally done via LASSO, which minimizes the \ell_1 convex optimization problem \hat { x } = \arg \min _ { x } \| x \| _ { 1 } + \lambda \| y - W x \| _ { 2 } ^ { 2 }. In contrast, statistical compressive sensing is a set of methods which recovers x from y using a procedure that results from training on many examples of x. That is, the decoding of x from y is done in a way that utilizes statistical properties learned from training data X. The encoding, on the other hand, may either involve a known random Gaussian matrix W, as is the case above, or it may also be learned as a part of the training procedure.

There are two statistical compressive sensing methods discussed in this paper. One is prior work that leverages the generative stage of variational autoencoders to reconstruct x from y (Bora et al. 2017). In this paper, the decoding network of learned VAE creates an x from a sample z. This is adapted to a compressive sensing framework of recreating the best x from measurement y by solving the following minimization problem:

\widehat { x } = G \left( \arg \min _ { z } \| y - W G ( z ) \| _ { 2 } \right)

Here, G(z) is the VAE decoding mapping. The optimization is: given some measurements y, we seek to find the z that generates an x such that, when it is compressed, best matches the measurements y.

This VAE-based decompression method is used as a comparison to a new method presented in this paper, called the Uncertainty Autoencoder (UAE) which is able to optionally learn both an encoding process, the mapping of x to y, and a decoding process, the recovery of a reconstructed x from a y. It learns these encoding and decoding mappings by maximizing the mutual information between X and Y, parameterized by encoding and decoding distributions. The objective, derived through information theoretic principles, can be written as:

\max _ { \phi , \theta } \sum _ { x \in \mathcal { D } } \mathbb { E } _ { Q _ { \phi } ( Y | x ) } \left[ \log p _ { \theta } ( x | y ) \right] \stackrel { \mathrm { def } } { = } \mathcal { L } ( \phi , \theta ; \mathcal { D } )

Here, Q _ { \phi } ( Y | X) is an encoding distribution parameterized like the original sparse coding compression mapping \mathcal { N } \left( W  ( X ) , \sigma ^ { 2 } I _ { m } \right), and  \log p _ { \theta } ( x | y ) is a variational distribution that decodes x from y using a deep neural network parameterized by \theta. This objective is maximized by drawing training data samples from what is asserted as a uniform prior over Q, which is simply the data itself, Q_{data}(X).

Using this objective, it is possible to derive an optimal linear encoder W under a Gaussian noise assumption in the limit of infinite noise. This linear encoding, however, is not the same as PCA, which is derived under the assumption of linear decoding. UAE linear compression, instead, makes no assumptions about the nature of the decoding. The authors use this optimal linear compressor W* on MNIST, and use a variety of classification algorithms (k-nearest neighbors, SVMs, etc) on this low-d representation to test classification performance. They find that the UAE linear compression better separates clusters of digits than PCA. It is unclear, however, how this UAE classification performance would compare to linear compression algorithms that are known to work better for classification, such as random projections and ICA. I suspect it will not do as well. Without these comparisons, unclear what use this particular linear mapping provides.

The authors demonstrate that UAE outperforms both LASSO and VAE decoding on reconstruction of MNIST and omniglot images from compressed measurements. They implement two versions of the UAE, one that includes a trained encoding W, and another where W is a random Gaussian matrix, as it is for the other decoding conditions. This allows the reader to distinguish to what extent the UAE decoder p(x|y) does a better job at reconstructing x under the same encoding as the alternative algorithms.

Most interestingly, the authors test the UAE in a transfer compressive sensing condition, where an encoder and decoder is learned on some data, say omniglot, and the decoder is re-learned using compressed measurements from different data, MNIST. In this condition, the training algorithm has no access to the test MNIST signals, but still manages to accurately recover the test signals given their compressed representations. This suggests that reasonably differently structured data may have similar optimal information-preserving linear encodings, and an encoder learned from one type of data can be utilized to create effective reconstruction mappings across a variety of data domains. The applications here may be useful.

It is unclear, however, how well these comparisons hold up in a context where traditional compressive sensing has proven to be very effective, like MRI imaging.

There are interesting parallels to be drawn between the UAE and other recent work in generative modeling and their connections to information theory. As the author’s point out, their objective function corresponds to last week’s \betaVAE objective function with \beta = 0. That is, the UAE objective written above is the same as the VAE objective minus the KL term. Though a VAE with \beta = 0 does not actually satisfy the bound derived from the marginal distribution, and hence is not a valid ELBO, is does satisfy the bound of maximal mutual information. And as Mike derives in the post below, there is an additional connection between rate-distortion theory and the \betaVAE  objective. This suggests that powerful generative models can be derived from either principles of Bayesian modeling or information theory, and connections between the two are just now beginning to be explored.

