Here some of the projects I participated in are listed.
Project: Voicebot for sociological studies
In the last decade artificial intelligence (AI) has
greatly evolved due to the combination of neural networks and
increased computational power. This has led to a variety of
new AI applications like chat- and voicebots. Even though such
applications have become increasingly visible, the moral impacts
they have on humans are relatively unknown. The analysis of
these impacts is an emerging research field in sociology and
psychology. Researchers therefore need easy-to-use chat- and
voicebots to carry out their experiments. The frameworks and
applications needed for such bots should be open-source and free
to use under a public licence so that the approaches can be shared
freely among researchers and be further developed.
In this paper, we develop a chatbot with voice support using
only open-source frameworks with public licences. It is used in a
social experiment about moral decisions to prove its applicability
to real world experiments. Our framework is capable of detecting
and processing speech in an online manner with a latency below
one second.
Checkout the report here .
Project: Local Error Signals
In the recent boom of Deep Learning, Backpropa-
gation of error has become the dominant method for optimizing
Deep Neural Networks and is the only optimization method that
is widely applied in the field. It has been argued however, that
the brain, which has initially inspired some of the development of
DNNs, is biologically unable to optimize networks in such a global
way. Therefore, many training methods have been suggested, in
which a loss and the subsequent weight optimization is calculated
at the level of each individual layer. This raises the question, of
whether these networks learn representations in their layers with
similar properties to Backpropagation trained networks.
In this work, two of these local loss methods, the Prediction
Similarity loss and the Direct Difference Target Propagation loss
have been used to train simple convolutional neural networks and
these networks have been compared to their Backpropagation
trained counterparts in terms of accuracies and layer-wise data
representation.
Our results give some initial evidence towards the assumption
that the Prediction Similarity loss might be a training method
that is able to approximate Backpropagation both in terms of
performance as well as layer-wise data representations. Direct
Difference Target Propagation however did not perform well
with convolutional models and therefore no statements about its
similarity to Backpropagation can be made.
Checkout the report here .
Master Thesis: Hypernetworks: Generating Neural Networks with Neural Networks
In reinforcement learning (RL) gradient based methods have shown to efficiently master many
games in a superhuman manner. New advances in meta learning aim to improve these playing
agents in terms of sample efficiency by introducing new methods of ’learning how to learn’.
They also try to enhance the adaptation capabilities of the agents to quickly changing and
diverse playing styles of their opponents. To study and evaluate these approaches, a diverse
set of agents is needed.
In this work hypernetworks will be evaluated as a method to create a diverse set of agents.
Hypernetworks are neural networks that generate weights for another neural network. They
can be used in a similar fashion as variational autoencoders (VAEs) or generative adversarial
networks (GANs) to generate a mapping from a low dimensional (random) input vector to high
dimensional weight space. The loss function of such a hypernetwork contains two properties:
accuracy and diversity. The latter is used to ensure that the network generates a distribution of
agents.
The approach is bench marked against individually trained agents as well as an evolution-
ary method, where the selection criterion consists also of both, an accuracy and a diversity
term. The different approaches are tested on an image classification task, a graph game and an
imperfect information card game.
Checkout the code and thesis .
Project: Machine Learning Turtle Bot
This project has been carried out as part of the Industrial Distributed Control Systems
project module under the Department of Industry Grade Networks and Clouds in the
Faculty for Electrical Engineering and Computer Science (IV) at the Technical University of Berlin.
This project attempts to demonstrate the feasibility of using a Deep
Reinforcement Learning algorithm such as Deep Q Networks (DQN) for Autonomous
Navigation using the latest Open Source Techstack available such as ROS2, gazebo9,
and turtlebot3 Machine Learning packages (Keras and Tensorflow). It lays out the various processes
that were followed to install, setup the code, and discusses any limitations
faced in the process. The packages used are studied for their scalability and flexibility to tweak the
various parameters and functions, and thereby enhance the learning
capabilities of the model. The training process was conducted on the DQN model in
two different simulation environments, solely relying on Laser Distance Sensor data for
navigation. The results recorded show successful navigation attempts in a relatively low
number of episodes and includes several observations regarding how the various parameters chosen,
affect the training outcomes. The scope of modifying these values and
extended training and simulation capabilities have been briefly explored in the report.
Checkout the report here .