Ethem Alpaydın

Ethem Alpaydin

Many researchers investigated neural networks with some of their weights fixed to values randomly drawn from a given distribution, e.g., $N(0, I)$. Our proposed HalfNet draws random weights from $N(0, Σ)$, where $Σ$, which defines the geometry of the distribution, has a low-rank factorization that we learn from data. Experiments on MNIST and CIFAR-10 demonstrate that HalfNet can match the performance of fully trained multilayer perceptrons while using substantially fewer parameters. Spectral analysis indicates that much of the predictive power of neural networks lies in the geometry of their weight space rather than in the precise values of individual parameters, and we observe that accuracy scales smoothly with rank. HalfNet is not a neural architecture trick for low-rank structure; it implements a data-dependent random embedding that can also be interpreted through supervised metric learning, or random-feature and kernel perspectives.

Ozan İrsoy, Ethem Alpaydın

Explainability is becoming an increasingly important topic for deep neural networks. Though the operation in convolutional layers is easier to understand, processing becomes opaque in fully-connected layers. The basic idea in our work is that each instance, as it flows through the layers, causes a different activation pattern in the hidden layers and in our Paths methodology, we cluster these activation vectors for each hidden layer and then see how the clusters in successive layers connect to one another as activation flows from the input layer to the output. We find that instances of the same class follow a small number of cluster sequences over the layers, which we name ``decision paths." Such paths explain how classification decisions are typically made, and also help us determine outliers that follow unusual paths. We also propose using the Sankey diagram to visualize such pathways. We validate our method with experiments on two feed-forward networks trained on MNIST and CELEB data sets, and one recurrent network trained on PenDigits.

Ethem Alpaydin

We propose a ``half'' layer of hidden units that has some of its weights randomly set and some of them trained. A half unit is composed of two stages: First, it takes a weighted sum of its inputs with fixed random weights, and second, the total activation is multiplied and then translated using two modifiable weights, before the result is passed through a nonlinearity. The number of modifiable weights of each hidden unit is thus two and does not depend on the fan-in. We show how such half units can be used in the first or any later layer in a deep network, possibly following convolutional layers. Our experiments on MNIST and FashionMNIST data sets indicate the promise of half layers, where we can achieve reasonable accuracy with a reduced number of parameters due to the regularizing effect of the randomized connections.

Ozan İrsoy, Ethem Alpaydın

Dropout is a very effective method in preventing overfitting and has become the go-to regularizer for multi-layer neural networks in recent years. Hierarchical mixture of experts is a hierarchically gated model that defines a soft decision tree where leaves correspond to experts and decision nodes correspond to gating models that softly choose between its children, and as such, the model defines a soft hierarchical partitioning of the input space. In this work, we propose a variant of dropout for hierarchical mixture of experts that is faithful to the tree hierarchy defined by the model, as opposed to having a flat, unitwise independent application of dropout as one has with multi-layer perceptrons. We show that on a synthetic regression data and on MNIST and CIFAR-10 datasets, our proposed dropout mechanism prevents overfitting on trees with many levels improving generalization and providing smoother fits.

Ozan İrsoy, Ethem Alpaydın

Traditionally, deep learning algorithms update the network weights whereas the network architecture is chosen manually, using a process of trial and error. In this work, we propose two novel approaches that automatically update the network structure while also learning its weights. The novelty of our approach lies in our parameterization where the depth, or additional complexity, is encapsulated continuously in the parameter space through control parameters that add additional complexity. We propose two methods: In tunnel networks, this selection is done at the level of a hidden unit, and in budding perceptrons, this is done at the level of a network layer; updating this control parameter introduces either another hidden unit or another hidden layer. We show the effectiveness of our methods on the synthetic two-spirals data and on two real data sets of MNIST and MIRFLICKR, where we see that our proposed methods, with the same set of hyperparameters, can correctly adjust the network complexity to the task complexity.

Ozan İrsoy, Ethem Alpaydın

Recently proposed budding tree is a decision tree algorithm in which every node is part internal node and part leaf. This allows representing every decision tree in a continuous parameter space, and therefore a budding tree can be jointly trained with backpropagation, like a neural network. Even though this continuity allows it to be used in hierarchical representation learning, the learned representations are local: Activation makes a soft selection among all root-to-leaf paths in a tree. In this work we extend the budding tree and propose the distributed tree where the children use different and independent splits and hence multiple paths in a tree can be traversed at the same time. This ability to combine multiple paths gives the power of a distributed representation, as in a traditional perceptron layer. We show that distributed trees perform comparably or better than budding and traditional hard trees on classification and regression tasks.

Ozan İrsoy, Ethem Alpaydın

We discuss an autoencoder model in which the encoding and decoding functions are implemented by decision trees. We use the soft decision tree where internal nodes realize soft multivariate splits given by a gating function and the overall output is the average of all leaves weighted by the gating values on their path. The encoder tree takes the input and generates a lower dimensional representation in the leaves and the decoder tree takes this and reconstructs the original input. Exploiting the continuity of the trees, autoencoder trees are trained with stochastic gradient descent. On handwritten digit and news data, we see that the autoencoder trees yield good reconstruction error compared to traditional autoencoder perceptrons. We also see that the autoencoder tree captures hierarchical representations at different granularities of the data on its different levels and the leaves capture the localities in the input space.

Olcay Taner Yildiz, Ethem Alpaydin

Statistical tests that compare classification algorithms are univariate and use a single performance measure, e.g., misclassification error, $F$ measure, AUC, and so on. In multivariate tests, comparison is done using multiple measures simultaneously. For example, error is the sum of false positives and false negatives and a univariate test on error cannot make a distinction between these two sources, but a 2-variate test can. Similarly, instead of combining precision and recall in $F$ measure, we can have a 2-variate test on (precision, recall). We use Hotelling's multivariate $T^2$ test for comparing two algorithms, and when we have three or more algorithms we use the multivariate analysis of variance (MANOVA) followed by pairwise post hoc tests. In our experiments, we see that multivariate tests have higher power than univariate tests, that is, they can detect differences that univariate tests cannot. We also discuss how multivariate analysis allows us to automatically extract performance measures that best distinguish the behavior of multiple algorithms.