 

 

 

 

Reductions in representation learning with rate-distortion theory

In lab meeting this week, we discussed unsupervised learning in the context of deep generative models, namely \beta-variational auto-encoders (\beta-VAEs), drawing from the original, Higgins et al. 2017 (ICLR), and its follow-up, Burgess et al. 2018. The classic VAE represents a clever approach to learning highly expressive generative models, defined by a deep neural network that transforms samples from a standard normal distribution to some distribution of interest (e.g., natural images).  Technically, VAE training seeks to maximize a lower bound on the likelihood p_\theta(x) = \int p_\theta(x\mid z) p(z) dz, where p_\theta(x|z) defines the generative mapping from latents z to data x. This “evidence lower bound” (ELBO) depends on a variational approximation to the posterior, q_\phi(z\mid x), which is also parametrized by a deep neural network (the so-called “encoder”).

A crucial drawback to the classic VAE, however, is that the learned latent representations tend to lack interpretability. The \beta-VAE seeks to overcome this limitation by learning “disentangled” representations, in which single latents are sensitive to single generative factors in the data and relatively invariant to others (Bengio et al. 2013). I would call these “intuitively robust” — rotating an apple (orientation) shouldn’t make its latent representation any less red (color) or any less fruity (type). To overcome this challenge, \beta-VAEs optimize a modified ELBO given by:

\underset{\theta,\phi}{\text{maximize}}\:\:\mathbb{E}_{q_\phi(z\mid x)}\left[\log p_\theta(x\mid z)\right]-\beta D_{KL}(q_\phi(z\mid x)\Vert\, p(z))

with and standard VAEs corresponding to \beta=1. The new hyperparameter \beta controls the optimization’s tension between maximizing the data likelihood and limiting the expressiveness of the variational posterior relative to a fixed latent prior p(z)=\mathcal{N}(0,I).

Recent work has been interested in tuning the latent representations of deep generative models (Adversarial Autoencoders (Makhzani et al. 2016), InfoGANS (Chen et al. 2016), Total Correlation VAEs (Chen et al. 2019), among others), but the generalization used by \beta-VAEs in particular looked somehow familiar to me. This is because \beta-VAEs recapitulate the classical rate-distortion theory problem. This was observed briefly also in recent work by Alemi et al. 2018, but I would like to elaborate and show explicitly how \beta-VAEs are reducible to a distortion-rate minimization using deep generative models.

Rate-distortion theory is a theoretical framework for lossy data compression through a noisy channel. This fundamental problem in information theory balances the minimum permissible amount of information (in bits) transmitted across the channel, the “rate”, against the corruption of the original signal, a penalty measured by a “distortion” function d(x,z). Our terminology changes, but the fundamental problem is the same; I made that comparison as obvious as possible in the figure below.

Derivation. Given a dataset \mathcal{D} with a distribution p^*(x), define any statistical mapping q_\phi(z\mid x) that encodes x into a code z. Note that q_\phi is just an encoder, and together they induce a joint distribution p(x,z)=q_\phi(z\mid x)p^*(x) with a marginal p(z)=\int dx\, q_\phi(z\mid x)p^*(x). The distortion-rate optimization would minimize distortion d(\cdot,\cdot) subject to a maximum rate R, i.e.

\underset{q_\phi(z\mid x),p(z)}{\text{minimize}}\:\:\mathbb{E}_{p(x,z)}[d(x,z)]\:\:\text{subject to}\:\: I(x,z)\le R

\Longrightarrow \underset{q_\phi(z\mid x),p(z)}{\text{minimize}}\:\:\mathbb{E}_{p(x,z)}[d(x,z)]-\beta I(x,z)

Consider first the mutual information. We leverage a more tractable upper bound with

I(x,z)=\int dx\,p^*(x)\int dz\,q_\phi(z\mid x)\log\frac{q_\phi(z\mid x)}{p(z)}\le \int dx\,p^*(x)\int dz\,q_\phi(z\mid x)\log\frac{q_\phi(z\mid x)}{m(z)}

\text{since}\:\: D_{KL}\left(p(z)\Vert\, m(z)\right)\Longrightarrow -\int dz\,p(z)\log p(z) \le -\int dz\,p(z)\log m(z)

