Dermatologist versus artificial intelligence confidence in dermoscopy diagnosis: Complementary information that may affect decision-making.

In dermatology, deep learning may be applied for skin lesion classification. However, for a given input image, a neural network only outputs a label, obtained using the class probabilities, which do not model uncertainty. Our group developed a novel method to quantify uncertainty in stochastic neural networks. In this study, we aimed to train such network for skin lesion classification and evaluate its diagnostic performance and uncertainty, and compare the results to the assessments by a group of dermatologists. By passing duplicates of an image through such a stochastic neural network, we obtained distributions per class, rather than a single probability value. We interpreted the overlap between these distributions as the output uncertainty, where a high overlap indicated a high uncertainty, and vice versa. We had 29 dermatologists diagnose a series of skin lesions and rate their confidence. We compared these results to those of the network. The network achieved a sensitivity and specificity of 50% and 88%, comparable to the average dermatologist (respectively 68% and 73%). Higher confidence/less uncertainty was associated with better diagnostic performance both in the neural network and in dermatologists. We found no correlation between the uncertainty of the neural network and the confidence of dermatologists (R = -0.06, p = 0.77). Dermatologists should not blindly trust the output of a neural network, especially when its uncertainty is high. The addition of an uncertainty score may stimulate the human-computer interaction.

Model Reduction Through Progressive Latent Space Pruning in Deep Active Inference.

Although still not fully understood, sleep is known to play an important role in learning and in pruning synaptic connections. From the active inference perspective, this can be cast as learning parameters of a generative model and Bayesian model reduction, respectively. In this article, we show how to reduce dimensionality of the latent space of such a generative model, and hence model complexity, in deep active inference during training through a similar process. While deep active inference uses deep neural networks for state space construction, an issue remains in that the dimensionality of the latent space must be specified beforehand. We investigate two methods that are able to prune the latent space of deep active inference models. The first approach functions similar to sleep and performs model reduction post hoc. The second approach is a novel method which is more similar to reflection, operates during training and displays "aha" moments when the model is able to reduce latent space dimensionality. We show for two well-known simulated environments that model performance is retained in the first approach and only diminishes slightly in the second approach. We also show that reconstructions from a real world example are indistinguishable before and after reduction. We conclude that the most important difference constitutes a trade-off between training time and model performance in terms of accuracy and the ability to generalize, via minimization of model complexity.

Data-Efficient Sensor Upgrade Path Using Knowledge Distillation.

Deep neural networks have achieved state-of-the-art performance in image classification. Due to this success, deep learning is now also being applied to other data modalities such as multispectral images, lidar and radar data. However, successfully training a deep neural network requires a large reddataset. Therefore, transitioning to a new sensor modality (e.g., from regular camera images to multispectral camera images) might result in a drop in performance, due to the limited availability of data in the new modality. This might hinder the adoption rate and time to market for new sensor technologies. In this paper, we present an approach to leverage the knowledge of a teacher network, that was trained using the original data modality, to improve the performance of a student network on a new data modality: a technique known in literature as knowledge distillation. By applying knowledge distillation to the problem of sensor transition, we can greatly speed up this process. We validate this approach using a multimodal version of the MNIST dataset. Especially when little data is available in the new modality (i.e., 10 images), training with additional teacher supervision results in increased performance, with the student network scoring a test set accuracy of 0.77, compared to an accuracy of 0.37 for the baseline. We also explore two extensions to the default method of knowledge distillation, which we evaluate on a multimodal version of the CIFAR-10 dataset: an annealing scheme for the hyperparameter α and selective knowledge distillation. Of these two, the first yields the best results. Choosing the optimal annealing scheme results in an increase in test set accuracy of 6%. Finally, we apply our method to the real-world use case of skin lesion classification.

Active Vision for Robot Manipulators Using the Free Energy Principle.

Occlusions, restricted field of view and limited resolution all constrain a robot's ability to sense its environment from a single observation. In these cases, the robot first needs to actively query multiple observations and accumulate information before it can complete a task. In this paper, we cast this problem of active vision as active inference, which states that an intelligent agent maintains a generative model of its environment and acts in order to minimize its surprise, or expected free energy according to this model. We apply this to an object-reaching task for a 7-DOF robotic manipulator with an in-hand camera to scan the workspace. A novel generative model using deep neural networks is proposed that is able to fuse multiple views into an abstract representation and is trained from data by minimizing variational free energy. We validate our approach experimentally for a reaching task in simulation in which a robotic agent starts without any knowledge about its workspace. Each step, the next view pose is chosen by evaluating the expected free energy. We find that by minimizing the expected free energy, exploratory behavior emerges when the target object to reach is not in view, and the end effector is moved to the correct reach position once the target is located. Similar to an owl scavenging for prey, the robot naturally prefers higher ground for exploring, approaching its target once located.

Learning Generative State Space Models for Active Inference.

In this paper we investigate the active inference framework as a means to enable autonomous behavior in artificial agents. Active inference is a theoretical framework underpinning the way organisms act and observe in the real world. In active inference, agents act in order to minimize their so called free energy, or prediction error. Besides being biologically plausible, active inference has been shown to solve hard exploration problems in various simulated environments. However, these simulations typically require handcrafting a generative model for the agent. Therefore we propose to use recent advances in deep artificial neural networks to learn generative state space models from scratch, using only observation-action sequences. This way we are able to scale active inference to new and challenging problem domains, whilst still building on the theoretical backing of the free energy principle. We validate our approach on the mountain car problem to illustrate that our learnt models can indeed trade-off instrumental value and ambiguity. Furthermore, we show that generative models can also be learnt using high-dimensional pixel observations, both in the OpenAI Gym car racing environment and a real-world robotic navigation task. Finally we show that active inference based policies are an order of magnitude more sample efficient than Deep Q Networks on RL tasks.