We’ve replaced the marginal p(z) induced by our choice of encoder q_\phi with another distribution m(z) that makes the optimization more tractable, e.g. \mathcal{N}(0,I) in the VAE. Our objective can be rewritten as

\underset{q_\phi(z\mid x),m(z)}{\text{minimize}}\:\:\mathbb{E}_{p(x,z)}[d(x,z)]-\beta\, \mathbb{E}_{x\sim\mathcal{D}}\left[D_{KL}(q_\phi(z\mid x)\Vert\, m(z))\right]

Suppose the distortion of interest is posterior density (mis)estimation, d(x,z)=-\log p_\theta(x\mid z). Such a function penalizes representations z from which we cannot regenerate an observed data vector x through the decoding network p_\theta with high probability. A typical distortion-rate problem would fix the distortion function, but we choose to learn this decoder. We can optimize the objective for each x to eliminate the outer expectation over the data \mathcal{D}, fix m(z)=\mathcal{N}(0,I), and recover the \beta-VAE objective precisely:

\underset{q_\phi(z\mid x),p_\theta(x\mid z)}{\text{maximize}}\:\:\mathbb{E}_{q_\phi(z\mid x)}\left[\log p_\theta(x\mid z)\right]-\beta\, D_{KL}(q_\phi(z\mid x)\Vert\, m(z))

When \beta> 1, our optimization prioritizes minimizing the second term (rate) over maximizing the first one (distortion). In this sense, the authors’ argument for large \beta can be reinterpreted as an argument for higher-distortion, lower-rate codes (read: latent representations) to encourage interpretability. I edited a figure below from Alemi et al. 2018 to clarify this.

Distortion (D) vs. Rate (R) as a function of free parameters in the rate-distortion problem (and \beta-VAEs) — the proposed method privileges solutions in the top-left quadrant (adapted from Alemi et al. 2018).

Information-theoretic hypotheses abound. Perhaps enforcing optimization in this region could discourage solutions that depend on learning an ultra-powerful decoder (VAE: generator) p_\theta(x\mid z), in other words solutions that depend on a good code, not necessarily a good decode. Does eliminating this possibility simply make room to fish out an ad-hoc interpretable representation, or is there a more sophisticated explanation waiting to be found? We’ll see.

Machine Theory of Mind

In lab meeting last week, we read Machine Theory of Mind, a recent paper from Neil Rabinowitz and his collaborators at DeepMind & Google Brain (a trimmed version of the paper was presented at ICML 2018). Here, Theory of Mind (ToM) is broadly defined as the ability to represent the mental states of others. This paper aims to demonstrate ToM in an artificial agent. While designing & training such an agent constitutes one challenge, the authors must first devise a scenario in which ToM can be convincingly shown. Inspired by the Sally-Anne test — a classic test of ToM from developmental psychology that evaluates whether a child understands that others can hold false beliefs — the authors construct an analogous test, then train an agent to successfully pass it.

The paper is composed of a series of experiments that build in complexity to this final test. Within each experiment are three key parts. First, the environment: a simple 11×11 grid-world containing walls and 4 colored boxes that are all randomly located within each new world. Second, the agents: an individual agent belongs to a particular “species” according to its policy for acting within an environment. Agents can behave randomly, algorithmically, or with a learned policy (via deep RL). The trajectory of a particular agent within a particular environment constitutes an episode. Reward within an episode is generally maximized by navigating to a box of a particular color as fast as possible. However, limitations on the sightedness and statefulness of the agents, as well as the inclusion of more complex subgoals, are adjusted per the needs of each experiment. Finally, the observer: a meta-learning agent, called ToMnet, that parses the episodes of many agents in many environments so as to learn a prior over the behavior of an agent species. At test time, ToMnet uses a novel agent’s recent episodes & its trajectory on a current episode thus far to infer a posterior and make predictions regarding the agent’s future behavior.

From Figure 4 in the paper. (a) An example environment. The red arrows represent the trajectory of a goal-driven agent on a previous episode. (b) A new environment, but the same agent as in (a). Using a prior learned over the agent’s species (e.g., agents of this species seek to consume a box of a specific color) in combination with this agent’s past episodes (e.g., this agent consumed a green box), ToMnet is tasked with predicting the agent’s future behavior. (c) For the new environment in (b), ToMnet outputs the next predicted action (go right) and the predicted color of the box that will eventually be consumed (green). (d) ToMnet also predicts the agent’s successor representation.

To probe ToM in ToMnet, the authors introduce a species of agent with both a limited field of view and a subgoal. For example, an agent that can only see the squares adjacent to it must first navigate to a “star” in the grid-world before finally navigating to the blue box to achieve maximum reward. In certain environments, the agent passes the blue box early in its initial search and so knows directly where to go after finding the star, even if the blue box is not visible to the agent from the star’s location. The test comes when the experimenter now swaps the locations of the boxes while the agent is on the star and the boxes are out of view. While the agent is blind to the swap, ToMnet is not. And so, the analogous Sally-Anne test arises: Will ToMnet not recognize that the swap occurred outside of the agent’s field of view, and thus mistakenly predict that it will move toward the new location of the blue box? Or, will ToMnet recognize that the agent maintains the false belief that no swap has occurred, and thus correctly predict that it will move toward the old location of the blue box?

ToMnet predicts behavior reflecting the agent’s false belief, successfully passing the test. Importantly, this finding is supplemented with results that show that ToMnet is sensitive to how different fields of view make an agent “differentially vulnerable to acquire false beliefs,” and that ToMnet still passes the test even if it had never seen swap events during training. Thus, ToMnet “learns that agents can act based on false beliefs,” providing a compelling proof-of-concept for Machine Theory of Mind.

Insights on representational similarity in neural networks with canonical correlation

For this week’s journal club, we covered “Insights on representational similarity in neural networks with canonical correlation” by Morcos, Raghu, and Bengio, NeurIPS, 2018.  To date, many different convolutional neural networks (CNNs) have been proposed to tackle the object recognition problem, including Inception (Szegedy et al., 2015), ResNet (He et al., 2016), and VGG (Simonyan and Zisserman, 2015). These networks have vastly different architectures but all achieve high accuracy. How can this be the case? One possibility is that although the architectures vary, the representations (i.e., the way these networks encode information about the objects of natural images) are very similar. 

To test this, we first need a metric of similarity. One approach has been “representation similarity analysis” (RSA) (Kriegskorte et al., 2008) which relies on distance matrices to test if two representations are similar. One potential problem with RSA is that some dimensions of the representations may be “noisy” (i.e., dimensions that do not pertain to encoding the input information). For example, during training, some dimensions of the activity of CNN neurons may vary substantially across epochs but are not relevant to encoding object information. These dimensions could mask the signal of relevant dimensions when analyzing a distance matrix. 

One way to avoid this is to try to directly identify the relevant dimensions, allowing us to ignore the noisy dimensions. The authors relied on an old but trusted method called canonical correlation analysis (CCA), which was developed way back in the 1930s (Hotelling, 1936)! CCA has been a handy tool in computational neuroscience, relating the activity of neurons across two populations (Semedo et al., 2014) as well as relating population activity to the output of model neurons (Susillo et al., 2015). Newer methods have been developed that are more appropriate for various problems. These include partial least squares (Höskuldsson, 1988), kernel CCA (Hardoon et al., 2004), as well as a method I developed for my own work called distance covariance analysis (DCA) (Cowley et al., 2017).  The common thread among all of these methods is that they identify dimensions that encode similar information among two or more datasets.

Overview of CCA. CCA is a close relative to linear regression, but whereas linear regression aims at prediction, CCA focuses on correlation—and thus is most suitable for cases in which the investigator seeks intuition of the data.  Given two datasets (e.g., \mathbf{X} \in \mathcal{R}^{k \times N} \textrm{ and }  \mathbf{Y} \in \mathcal{R}^{p \times N}, both centered, where N is the number of samples), CCA seeks to identify a pair of dimensions \mathbf{u} \in \mathcal{R}^k \textrm{ and } \mathbf{v} \in \mathcal{R}^p such that the Pearson’s correlation between the projections \mathbf{u}^T \mathbf{X} \textrm{ and } \mathbf{v}^T \mathbf{Y} is the largest. In other words, CCA identifies linear combinations of the variables in \mathbf{X} \textrm{ and } \mathbf{Y} that are the most linearly-related. CCA need not stop there—it can identify pairs of dimensions that monotonically decrease in correlation. In this way, we can ignore the dimensions with the smallest correlations (which likely are spurious). One fun fact about CCA is that any two identified dimensions in \mathbf{X} are uncorrelated: \textrm{corr}(\mathbf{u}_i^T \mathbf{X}, \mathbf{u}_j^T \mathbf{X}) = 0 \textrm{ for } i \neq j (and the same for \mathbf{v}_i, \mathbf{v}_j). This is different from PCA, whose identified dimensions are both uncorrelated and orthogonal.  The uncorrelatedness of CCA dimensions ensures that we do not include dimensions that contain redundant information. (Implementation details: CCA is solved with singular-value decomposition, but be sure to use a regularized form akin to ridge regression—it was unclear if the authors used regularization). 

Figure 1. Generalizing networks converge to more similar solutions than memorizing networks.

Onto the results. The authors proposed a distance metric of CCA to uncover some intuitive characteristics about deep neural networks. First, they found that different initializations of generalizing networks (i.e., networks trained on labeled natural images) were more similar than different initializations of memorizing networks (i.e., networks trained on the same dataset with randomly-shuffled labels). This is expected, as natural labels likely put a constraint on generalizing networks. Interestingly, when comparing generalizing and memorizing networks (Fig. 1, yellow line, ‘Inter’), they found that generalizing and memorizing networks were as similar as different memorizing networks trained on the same fixed dataset. This suggests that overfitted networks converge on very different solutions for the same problem. Also interesting was that earlier layers of both generalizing and memorizing networks seem to converge on similar solutions, while the later layers diverged. This suggests that earlier layers rely more on the structure of natural images while the later layers rely more on the structure of the labels. Second, they found that wider networks (i.e., networks with more filters per layer) converge to more similar solutions than those of narrower networks.  They argue that this supports the “lottery-ticket” hypothesis that wider networks are more likely to have a sub-network that fits the desired function.  Finally, they found that networks trained with different initializations and learning rates on the same problem converge to different groups of solutions. This highlights the need to try different initializations when training neural networks.

This paper left me thinking a lot about representation in the visual cortex of the brain. Does visual cortical population activity have stable and “noisy” dimensions?  If we reduced the number of visual cortical neurons per visual cortical area (either via lesion or pharmacological intervention) in a developing animal, would these animals have severe perceptual deficits (i.e., their visual system did not have the right lottery ticket when developing)?  Lastly, it seems plausible that humans start out with different initializations of their visual cortices—does that suggest different humans have converged on different solutions to solving visual perception?  If so, it suggests that inter-subject variability may be larger than previously thought.

The Loss-Calibrated Bayesian

By Farhan Damani

In lab meeting this week, we discussed loss-calibrated approximate inference in the context of Bayesian decision theory (Lacoste-Julien et. al. 2011, Cobb et. al. 2018). For many applications, the cost of an incorrect prediction can vary depending on the nature of the mistake. Suppose you are in charge of controlling a nuclear power plant with an unknown temperature \theta. We observe indirect measurements of the temperature D, and we use Bayesian inference to infer a posterior distribution over the temperature given the observations p(\theta|D). The plant is in danger of over-heating and as the operator, you can either keep the plant running or shut it down. Keeping the plant running while the plant’s temperature exceeds a critical threshold T_{\text{critic}} will cause a nuclear meltdown, incurring a huge loss L(\theta > T_{\text{critic}}, \text{'on'}) while shutting off the plant for benign temperatures incurs a minor loss L(\theta < T_{\text{critic}}, \text{'off'})

In figure 1 we observe the true posterior p(\theta|D) is multi-modal. Our suite of approximate inference techniques characterize general properties of the posterior, attempting to match either the first or second moment of p. Both strategies underestimate the posterior mass for the safety-critical region. Instead, the dash-dotted line, while failing to characterize typical properties of the posterior, results in the same decision as the true posterior by optimizing for task-specific utility. The point is the “best” approximate posterior is subjective, and therefore, we should tailor our inferential resources to find an approximation that is well suited for the decision task at hand.

Bayesian decision theory extends the Bayesian paradigm by including a task-specific utility function U(\theta, a), which tells us the utility of taking action a \in \mathcal{A} when the world is in state \theta. According to this view, the optimal action minimizes the posterior risk: \underset{a}{\arg \min} \text{ } \mathcal{R}(a) = \mathbb{E}_{p(\theta|D)}[U(\theta, a)]. Typically, this is computed using a 2-step procedure. First approximate the posterior p(\theta|D) with a q(\theta|D) and then minimize the risk under q. This approach, however, assumes our approximate q measures properties of the posterior that we care about. This by definition requires our utility function, so therefore, we should jointly optimize the approximate posterior with the action that minimizes the posterior risk. Cobb et. al. 2018 show how to derive a variational lower bound that depends on a task-specific utility function. In their setup, they show that minimizing the KL divergence between an approximate posterior q and a calibrated posterior scaled by the utility function results in the standard ELBO loss plus an additional utility-dependent regularization term. This formulation is amenable to stochastic optimization, allowing for the practical deployment of this framework to supervised learning.